This PDF file contains the front matter associated with SPIE Proceedings Volume 12803, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Target localization is one of the most widespread applications of WSN (Wireless Sensor Network). However, in a number of specific harsh scenarios, physical location of anchor node may drift because of influence of external environment or human factors, which in turn leads to the estimated location of unknown node is not accurate. Therefore, a reliable node localization algorithm for WSN based on weighted AHP-GRA is proposed in this paper. Firstly, distance change among nodes in localization scenarios is counted and reference sequence is constructed, then comprehensive influence matrix is obtained by environmental factors. Next, the weighted comprehensive influence matrix is further obtained by combining the two. After that, the node confidence level is obtained based on AHP-GRA, and three reliable nodes with the highest confidence level are taken as anchor nodes to complete localization by using trilateral localization method. Simulation results show that the algorithm in this paper effectively improves localization effect of WSN in complex scenes. Compared with other algorithms, the algorithm in this paper has great advantages in selection of reliable nodes, localization accuracy, and localization coverage.

In this paper, an infrared and color image fusion algorithm based on guided filtering was proposed . The intensity and chromaticity component of the color image in a color space, e.g., HSV, was extracted first. Secondly, in order to avoid edge blurring and reduce computing costs, guided filtering was used to decompose the intensity component of the infrared image and color image to obtain their base layer and detail layer, respectively. Then, the two base and two detail layers were fused using the proposed method separately. The clear and supplementary areas were distinguished by the sum of gradients, the initial base layer was obtained by preliminary fusion, and the intensity level of the fused base layer was adjusted similarly to that of the color image. The two detail layers were fused simply by selecting their maximum absolute value of the gradient map in each pixel, and the intensity component of the fused image was obtained by inverse transformation. Finally, the fused color image was reconstructed by merging the fused intensity component and original chromaticity components. Multiple sets of images were selected for testing the proposed algorithm, and some state-of-art algorithms in the experiment, and the results show that the proposed algorithm had good image visibility, stable color, and fast fusion speed.

To accurately predict the existing tunnel deformation induced by shield undercrossing, a BO-Bi-LSTM deep learning network model is developed and trained using shield construction parameters, geological condition parameters, tunnel design parameters and existing tunnel deformation data. By using Bayesian Optimization (BO) algorithm to optimize hyperparameters of Bi-directional Long Short-Term Memory (Bi-LSTM), a prediction model of existing tunnel deformation induced by shield undercrossing based on BO-Bi-LSTM is established and compared with other neural network models. The main findings of this study are as follows: (1) The prediction models established for horizontal and vertical displacement induced by shield undercrossing have high accuracy, and the test sets R2 are 0.935 and 0.924, RMSE are 0.504 and 0.903, and MAE are 0.415 and 0.824, respectively. (2) Compared with Bi-LSTM and GRU models, the proposed existing tunnel deformation prediction model has the highest accuracy and accuracy, and realizes real-time and reliable prediction of tunnel deformation.

Tide images collected by outdoor cameras are susceptible to complex environmental factors and seriously affect the accuracy of object recognition algorithm. Therefore, an improved Stacked-Denoising Auto Encoder (SDAE) tide image denoising algorithm in complex environment is proposed. In order to increase the number of network layers and find features faster, an SDAE based on convolutional network is proposed. Inside the network architecture of SDAE, layers including input layer, convolutional layer, pooling layer, full connection layer and output convolutional pooling layer are introduced to effectively extract image noise features and reduce network training parameters. The experimental results show that the improved algorithm can effectively remove single noises such as blur, illumination imbalance, rain and fog in complex environment, and the PSNR value and SSIM value of images of various noise types are improved by at least 13% and 33% respectively, and the SSIM value of the denoised rain images is close to 1.

In this paper, in view of the problems of large computation and complex network model in the current corneal disease classification algorithm, a new residual network model Resnet-V is designed on the basis of feature extraction ResNet-50. The algorithm extracts and reduces the feature of images through the residual network, and maps the image feature vectors to multiple hash tables through E2LSH to achieve efficient image retrieval and classification. The residual network is used to extract and reduce the feature data to obtain a low-dimensional feature vector representation. The use of a pooling layer and downsampling in the residual network will reduce the spatial resolution of the feature map, so that many details are lost, thereby affecting the accuracy of the model for graph classification, based on the feature vector of the residual network, select the appropriate hash function and parameters to realize the hash encoding of the feature vector. Experiments were conducted on iChallenge-PM, a publicly available medical dataset. Experimental results show that the proposed algorithm achieves high accuracy and low loss value, and the classification recognition rate is better than that of other model algorithms.

This paper proposes a color image encryption strategy based on the Zig-Zag transform, DNA coding technology and fractional-order memristive hyper-chaotic circuit system (FOMHCC). First, the chaotic sequences are generated from FOMHCC using control parameters and initial values. Next, a modified Zig-Zag method is used to confuse the tricolour of the colour image, and then DNA encoding is performed. The DNA coded matrix is then obfuscated and diffused through the chaotic sequences. Finally, DNA decoding is performed to get an encrypted image. The security analysis by simulation confirm that the algorithm is able to withstand multiple attacks and provides excellent image encryption performance.

When the demand for network resources exceeds the available capacity of the network, network congestion occurs on the Internet. Network congestion can cause transmission delays, packet loss, and even congestion collapse. When this happens, all communication across the network comes to a halt. Therefore, congestion has long been a concern. Although traditional congestion control algorithms have performed well in early simple networks, their limitations prevent them from being effective before network congestion occurs, making them difficult to cope with the challenges of network complexity and service diversification. In this paper, we model network congestion control as a Markov decision process and optimize congestion control strategies using a deep Q network in deep reinforcement learning, proposing a congestion control algorithm, QNCC, that is purely data-driven and does not rely on any assumptions. QNCC uses a fully connected neural network to approximate the value function, enabling it to automatically learn features and have good generalization ability. Experiments show that QNCC performs better overall than traditional congestion control algorithms such as TCP-Cubic and TCP-Vegas in multiple network scenarios.

Telecom fraud has become a pressing issue, with most research targeting effective countermeasures. However, extracting valuable rules from fraud data to improve case handling efficiency remains under-explored. The Apriori algorithm, commonly used for association rule mining, faces challenges due to its low efficiency and scalability arising from numerous frequent and candidate item sets. To address this problem, we propose a fast Apriori algorithm. The main idea is as follows: First, we establish a similarity measurement model based on information entropy, and combine similar items to significantly reduce the number of frequent item sets and candidate item sets. Second, we optimize the process of generating frequent item sets and candidate item sets and reduce the number of database scans. Third, we apply the improved algorithm to mine the data set of telecom fraud cases in a city, and obtain some meaningful association rules that reflect the relationships among the duration of the crime, the reporting time, and the amount of fraud. These rules provide a new perspective and idea for investigating telecom fraud cases.

The lithology identification of cuttings based on mineral element content data plays an important role in oil and gas exploration. Currently, the method for acquiring the element content of cuttings is to air-dry cuttings obtained in the mud logging process, and then use x-ray fluorescence (XRF) technology to obtain the types and contents of the main elements in the cuttings. In this paper, a method for identifying the lithology of cuttings based on channel attention mechanism is proposed for the mineral element content data of cuttings obtained by XRF technology. Specifically, the existing one-dimensional data composed of mineral elements are input into the network model. The channels are first expanded to introduce more features. Then, the features obtained by the multi-channels are fused to obtain features that are more conducive to the identification of cuttings lithology. To avoid introducing too much noise during channel change, the SE module is improved and applied to the one-dimensional convolutional neural network in this paper. Additionally, the features of different channels are weighted by autonomous learning, ensuring that the features related to the current task have a higher contribution to the network. By reducing the influence of invalid features caused by changing channels, this method safeguards the reliability of the features used for debris classification. The experiment results show that the cuttings recognition algorithm proposed in this paper has higher accuracy than the comparison algorithm.

With the continuous increase of drilling speed in rock debris logging, the corresponding rock identification speed is required to be constantly improved. Therefore, this paper proposes a rock identification algorithm based on rock debris images, which is based on multi-task learning and pseudo-labels. The algorithm first uses affine transformation matrix for data augmentation and builds a multi-task neural network framework based on hard sharing mechanism, which uses rock debris color and texture as two sub-task modules for classification. Secondly, the algorithm uses unlabeled rock debris images as pseudo-labels for the original images and proposes a pseudo-label filtering method for multi-task learning, which adds pseudo-labeled images to the original dataset for model iteration to improve the accuracy and generalization ability of the model. The experiments show that this algorithm can meet the current requirements of rock identification speed and accuracy in rock debris logging.

In China, the frequency and variety of roadbed diseases have been increasing, leading to a rise in road safety accidents caused by such issues. Identifying the location, morphology, and development stage of these hidden roadbed diseases is crucial for ensuring the safe operation and maintenance of highways. Current non-destructive testing methods, like ground-penetrating radar, heavily rely on subjective data processing and interpretation. This study focuses on common roadbed diseases—roadbed looseness and voids. It establishes training and test sets using existing images in a 1:4 ratio and labels disease types in each image. The Faster R-CNN algorithm is enhanced to faster_rcnn_resnet101 and rfcn_resnet101 versions. Training both algorithms and analyzing loss, test recognition area, and accuracy reveals faster_rcnn_resnet101’s superior performance. It achieves a total loss of 0.0392, a recognition accuracy of 86.0%, compared to rfcn_resnet101’s recognition count of 324 (higher than faster_rcnn_resnet101’s 284). Considering all aspects, the faster_rcnn_resnet101 algorithm is better suited for intelligently recognizing roadbed diseases on urban roads.

Fractures of the talar neck and body are associated with spine fractures and scoliosis deformity, which affect cosmetic appearance and cause difficulty in ambulation. The implant design for talus surgery is thriving as a functional alternative in case of severe talar destruction, focusing on segmentation and reconstruction of the talus’s shape. However, manual segmentation of the talus is time-consuming and subjective. In this study, we exploited the automatic segmentation framework to efficiently train a deep learning-based model to accurately segment the talus based on computed tomography imaging. We developed three model configurations with nnU-Net and investigated their Dice similarity coefficients (DSC) and 95% Hausdorff distances (HD95) for talus segmentation on a CT image dataset. The three configurations performed well (DSC>0.95, HD95<0.6). When tested on the same samples, one of the configurations was more efficient while ensuring higher accuracy. We propose to focus on talus anatomic variations with increasing age based on this framework and apply it to clinical trials at the next stage.

Order-preserving submatrix (OPSM) is an important qualitative biclustering method of gene expression data. OPSM first sorts the expression values of gene and replaces them with corresponding column labels, and then mines the local patterns of the column label sequence set, where some rows rise and fall together under some columns. This paper proposes an order-preserving subsequence mining method (Charm_Seq) based on the Charm algorithm, and Charm_Seq makes full use of Charm’s efficient Itemset-Tidset prefix search tree to mine frequent closed patterns of column label sequence set. Meanwhile, Charm_Cla can effectively improve classification performance by restoring frequent closed sequences to training samples. Experiments were conducted on actual gene expression datasets, and the experimental results verified the efficiency and effectiveness of this method.

Aiming at the fact that the watershed algorithm appears very sensitive to various noises, such as being easy to cause over segmentation, pseudo-edge, edge discontinuity, and other phenomena, the paper explores an image edge extraction algorithm with improved watershed algorithm model. This improved model adopts the morphological method to compute the optimal threshold for maximizing the separation of foreground color and background color of the image and then applies this threshold to limit the path cost function of the algorithm model to narrow the search range and improve the execution speed of the algorithm so that the edge information of the image can be extracted more clearly. Comparative experimental studies show that the edges of the images extracted by the proposed algorithm are clearer. The results show that the improved algorithm model is more suitable for edge image extraction than other algorithms.

In order to address the high cost and limited robustness of continuous lens pose calculation in AR application systems, this paper proposes an AR registration method based on the SA-PSO algorithm. This method mainly combines the advantages of SA and PSO algorithms to fit the camera’s position and posture in a top-down manner. By continuously optimizing multiple particles in the feature space, the computational complexity of the AR registration process is reduced, and the effectiveness of registration time and robustness in practical application scenarios is ensured. The experimental results show that the proposed SA-PSO registration method can effectively achieve AR registration fusion in both manual marking and texture feature modes, and has a certain degree of robustness against occlusion.

In recent years, personalized paper grouping is a hot research topic in the field of intelligence education. In this paper, we propose an optimized deep knowledge tracking model Mul-MAKT combined with a genetic algorithm for personalized paper grouping with English practice as the target. Firstly, we use students’ records (exercise labels, response results, learning behaviors, exercise attributes) as input to the knowledge tracking model to predict students’ performance at the next moment and also to obtain students’ mastery level of knowledge points. Based on the students’ mastery levels, we adjust the students’ ability in each type of English exercise and calculate the students’ weaknesses in each type of question, so as to extract the questions from the test bank with the difficulty matching the students’ ability and containing the students’ weaknesses.

The ACLPSO algorithm, based on dimension specification intervals, adaptively tunes maximum velocity, inertia weight, acceleration coefficient, and learning probability for each dimension. It has demonstrated excellent performance in benchmark function tests involving both single-modal and multi-modal functions. However, to obtain the global optimal or approximate optimal solutions for all executed benchmark functions, it is necessary to manually set appropriate values for the maximum velocity coefficient s and the learning probability coefficient v during the operation of the ACLPSO algorithm. This study introduces an automatic approach to assign values to s and v, relying on function iteration count and test function fitness convergence. This enhancement enables the improved ACLPSO algorithm to directly derive the global optimal or approximate optimal solutions for all benchmark functions, eliminating the need for manual parameter tuning.

With the widespread use of big data and artificial intelligence technologies, the complexity of prediction problems such as classification, clustering, and regression is increasing, and the requirements for prediction models generally call for the fusion of different individual learners to achieve these goals. Although the prediction accuracy of the new fusion model can be improved to some extent by model fusion, the structure of the fusion model is more complex, the computationally intensive prediction time increases, and the reliability suffers. In this paper, we firstly systematically sort out the deep learning model fusion methods, secondly, analyze the coupling types of different model fusion methods and the impact on the reliability of the prediction system, and finally construct a fusion model for false information multimodal detection using model fusion methods for the needs of false information multimodal detection application scenarios and analyze its reliability impact.

Neural networks, as powerful tools for information processing, have been widely applied in various engineering fields. In this paper, we propose a neural network-based solution method that combines the ReLU function for ordinary differential equation (ODE) problems in scientific and engineering applications. The fundamental idea is to approximate the analytical solution of the initial value problem using the output of the neural network model proposed in this study. Through different examples, we demonstrate the accuracy and stability of this algorithm.

The domain generalisation pedestrian re-identification problem aims to generalise features learned in a known pedestrian data domain to an unknown pedestrian target domain. Most traditional domain generalization methods assume that the statistical properties between features and categories remain consistent across data domains. However, actual pedestrian data is often susceptible to domain factors such as lighting, colour shifts, and camera differences. This results in the data distribution of pedestrians under different domains not being identical and hinders the generalisation of the model. To address the above issues, this paper proposes a representation learning algorithm based on causal strong and weak alignment by constructing a structural causal model for the pedestrian domain generalization problem from the perspective of causal inference. The algorithm first performs causal intervention on the input pedestrian data to obtain causally enhanced images, then the image features are fed into the strong alignment module to achieve feature alignment in each dimension and obtain a preliminary invariant representation, finally, the features are then subjected to the constraints of the weak alignment module contrastive loss to further optimise the causal features under different cameras and improve the stability of the model’s cross-domain causal prediction. The method was compared and ablation experiments were carried out on the Market-1501 and DukeMTMC-reID datasets, demonstrating the effectiveness of the proposed method.

With the continuous development of information technology, more and more scholars use machine learning and integrated learning algorithms to analyze and predict the text emotion of online text reviews, explore the emotional tendencies of online users, which has far-reaching theoretical guiding significance and practical value for the promotion of online book sales decisions in the later stage. In this paper, user emotion is mined and analyzed based on online book reviews. The user emotion tendency is obtained through the emotion analysis method, which can help the e-commerce understand the user preferences and the quality of books on time, recommend books according to the user preferences, and play an auxiliary role in decision-making for consumers to purchase. Because of the current single method model and the inability to compare the accuracy of emotional prediction, this paper uses the public dataset Book_review on the official website of Kaggle, using NLTK (Nature Language TooKit), a natural language processing tool, to clean text data, build machine learning models and integrated learning models based on n-gram text features, obtain the results of emotional prediction through experimental analysis, and compare the machine learning model with the integrated learning model method, selecting the Logistic Regression model with the highest accuracy for emotional analysis of online book reviews, identify users’ emotional tendencies. Finally, we found that the accuracy of online comment emotion prediction based on the n-gram text feature Logistic Regression model is about 5% higher than the existing research methods.

Extensive progress has been made in point cloud semantic segmentation, yet the heavy preprocessing and complex design of network model still drag the training efficiency. In view of this, this paper proposes an optimized graph convolutional framework model PM-SPG based on superpoint fusion to efficiently perform point cloud. After the input cloud is pre-segmented into superpoints, a fusion algorithm is designed to merge the superpoints with similar geometric structures, which enhances the sparse representation of point clouds, and greatly improves the efficiency for extracting better local features. Meanwhile, a feature joint learning module is introduced to extract point cloud features in 3D and 2D perspectives, respectively by PointNet and multi-view feature embedding network, forming 3D-2D joint features to make full use of the local geometric information in the point cloud. The experimental evaluations on datasets S3DIS and Semantic3D show that the proposed framework significantly improves the model training efficiency while achieving competitive accuracy.

To effectively control the shield attitude to avoid excessive influence on the tunnel caused by forward deformation, serpentine shape, axis deviation and correction, it is necessary to effectively control the shield attitude. Based on BO-CatBoost machine learning algorithm, an intelligent prediction framework for shield construction propulsion attitude is proposed to maintain construction safety and quality. 23 factors affecting the shield attitude are selected as input variables. The CatBoost hyperparameters are optimized by Bayesian optimization (BO) algorithm and the importance of the influencing factors is evaluated by BO-CatBoost. And the attitude prediction model is established. The applicability and effectiveness of this method are verified by an engineering example: (1) The BO-CatBoost model has an excellent prediction effect for shield attitude. The goodness of fit 𝑅² is all above 0.9, and the values of 𝑅𝑀𝑆𝐸 and 𝑀𝐴𝐸 are small. (2) The BO-CatBoost algorithm can effectively identify important construction parameters and provide guarantee for the safety control of shield construction.

Distributed measurement of underground rock and soil deformation is an important part of geological hazard monitoring. Using parallel helical transmission lines as sensors and combining frequency domain reflection (FDR) analysis enables continuous distributed measurement of larger deformation quantities. It is known that the characteristic impedance of the parallel helical transmission line increases with the increase of tensile force. Through FDR, the measurement and location of the tensile point can be determined. This paper explains the principle of FDR measurement of tensile deformation in parallel helical transmission lines. In response to practical application issues of FDR in parallel helical transmission lines, this paper analyzes the influence of sweep range, frequency, and number of sampling points on measurement results through experimental data. The paper also measures different incident signals under different circumstances, identifies and analyzes factors that may affect measurement results, and selects a better measurement scheme. The proposed method is used to simulate the actual use of parallel helical transmission lines and obtain actual measurement data, which is analyzed for better understanding.

The detection of cigar appearance defects is a critical step in the cigar production process and holds significant importance in ensuring the quality of cigar production. To address the issues of inefficient and unstable manual detection, this paper employed deep learning-based object detection models for cigar appearance defect detection and proposed an algorithm for cigar appearance detection based on YOLOv5. First, the cigar appearance defect acquisition equipment was constructed, the defect image data were collected, and the experimental dataset was established. Then, the deformable convolution was introduced to enhance the learning capability of the backbone network. Furthermore, the Bi-directional Feature Pyramid Network (BiFPN) was employed to improve the feature information of each layer. Lastly, the spatial context pyramid (SCP) was utilized to enable global spatial context learning within the feature layers, further enhancing the features. The model performance was evaluated by mean average precision (mAP). Experimental results demonstrated that the improved YOLOv5 achieved a mAP of 90.7% for cigar appearance defect detection and a detection speed of 10.6ms per image, showcasing excellent detection accuracy and speed. Moreover, this model exhibited significant improvement in detecting small defects and defects located at the edges. Therefore, the improved YOLOv5 model satisfied the requirements for automatic cigar appearance defect detection.

The distribution of emergency supplies after emergent public events is an important measure to reduce the losses. Vehicle routing for emergency supplies distribution has become an important research problem. Related studies mainly focus on general situations, and few studies focus on the vehicle routing problem for scarce emergency supplies distribution within a city. Vehicle routing problem within a city has its particularities, among which is city road network. The general road network used in relevant studies only contains the distribution center and demand points, which is different from the city road network, leading to inadequate applicability. This paper combines the vehicle routing problem for scarce emergency supplies distribution with the city road network. Aiming at minimizing the maximum loss of shortage at a single demand point, a vehicle routing model for scarce emergency supplies distribution within a city is established, and solution algorithms are designed according to different cases. The aim is to provide theoretical basis for emergency management departments to make decision on emergency supplies distribution scheme.

With the ever-growing number of vehicles in our world, enhancing fuel efficiency has become a paramount concern for gas stations. Traditional manual refueling process, which typically takes 5 to 6 minutes, or even longer during peak hours, no longer meets the demands of today’s fast-paced life. In light of this challenge, this paper proposes a 2D-3D fusion adaptive visual perception algorithm named iterative measurement using automatic exposure point cloud imaging (IM-AEPCI) to obtain the vehicle position. Specifically, we first construct an adaptive exposure prediction network (AEP-Net) based on deep convolutional network and attention mechanism to generate highly reliable 3D point clouds, and then propose the iterative-based point cloud registration algorithm to obtain 6D pose measurement. Experimentally, the maximum error of our method in the three directions of XYZ does not exceed 1.8mm under different lighting conditions, even with difficult measurement highly reflective vehicles, which proves that IM-AEPCI can greatly improve the accuracy and robustness of 6D pose measurement for robotic arm. Besides, our entire visual measurement process takes only 5s, which results in the entire robotic arm refueling system only takes 2-3 minutes to complete. Our research makes it possible to implement a non-contact refueling system, which prominently improves refueling efficiency and safety, and greatly saves labor costs.

Acupoint potential signal is a kind of bioelectric signal collected on the skin surface, which has the characteristics of weak signal, strong noise and strong randomness. These characteristics make it difficult to extract acupoint features, which further affects the accuracy of acupoint classification. This paper proposes an acupoint classification method combining signal processing with intelligent algorithm. Firstly, the signal is decomposed by CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise), and the decomposed IMFs (Intrinsic Mode Functions are obtained). Secondly, the filtered modal set is taken as input, and the features are extracted by using Convolutional Neural Networks (CNN) to get its deeper features. Finally, the obtained feature parameters are input into the support vector machine (SVM) optimized DBO:(Dung Beetle Optimizer) for classification. The experimental results show that CEEEMDAN-CNN-DBO-SVM model can effectively identify the types of acupoints, with an average accuracy rate of 93.01% at rest and 90.6% at click stimulation. The effect is better than CNN, SVM, CEEMDAN-CNN, CEEMDAN-SVM, CNN-SVM and other five classification methods.

Most regularization-based continual learning methods are based on synapse importance. In contrast to them, we propose a neuron importance algorithm that uses Taylor criterion to calculate the importance of neurons. Then, for the fully connected layer and the convolutional layer, we propose two different approaches to convert neuron importance to synapse importance. Compared with existing neuron importance based method, the proposed algorithm is simple to implement, requiring only one forward propagation and one backward propagation to calculate the neuron importance. The effectiveness of the algorithm was validated on two datasets, 5-Split-MNIST and 5-Split-CIFAR10.

This paper analyzes and studies the problem of large consumption of computing and bandwidth resources in the application of the security onboard communication scheme in the vehicle-mounted embedded system, proposes a possible solution, and uses software encryption in the vehicle-mounted embedded ECU to verify and test the proposed security protection scheme, and compares the performance of applying several commonly used encryption algorithms to secure the bus communication data, At last, the paper gives a relatively more suitable protection scheme for the classical CAN and CAN-FD communication.

One of the advanced algorithms for image foreground segmentation is DIM (Deep Image Matting), which applies convolutional neural networks. However, for complex foreground shapes, fixed convolutional kernels are not the optimal approach. Therefore, Deformable Convolution Networks (DCN) were introduced as a replacement, which partially address the issue of losing edge details in deep layers. However, DCN suffers from the extraction of undesired features, significantly reducing the model’s flexibility. To address this, the bias is reshaped using a balanced refining bias with an adaptive weight matrix, instead of a fixed 9-grid bias. To evaluate the effectiveness of the bias, a simple model called DIM_RDC is designed and trained, which utilizes the balanced refining bias. The model is tested on the Adobe dataset, and the results show that DIM_RDC improves the evaluation metrics of SAD by 1.62% and MSE by 11.05% compared to DIM. Therefore, DIM_RDC demonstrates a certain level of competitiveness over DIM.

Automatic valet parking is a popular service provided by many car parks, which allows customers to park their cars without having to physically handle the parking process. In order to ensure the safety and efficiency of automatic valet parking, it is necessary to generate a visual feature map that reflects the current scene status. The visual feature map can be used to identify potential hazards and ensure that the car park is free of obstacles. The construction of visual feature maps is generally obtained offline, and can be acquired and processed after collecting data from professional survey vehicles equipped with multiple sensors including laser, vision, and IMU. In this paper, we propose an automated valet parking scene visual feature map generation algorithm based on LiDAR SLAM. The key steps include: calibration of visual and laser systems, generation of trajectories by fusing laser data with IMU measurements, and generation of visual feature map based on laser trajectory and visual images. The algorithm integrates the advantages of both LiDAR SLAM and visual feature mapping, and achieves high accuracy and efficiency in map generation.

Traffic flow prediction has a good guiding effect on traffic control. In response to the current inability of road traffic flow prediction methods to fully reveal the inherent laws of traffic flow, and considering the issue of fully considering spatiotemporal correlation in traffic flow prediction, this paper proposes an LSTM (Long Short-Term Memory) model based on Bayesian optimization. Experimental studies have shown that the LSTM model based on Bayesian optimization has good performance and high prediction accuracy.

Aiming at the localization and composition of mobile robots in unknown environments, this paper briefly summarizes the connotation of SLAM, analyzes the principle of extended Kalman filter theory, which is an algorithm process of gradually approaching the real state of the system through recursive estimation of the state of nonlinear random dynamic system, and finally verifies the relevant theories and algorithms through two-dimensional EKF vSLAM simulation experiments. The experimental results show that the EKF-vSLAM algorithm can effectively improve the accuracy of robot positioning and landmark positioning in two-dimensional environments, reduce the uncertainty of mobile robot positioning and composition in unknown environments, and to some extent improve the autonomy of mobile robots.

It is shown how a dynamic road network conversion method can be used to address the shortest reliable path problem. Firstly, the impact of stochastic time-varying characteristics on commute time in urban traffic systems is analyzed. Second, the function of waiting time at signalized intersection is proposed and the applicable conditions of the function in different scenarios are analyzed. Afterward, by using the first mathematical induction, it is proved that stochastic time-varying networks are able to be transformed to deterministic time-varying traffic networks. Finally, a numerical test is performed to prove that the proposed method is feasible. The test results show that the method can search for the shortest reliable path based on historical transportation data and personal travel experiences without obtaining the complicated probability distribution of travel time.

Social recommendation algorithms that integrate social information have become a research hotspot. However, the existing social recommendation has the problem that it is difficult to obtain reliable and clear friend relationships. Therefore, this paper proposes a social friend selection method FSISB based on interests and social behaviors. This method first defines four types of scenarios to find the explicit/hidden social friends of the target user, and alleviates the sparsity problem of social data by expanding the social scope; then by combining the user’s social structure and long-term/short-term interests to calculate the similarity of social friends, gets reliable and unambiguous social information and complete friend selection. Experiments show that the baseline model integrated with the FSISB method improves the recommendation performance by 0.4%-1.5% compared with the original method, and the performance improvement is more obvious in datasets with sparse social data, which verifies the positive impact of the FSISB method on alleviating data sparseness.

In this paper, the definition of a generalized fuzzy k-pseudo metric and some typical examples are given. Then we point out that a generalized fuzzy k-pseudo metric can be constructed from a nest of k-pseudo metrics. Besides, we verify that a nest of k-pseudo metrics can also be constructed from a generalized fuzzy k-pseudo metric. Therefore, it can be shown that there exists a one-to-one correspondence between generalized fuzzy k-pseudo metrics and a nest of k-pseudo metrics.

Memristor-based stateful logic is considered to be one of the effective techniques to realize in-memory computing and break through the bottleneck of von Neumann computing structure. Up to now, many methods have been proposed to implement stateful logic, such as implication (IMPLY) and anti-parallel bipolar memristor (APBM). Based on the circuit structures of these two methods, we replace memristors in the original circuits with ternary memristors to obtain a ternary memristor-based IMPLY circuit structure and a ternary memristor-based APBM circuit structure. On this basis, we analyze in detail the operation voltages and constraints required to implement the ternary logic operations NOT-implication (TNIMP) and ternary AND (TAND), respectively, using these two circuits. In addition, we use the logic operations TNIMP and TAND to design a TNIMP logic circuit without loss of input, and demonstrate it by simulation.

As industrial manufacturing booms and data continues to grow in the IoT era, the number of edge computing nodes needed is growing. However, the resources of these edge computing nodes are often constrained. Under the conditions that the storage space of edge computing nodes is insufficient and the network environment is more stable, we propose a lightweight edge container image storage method with cloud-edge collaboration. This approach leverages the CernVM-FS (Cern Virtual Machine File System) to enable edge computing nodes to remotely mount the container root filesystem. In the experimental environment set up in this paper, the total storage space consumption of the conventional mounting method (local filesystem mount) is 2.67 times higher than that of the proposed method (remote filesystem mount). In addition, by measuring and comparing the time overhead of these two methods in the phases of pulling images, creating containers, and running containers, we believe that this remote mounting method has some applicability in the edge application scenario of this paper.

In order to better guarantee different quality of service requirements for different types of 5G network traffic, the characteristics should be studied and modeled for each type of traffic carried via 5G network. In this paper, we report a measurement-based study, which includes 5G network traffic data collection, analysis and modeling. To better extract the behavioral characteristics of 5G network traffic, we conduct the collection of traffic data from the 5G new radio (NR) system based on OpenAirInterface (OAI) platform and universal software radio peripheral (USRP). In this paper, three types of traffic carried via 5G network are collected and analyzed. For basic traffic characteristics, the traffic rates and the probability density distribution of packet length and packet interval are analyzed. For traffic modeling, the concept of cumulative arrival process is adopted and the arrival curve model is derived based on network calculus.

Inverse synthetic aperture radar (ISAR) image quality assessment is a prerequisite and key for the development and application of ISAR images. In order to solve the problem of low correspondence between objective evaluation indexes and subjective feelings, this paper proposes an algorithm model for evaluation based on a convolutional neural network. Based on the input of the original image, this paper further combines eye-tracking-based thermal maps to construct a dual channel input residual network evaluation algorithm, which improves the evaluation ability of ISAR image quality. The rationality and effectiveness of this method have been proven through experimental testing.

Lane-line detection is one of the key technologies in the field of intelligent driving. This paper changes the anchor box regression loss function of the YOLOv5s to the EIoU function to construct a kind of lane-line detection model. The experimental results show that the improved detection model is more accurate in lane-line detection. The improved model improved mAP by about 11.1% compared to the conventional YOLOv5 model. At the same time, it can solve the problems of traditional models such as low real-time performance and underfitting to curved lane lines.

During the coal mine safety production process, a significant amount of text data containing information about coal mine safety hazards, such as working face, hazard location, hazard subject, and hazard problem description, is accumulated. The extraction of named entities from coal mine safety hazard text serves as the foundation for conducting a study on the early detection of coal mine safety hazards. Since the single modeling strategy is being used, the current named entity recognition (NER) model technique has a low recognition precision and ratio. Firstly, a character-level encoding of coal mine safety hazard text is performed by a BERT pre-training language model to generate word vectors based on contextual information, followed by local and global deep feature extraction of coal mine safety hazard word vectors by a convolutional neural network (CNN) with multi-layer bi-directional gated recurrent neural networks (BiGRUs), and finally decoding by conditional random fields (CRF) to generate global optimal label. On tasks of the NER for coal mine safety hazards, by comparing and analyzing with the mainstream deep learning entity recognition models, As shown by the outcomes that the precision of the NER model for coal mine safety hazards proposed in this paper reaches 91.74%, the recall reaches 93.20%, and the 𝐹1-measure reaches 92.45%, which shows a better performance. The NER task of precisely obtaining key information such as hazard location and hazard subject from unstructured coal mine safety hazards text data is achieved, which provides important information for hazard investigation and management.

In this research, we present a novel approach for relation extraction using the multiple kernel support vector machine model. The aim is to improve the comprehension of Chinese Instructions by family service robots. Our approach focuses on extracting the people-item relation from Chinese instructions. We start by defining four categories of people-item relations: sequential, belong to, equivalent, and direction. Next, we construct a feature combination for the entity using lexical, phrase, order, and property features. We then generate multiple kernel functions using a weighted sum method and selected foundation kernel functions (lexical, syntactic, and property). The multiple kernel support vector machine model is constructed using a simple multiple kernel learning technique. The experimental findings validate that our proposed approach outperforms current methodologies in terms of accuracy, retrieval rate, and F-measure.

In the process of oil field production, wax deposition in oil well will lead to lower pumping efficiency and even wax sticking in well. In response to the problem of low accuracy in determining the flushing cycle of oil well wax removal and prevention work relying on manual experience, the degree of wax deposition in the oil well is obtained by analyzing the indicator diagram of the oil well working conditions and using the load difference. The Pearson correlation coefficient is used to analyze the correlation of oil well parameters and establish a dataset of oil well load. On this basis, the oil well wax deposition prediction model is constructed by using the Long Short-Term Memory network, and the optimal parameters matching the network structure are searched by Particle Swarm Optimization algorithm. The prediction model is utilized to train the sample dataset, obtain the prediction data and realize the pre-warning of oil well wax deposition. Finally, the least square curve fitting method is used to obtain the growth rate of load difference and the wax deposition state of the well. The experimental results indicate that the constructed oil well wax deposition prediction model can obtain the wax deposition status in advance, guide the wax removal and prevention work, and provide decision support for the formulation of the well flushing cycle in the oil field.

This work introduced an AI-based simulation method as an innovative solution to address the limitations of conventional numerical reservoir simulation methods. The primary motivations behind this research were the time-consuming nature of finite difference (FD) simulation and the considerably lower accuracy of streamline (SL) simulation. We adopted a novel approach to overcome these challenges by incorporating erroneous yet rapidly supplied streamline data as inputs to our model. This unique input selection allowed us to leverage the direct training strategy in a deep learning-based network. Furthermore, by integrating FD grid data as training outputs, we improved the model’s accuracy and expanded its functionalities. The developed proxy model demonstrates remarkable performance, surpassing the FD simulator’s computational efficiency while outperforming the SL simulator ’s accuracy. Through validation of the benchmark Egg model, our research confirmed the potential of this AI-based proxy model as a reliable, accurate, and efficient alternative in dynamic reservoir modeling.

In view of the problems in the field of non-coal mine, such as high pressure of perception data acquisition and transmission, cloud analysis and processing, insufficient ability of active safety risk analysis, and poor timeliness, this paper discusses the necessity of building a non-coal mine safety risk monitoring and early warning system using edge computing technology against the requirements of major disaster risk prevention and control construction in non-coal mine. It analyzes the technical requirements of edge computing for data security and data intelligent analysis of non-coal mine safety risks, propose the edge computing architecture of non-coal mine safety risk. Designed intelligent risk analysis equipment for non-coal mine safety risks, implementing functions such as data collection and aggregation, data association analysis, video intelligent analysis, and data security control. This research improved the level of intelligent warning on the edge of non-coal mine safety risks, with great significance in preventing and extricating major safety production accidents in noncoal mining enterprises.

In the Internet era, network security faces challenges from various attacks like data modification, compromised applications, and DoS. To tackle abnormal traffic caused by malicious programs, we propose an improved Convolutional Neural Network (CNN) learning model for classifying traffic as normal or attack. Our model enhances feature extraction in both width and depth, resulting in improved accuracy and robustness of the network. Compared to commonly used detection models, our approach achieves an average accuracy improvement of 4.55%. Ablation experiments also show that employing our IDA-CNN-based intrusion classification detection model reduces the average error rate compared to using all feature subsets.

With the exponent risen of intelligent devices, many applications should process large scale data and provide real-time services with sufficient computing resources and high energy consumption. Traditional cloud computing may bring the great burden to the core network, and cannot meet the requirements of low latency, security, and high quality of service. Thus, edge computing can provide services, with the purpose of reducing latency, saving bandwidth resources, improving efficiency and high quality of service for users. Due to the limitation of the service coverage for each edge servers. When the vehicles move at the road, it may exceed the service range at the edge. Therefore, service migration stratagem should be designed carefully to guarantee the continuity of the service. In sight of the studies of service migration, the mobility and the trajectory are the important factor for service migration, which may reflect the consumption of the service migration. Thus, we propose a trajectory-aware service migration approach with deep reinforcement learning. In this paper, the vehicle trajectory is predicted by the deep spatiotemporal residual network model, and then a service migration algorithm based on Deep reinforcement learning is proposed according to the prediction results. The experimental results show that our algorithm can achieve lower latency comparing with other algorithms.

The Cross-Project Effort-Aware Defect Prediction (CPEADP) model can effectively use detection resources and the data from different projects to build models. One factor affecting the performance of CPEADP is the problem of data distribution differences in cross-project settings. The classification performance also greatly impacts the predictive capabilities of the models. Therefore, we propose the BDA-DF model and conduct experiments on 11 cross-project datasets from the PROMISE repository. Compared to traditional data filtering and transfer learning methods, our approach exhibits significant improvements across five effort-aware metrics, including Precision@20%, Recall@20%, F1@20%, PofB@20%, and IFA. To explore the optimal classifier for CPEADP, we embed seven different classifiers into BDA. The experimental results on the BDA embedded with different classifiers reveal that DF exhibits the best overall performance.

Brain-computer interface (BCI) provides a direct link for the interaction between the human brain and external devices. In the meanwhile, virtual reality (VR) offers people an immersive virtual experience with simulated environments. By combining these two technologies, integrated solutions have been raised, but most of the current solutions are only applicable in the laboratory setting. Therefore, in this study, we proposed an integrated BCI-VR system for daily use based on dry electrodes. A game for motor imagery was also designed in the system through the classification of motor imagery signals. The proposed system shows potential as a practical platform in rehabilitation and other healthcare fields.

Attribute-based access control models are widely used in permission management for resource access. By mining access control lists of policies, it can significantly reduce the cost of policy management and streamline the composition of policy rules. However, as resources increase, access policies will become complex. Uncontrolled attribute will lead to policy conflicts and thus policy mining will no longer be reliable. To address this issue, we propose a dynamic access control model based on attribute reachability. Firstly, we analysed the accessibility of attributes to ensure the reliability of authorised attributes. Secondly, we propose a multi-dimensional attribute management mechanism based on precondition-limited attributes, which enables permission passing and inheritance. On the basis of reachability access policy, we finally achieved secure policy mining changes by traversing user permission relationship tuples and constructing candidate rule seeds.

The bike-sharing scheme has become extremely prevalent in large countries around the world. The appearance of it raises a lot of advantages such as reducing greenhouse gas emissions, alleviating traffic congestion, especially implicitly increasing exercise and enhancing health. However, the management of shared bicycles scattered everywhere in the city has become a serious problem. Placing enough bicycles at a certain time in high-demand places can maximize the utilization of bicycles and improve the convenience of people. So the forecast for bike-sharing demand is quite necessary to improve the distribution of bicycles which ensures enough bicycles for the public all the time. In this paper, we study the prediction of bike-sharing demand in London using multiple linear regression and random forest methods based on historical rental bicycle data. We analyze the descriptive statistics and conduct feature engineering using rich relevant factors. The experimental results demonstrate that the random forest model achieves a superb performance with an R-squared value of 0.95 on the test set. This research can be applied in bicycle management to increase bicycle utilization and improve convenience.

A study has been conducted on the effectiveness assessment methods of high-orbit infrared remote sensing satellites, and by analyzing the difficulties in evaluating the effectiveness of the complex star-ground system of the “human-in-the-loop”, an assessment criterion centered on the degree of mission accomplishment has been proposed. Drawing on the concept of equipment technology maturity and grading, and combining practical experience, an efficiency assessment model has been established around the construction of mission scenarios of different grades, and a class of assessment processes and assessment methods have been developed and provided as samples. A calculation method based on system reliability, repairability and availability is proposed for the comprehensive assessment of mission completion, which can improve the current assessment method in practical application. A generic quantitative method for evaluating the effectiveness of the system based on the unfulfilled or unintended tasks is investigated, which can effectively reflect the bottleneck capacity of the system and support the subsequent optimization and improvement of the system.

Beach erosion monitoring is crucial for coastal protection and beach maintenance. Field survey is hard to collect longterm beach change information due to high cost and time consuming. And it is difficult to observe three-dimensional deformation information on beaches by the optical remote sensing that is likely to be influenced by bad weather and few feature points on beach. Therefore, this paper employs the time series radar images of Sentinel-1A satellite to monitor the morphological changes of the survey area of Qingshan Bay beach in Quanzhou City, Fujian Province, China. The Differential Interferometric Synthetic Aperture Radar (DInSAR) technology is utilized to obtain the deformation information of the research area. The time-phased cumulative deformation analysis is conducted for different beach locations. Moreover, the high spatial resolution drone images and beach soil moisture data are employed to analyze the results of erosion monitoring. The current research results indicate that: (1) The results obtained by DInSAR technology can achieve the observation accuracy of centimeter level in the beach area; (2) The quality of monitoring results based on the DInSAR technology is related to beach humidity.

To explore the influence factors of different step practices on the human balance ability, and provide data support and theoretical basis for the practice of Tai Chi and Baduanjin. Sports biomechanics parameters of the lunge step support period of Tai Chi group, horse step support period of Baduanjin group and walking gait support period of the beginner group were compared, scanning and collecting knee CT data from the subjects of lunge step test, horse step test and walking gait test, the tibia model was treated with Geomagic Studio, the right tibia of the musculoskeletal model was replaced by mirror image, adding the net force of the knee joint derived from AnyBody to Ansys, after assigning values to materials, setting boundary conditions and adding loads, the equivalent stresses of lunge step, horse step and upright step were analyzed. Compared with the walking gait, in lunge step, the hip flexion angle and knee flexion angle were significantly larger (P<0.01), the hip abduction torque, hip rotation torque and ankle plantar flexion torque were significantly larger (P<0.01); in horse step, the hip abduction angle, hip flexion angle and knee flexion angle were significantly larger (P<0.01). Lunge step exercise is conducive to improving the deep anterior thigh muscle strength, horse step exercise is conducive to improving the deep medial hip muscle strength and shallow posterior calf muscle strength, walking step is conducive to improving the shallow anterior calf muscle strength and ankle dorsal flexion muscle strength.

An improved method based on binocular vision ranging is proposed for a certain monocular vision position detection model with complex structure, large limitations and low accuracy. Firstly, the internal and external parameters of the binocular camera are obtained by using the Matlab camera calibration tool, and then the obtained parameters are input into the compiled program for stereo correction of the images acquired by the binocular camera, and finally the BM stereo matching algorithm is used to generate the parallax map for binocular ranging, and the distance information of the target and the 3D coordinates of the feature points are obtained. The experimental results show that the ranging method has high accuracy and low error.

Hyperspectral remote sensing images (HRSI) contain rich spectral information and spatial structure information which can be effectively used to classify ground objects. In this paper, we propose a feature construction method that combines spectral features, Weber Local Descriptor (WLD) texture features and Gabor texture features. To start with, the spectral information of each pixel is extracted as the basic feature of this pixel. Followed by the local WLD texture feature of each pixel is extracted. Considering the limited expression ability of a single local texture feature, hence the Gabor filtering is carried out on each band respectively to form the Gabor feature map. Singular value decomposition (SVD) is adopted in the local neighborhood, and the singular value is used as the Gabor texture feature of the local neighborhood. Finally, the spectral feature, WLD texture feature and Gabor texture feature are integrated together as the final feature of the pixel, and XGBoost classifier is used for training and classification. Conducted experiments validate that the combination of these spectral features has better expression ability and classification performance.

Labanotation is an approach to record human movement with different distributions in space that have different representations of symbols. Motion capture is a technique that records the trajectory of an object in the time domain by tracking the motion of key points, observed using different sensors, and recorded in files. This paper is to manually mark the Labanotation based on 3D human motion capture data, to divide the 3D human motion into gaits and poses, and to record and represent different human movements and positions in space in the notation. This is an exploration of motion capture data to generate Labanotation, while also contributing to the preservation of folk-art forms.

For a long time, blind or visually impaired people have relied on traditional aids such as canes or guide dogs to assist them in walking, but these aids have significant limitations. In order to help visually impaired individuals walk independently, this article improves the method of mapping images to sound in the literature, making it more applicable to blind navigation. By limiting the range and size of captured images, as well as improving the mapping relationship between captured images and transformed sound, the accuracy of the method has been improved from 84.2% to 88.8%, an increase of nearly 5% compared to the method in the literature, which can better assist in guiding the blind.

Mammography is main tests used for breast cancer risk assessment. However, the mass segmentation and classification of mammograms are extremely challenging. To reduce computational costs and the workloads of radiologists, deep neural networks have been widely used in various medical image segmentation and classification tasks. Since there are some common features between these two tasks, a multi-task learning approach to solve both tasks are a promising direction. We propose a mix Transformer depth-wise separable convolution U-Network (MTDUNet) for mass segmentation of mammograms. We introduce depth-wise separable convolutions to replace traditional convolutions and improve the network’s perception of multi-scale features within the receptive field. Additionally, due to the inherent limitations of convolutional networks, we introduce mix transformer to model remote contextual information. We conducted evaluations the proposed GATNet on two publicly available breast mass segmentation datasets. The average Dice similarity coefficients between the MTDUNet results and INBreast and CBIS-DDSM data were 89.90% and 83.63%, respectively. The experimental results indicate that MTDUNet can significantly reduce the spatial complexity of medical image segmentation networks and effectively save computational resources.

Traditional person re-identification (ReID) methods based on generative adversarial networks often neglect the impact of background clutter during the image generation stage of data augmentation, which can have a negative effect on the accuracy of subsequent recognition results. To address this problem, this paper proposes a mask-guided generation learning loss optimization method, which consists of reconstructed image loss optimization and composite image loss optimization. By optimizing the loss function in the image generation stage, the proposed method generates more real images of the person’s body part to eliminate the impact of background clutter on recognition results in subsequent stages. The reconstructed image loss optimization uses a body mask to construct an attention transfer matrix, mapping the original image into a new feature map containing person-related but background-independent features. The composite image loss optimization uses VGG16 to extract fine-grained features from the composite image and calculates the local loss so that the generation network can further learn the fine-grained features of the body part. Extensive experiments on the Market- 1501 dataset show that the proposed method achieves 94.8% in Rank-1 and 85.5% in mAP, increasing the baseline results by 0.4% and 0.5%, respectively. The proposed method significantly improves the authenticity of the person’s body part in the generated image and the accuracy of ReID, providing high-quality training data for the model to discover more finegrained features. The proposed method provides a new idea for the application of body masks in ReID as well as a feasible technical route for future research.

How to transmit better the color distribution of one image to another image is an interesting problem in computer image generation, and the optimal transmission mathematical theory has been applied to the research field of this type of problem. In response to the common problems of color distribution differences or lack of hierarchy between the transmitted result image and the reference image in existing methods, this paper proposes a new color transfer method: first, we segment the foreground and background regions of the reference image and the specify image; Then we use the optimal transmission theory to transfer the colors of the foreground and background regions separately; Finally, we merge and enhance the transmitted regions to obtain the resulting image. The experimental results show that compared to the two image color transfer methods in the literature, the method proposed in this paper can better transfer the color distribution of reference images to specify image, preserve the hierarchical sense of the original specify image, and achieve better visual effects.

The classification and grading of flue-cured tobacco leaf is an essential step in tobacco production. In order to carry out relevant research, a new flue-cured tobacco leaf image database (FTIDB) is established to provide high-quality and annotated flue-cured tobacco leaf image data. At present, the database consists of 2,113 flue-cured tobacco leaf images from major production regions across China. To ensure the quality of the database, a series of procedures are implemented. Firstly, the image acquisition system was constructed and the acquisition environment was calibrated. Secondly, Image pre-processing was adopted to improve the image quality. Thirdly, according to the characteristics and application requirements of tobacco leaf image, the image quality was objectively evaluated through quantitative metrics in terms of color, texture and greyscale. Fourthly, XML file was used to organize annotated information from tobacco experts. Finally, the image data dictionary was created to complete the data storage management using Microsoft Structured Query Language Server. This database can provide basic tobacco image resources for researchers and agricultural technicians, so that extensive studies can be performed and the automatic level of tobacco classification and grading will increase effectively.

There are numerous practical applications for animal age estimation, such as age simulation, population structure estimation, wildlife conservation, biometrics, and more. Most of the current methods either concentrate solely on animal detection or animal age estimation. However, animal detection provides limited information, which is not adequate for more sophisticated computer vision tasks. On the other hand, existing methods for estimating animal age are based primarily on facial images, but capturing an accurate image of an animal’s face in a wild environment is extremely difficult and impractical for animals without discriminative facial features. In this paper, we propose a novel task, called Animal Age Group Detection. This task is a combination of animal detection and animal age estimation, with the primary goal of obtaining more information beyond the location of the animal by also providing crude age information, which is simple to accomplish with computer vision methods in a variety of open-world applications. To facilitate the study of this challenging, we present a benchmark for Animal Age Group Detection (AAGD), which contains a diverse range of 22 animal types, comprising a total of 4,778 images. These images are well-annotated with axis-aligned bounding boxes and include information about the age groups of the animals. In addition, we establish a baseline detector for this task, called AG-YOLOF. The release of AAGD aims to facilitate future research on the identification of animal age groups and enhance awareness of this significant area of study.

We present a 3D point cloud semantic segmentation method for power transmission line based on PointNet++. The method aims to identify pylons, ground, vegetation, insulators, earth wires, conductors, bypass jumpers, etc. We created a dataset of 54,733,560 points, collected from real power transmission lines by airborne lidars. We modified the feature extraction network of PointNet++, the radius, and the number of samplings to fit the complicated structures of the power transmission system components. The experimental results showed accuracy improvement in all 8 categories, the highest improvement is with the insulators, by 2.73%, and the overall improvement is by 0.72%, compared to the original model.

In order to significantly improve the performance of resonant temperature sensor the mode localization phenomenon is first applied to the resonant temperature sensor and a resonate temperature sensor composed of 2-DoF (degree-of-freedom) resonator is proposed in this paper. The sensing scheme and 2-DoF resonator of the sensor were analyzed theoretically. The temperature response and model shape were analyzed by finite software. The simulation results showed the AR (amplitude ratio) based sensitivity was enhanced by four orders of magnitude compared with the frequency-based sensitivity.

This paper presents the development of a wireless detector specifically designed for weathering test chambers, which are employed to assess the effects of solar radiation on various materials. Accurate evaluation of internal irradiance parameters within these chambers is critical for precise analysis. However, the absence of detection windows in many chambers poses a challenge to the measurement process. To address this limitation, a wireless detector is introduced, overcoming structural and operational constraints. Existing wired radiometers are unsuitable for the high-temperature conditions inside the chambers, while the few available wireless options fail to meet measurement requirements and stability criteria. The proposed wireless calibration apparatus incorporates national calibration standards and wireless transmission technology to enable real-time measurement of irradiance. It comprises a wireless calibration module, a handheld terminal, and a PC terminal. Finite element simulation analysis is utilized to investigate temperature rise and distribution within the aging test chamber, thereby informing the design of temperature control systems. Furthermore, this paper explores the shell design of the wireless calibration device, emphasizing circuit board protection. Overall, the proposed wireless detector and calibration apparatus overcome the limitations associated with existing wired and wireless radiometers, offering a more accessible and accurate solution for weathering test chambers.

With the increasing use of robots in various industries and daily life, researchers have been focusing on studying this area. In the chemical industry, where there are potential risks, using robots instead of manual labor in hazardous environments can effectively decrease the chances of accidents. However, there are still challenges in terms of low efficiency, poor accuracy, and incorrect positioning in robot sorting operations. To tackle these issues, a method based on the visual perception of information has been suggested. A prototype of a sorting robot verification experiment platform has been developed, which achieves precise sorting through graphic recognition and positioning of parcels. The experiment for robot parcel sorting has been conducted, and the results are promising. The adaptive recognition rate for miscellaneous feature package images is 96.31%, with an average time of 0.12 seconds per image. These outcomes demonstrate the successful implementation of robot parcel sorting. Overall, the use of robots in hazardous environments significantly reduces the risk of accidents. The proposed method based on visual information perception has shown promising results in enhancing the efficiency, accuracy, and positioning of robot sorting operations.

With the continuous development of technology, pico-projectors are widely used as a portable display device in business presentations and home entertainment. However, due to their miniaturized design and high-power operation, pico-projectors commonly suffer from heat dissipation problems, which directly affect their performance and lifetime. In this paper, we analyze the working principle and heat dissipation mechanism of pico-projectors, deeply investigate the causes of heat dissipation problems, propose two heat sink solutions based on numerical analysis of existing heat sink structures, verify the effectiveness of these improvement strategies through experimental design and result analysis, and provide an outlook on future research directions.

The garbage cleaning methods in small waters such as parks and aquaculture waters are mainly manual cleaning, which has problems such as time-consuming, low efficiency, and high labor costs. Therefore, according to the needs of cleaning work and multi-turn optimal path planning, this paper proposes a low-cost, high-efficiency water surface garbage cleaning robot design scheme for small water areas, and adopts 4-degree-of-freedom mechanical arm actuator, pontoon propeller walking mode. It realizes the combination of robotic arm type and conveyor belt type, garbage disposal device and power unit, etc. And the breadth-first search algorithm is used to plan the path of the robot.

The research and design of mobile mechanisms that can adapt to a variety of complex terrains is a key research direction in the field of mobile device research. Because having both good mobility and barrier crossing is the primary performance indicator for mobile agencies to quickly adapt to the unstructured environment, the combination of strong mobility of the wheeled walking mechanism and high barrier crossing of the legged walking mechanism of the composite deformed wheel walking structure is generally favored. In this paper, in view of the demand for mobile equipment walking structure in complex environment, comparing the advantages and disadvantages of domestic and foreign composite deformed wheel moving mechanism, analyzing the characteristics of each walking mechanism and its design points, and designing a composite deformed wheel combining wheeled walking and tracked walking; The action principle of the deformed wheel is studied, and the spokes of the wheel are expanded by the motor-driven hydraulic cylinder to realize the deformation; The working conditions after deformation of the circular wheel body are analyzed, and it is proved that the structure can accomplish the deformation under different working conditions, and the fields to which the composite mobile structure is applicable are analyzed, which can provide reference for the subsequent research of related walking structures.

This paper presents a comprehensive calibration system, including intrinsic and extrinsic parameters of the camera, hand-eye, and Tool Center Point (TCP) calibration. To improve efficiency and simplify the process, the system only requires two steps to calibrate four parameters and provides errors for each parameter. The entire process does not require any other devices except for one calibration board. This paper suggests using 3D distance as the error for stereo camera calibration and provides suggestions for capturing 15 different angles of images to cover the entire image. In terms of hand-eye calibration, this paper effectively utilizes the stereo camera's 3D reconstruction results and provides intuitive error feedback for TCP calibration. Through practical engineering practice, the calibration system presented in this paper has good application value. The system can meet the application needs of refueling robots and effectively reduce the calibration time to less than 10 minutes.

Lower limb exoskeletons can enhance human locomotion performance and provide aid in rehabilitation. Due to human interpersonal differences, identifying a proper assistance strategy is challenging. The uptake of embodied intelligence that learns individuals’ needs and tasks’ requirements will help exoskeleton systems achieve their potential under different scenarios. Utilizing the evolution strategy to explore human reaction under exoskeleton assistance, “human-in-the- loop” (HIL) optimization is promising to obtain suitable assistance patterns. However, most current HIL optimizations use physiological signals, such as metabolic consumption and muscle activity, as the objective function to minimize, which need a long time to be evaluated and are inconvenient for real-life use. In the study, we aimed to construct a HIL optimization strategy to search effective exoskeleton assistance patterns based on the human-robot interactive force measured by wearable sensors. We first used a unilateral cable-driven ankle exoskeleton to explore the characteristics of human-robot interaction under 20 assistance patterns. A plantar-pressure-based cost function was constructed and real-timely evaluated for HIL optimization. A pilot experiment was conducted with a single participant. Optimized exoskeleton assistance can improve the individual walking economy by a 41.2% reduction in soleus muscle activity and a 21.3% decrease in metabolic cost. The proposed method is promising to improve the HIL optimization time efficiency and promote more effective real-life exoskeleton applications.

This paper presents a novel wireless calibration device for the irradiance of a solar radiation aging test chamber. The device is designed using the STM32 microprocessor, providing accurate and reliable measurement results. The wireless calibration device comprises several components, including a temperature sensor, a STM32F407VET6 main control chip, an irradiance sensor, an ATK-BLE02 Bluetooth module, and a mobile client. Upon acquiring temperature and irradiance data from the respective sensors, the data is then transmitted to the microcontroller unit (MCU). The developed system enables real-time monitoring of the temperature and ultraviolet (UV) intensity in the solar radiation aging test chamber, with the ability to remotely transmit the collected data to a mobile phone. Through the mobile client, operators have access to temperature and irradiance values and can collect data and generate curves. These features facilitate analysis of the solar radiation aging test chamber’s performance and enable convenient monitoring of its operational status. This device is characterized by its small size, high expandability, and ease of operation, and can monitor temperature and irradiance accurately in real time. It effectively avoids the adverse effects on the box caused by the uncertainty of temperature and improves work convenience while saving human and economic resources.

Considering the energy conservation and emission reduction is a key focus of attention. Owning to its special working condition, the quay crane usually undertake an important role. As a significant machinery in hoisting mechanism, the working states for gearbox is playing a significant role in machinery’s reliable operation. For extracting degradation indicators in huge vibration signals of quay crane, this paper proposes an improved KW symbol entropy for fault degradation feature. So as to keep united for the symbols, the technique introduces the root mean square for normal states signal and take the value as the standard and then combining a parameter named symbol coefficient in forming the unified symbols scale. Besides, a parameter named symbols number is proposed to expand the section of symbols set, the information expressing ability in improved then. Based on this, considering information entropy theory, some complexity index for symbol sequence is calculated, and then, two index are calculated as IKSE and IKDE. In order to verify the accuracy of the method proposed in this paper, an example is introduced to analyze the signals in actual working conditions. The results express that the two indexes can accurately represent the degradation trend of the signals, and for the parameter stability, has a certain application value.

This paper explores the natural language processing technology of news text, especially the algorithm and application of title generation and abstraction. Based on the analysis of the existing principle of automatic title generation and abstraction, the data sets are preprocessed by text cleaning and word quantization according to the objective and accurate features of news text. Then, the Roberta, GPT-2, and T5-pegasus models were evaluated, and the method of model fusion and integration based on the principle of accuracy priority was proposed to improve the accuracy of title generation by 3%. Finally, the news text processing system is constructed based on the optimized model. Experimental results show that the proposed algorithm has high reliability and robustness. This topic has been applied, and the application shows that the results of the news text processing system are accurate and objective, and have high adaptability.

The study and application of multimodal emotion recognition have gained significant popularity in recent years, representing one of the challenging tasks in the field of affective computing. We propose a multimodal speech emotion recognition model that utilizes multiple acoustic and textual information layers. This model incorporates transcribed textual data to complement speech data and enable accurate emotion recognition. In the unimodal model, we employ AlexNet, BiGRU, and HuBERT to extract multi-layer acoustic feature information from speech, and the RoBERTa encoder to extract text features. Additionally, we perform fusion between speech and text by utilizing the co-attentive mechanism to extract complementary information across modalities and eliminate inter-modality noise. This process ultimately enhances the emotional representation of the target modality. Finally, the fused features are utilized to predict the emotion category. Our model achieved a weighted sentiment recognition accuracy of 77.41% and an unweighted accuracy of 78.66%.

With the development of the metaverse, its position in cyberspace is becoming increasingly important. Due to its complexity and openness, the metaverse also faces many security threats and vulnerabilities. This paper focuses on the security vulnerability of the metaverse in cyberspace. First, the basic concepts and characteristics of the metaverse are introduced. Then, the architecture of the metaverse is hierarchically introduced, and the security vulnerability of the metaverse in cyberspace is analyzed. Furthermore, the security countermeasures of the metaverse in cyberspace are introduced. Finally, the importance of the metaverse is summarized and the corresponding security countermeasures in the future are discussed. This paper explores the security vulnerability of the metaverse in cyberspace from several technical perspectives and provides strong theoretical support and practical guidance for the development of the metaverse.

Vehicle trajectory information is important for intelligent traffic management. By considering the problems of vehicle trajectory acquisition, a trajectory data collection system that exploits the crowd sensing techniques is proposed in this paper. The smartphone is used to collect vehicle trajectory data. The system consists of trajectory data collection terminal, trajectory data server, and trajectory query terminal. The collection terminal uses a smartphone to take photos of the preceding vehicles and recognize the license plates. It is uploaded to the trajectory data server along with the timestamp and location. The server receives and stores the trajectory data from all mobile collection terminals. The query terminal queries the driving trajectory of a specified vehicle during a specified period from the trajectory database. The queried results are visualized on an electronic map. The proposed system is tested in urban scenes. The experimental results show its feasibility and availability. The designed system for vehicle trajectory collection and query expands the scene of vehicle perception and improves the real-time trajectory data collection. It can be applied to vehicle intelligence management in the field of intelligent transportation.

Grammar automatic error detection can help language learners identify whether there are errors in their own written text. For the grammatical error detection task of spoken Chinese in this paper, one of the most obvious shortcomings of this task is that the data set is small. In this case, there may be some sparse features in the training data set. To solve this problem, we propose a Chinese grammar error detection method combining Temporal Convolutional Network (TCN) and Threshold Recurrent Unit (GRU). To verify the effectiveness of the model framework proposed in this paper, we test our method on the Chinese spoken grammar dataset and Chinese text writing dataset respectively. By comparing the prediction effects of various deep learning models, the proposed model has a high F1 value in Chinese spoken grammar error detection and Chinese text writing grammar error detection.

With the development of networked combat and integrated combat ideas and technologies, military combat is increasingly emphasizing the system-system confrontation and the effectiveness of the system. In response to the problems faced in the assessment of the effectiveness of weapon and equipment systems, such as the large number of assessment indicators and the diversity of system combat tasks, a comprehensive assessment method for the combat effectiveness of weapon and equipment systems is given. A comprehensive assessment method based on improved principal component analysis is proposed to eliminate the influence of correlation between indicators, while retaining the differences in the importance of each effectiveness indicator. For typical combat tasks, multiple simulations are run, all the information of the original data is retained while eliminating the dimension, and all the effectiveness indicators are integrated to obtain the comprehensive combat effectiveness of the system, which makes the system effectiveness assessment results more reasonable and comprehensive.

In response to the need for quasi-real-time and high-confidence research and judgment of spacecraft abnormal proximity symptoms in space situation awareness, this paper proposes an intelligent detection method for orbital anomalies based on the high-dimesional representation of spacecraft behavior. Based on the target orbit element database to extract the spacecraft behavior time series characteristics and generate a high-dimensional representation matrix, a convolutional neural network structure integrating multi-dimensional characteristic classification and detection is designed, the orbital abnormal behavior characteristics is automatically learned, and it’s detected whether the spacecraft orbital is abnormal. The historical Two-Line Elements (TLE) data are used to generate a test set of abnormal orbital behavior of the spacecraft, and the Mahalanobis distance method and the intelligent detection method are used to jointly detect the test set. The test results show that the intelligent detection method provides a better orbital anomaly detection success rate on the self-built test set than the Mahalanobis distance method, which is increased from 85% to 98.5%. The intelligent detection method can be effectively used for detection of abnormal spacecraft orbital behavior.

With the operational advantages of unmanned combat platforms in modern war gradually appearing, the research of unmanned combat platforms has become the focus of all circles. In order to realize intelligent and autonomous unmanned operation in a real sense, a combat mission computer based on AI development board was proposed to be built as the control core of unmanned vehicles, simulate the operational mobility situation diagram of unmanned vehicles, and use the deep reinforcement learning network DQN to establish angle and distance decision-making network, so as to realize intelligent mobility decision-making of unmanned vehicles. The experiment verified that the unmanned vehicle can maneuver to the target area autonomously, which proved that the deep reinforcement learning network can realize the feasibility of platform autonomous and intelligent decision-making, and provided a feasible technical approach and theoretical support for the construction of combat mission computer to realize intelligent, autonomous and unmanned combat in a real sense.

The new incidence rate of rectal cancer accounts for 12.2% of all cancers, which seriously affects the health of the people. In addition, the early clinical diagnosis of T staging of rectal cancer will strongly support the formulation of effective treatment plans. However, it is difficult for clinicians to accurately determine the T stage of rectal cancer before operation because the site of rectal tumor invasion is not obvious. Therefore, this paper first proposes a bilinear feature fusion mechanism, which effectively avoids the problem of information loss in the process of convolution neural network training of rectal cancer MRI images; Secondly, the new weighted loss function designed can solve the problem of multi-example imbalance; Finally, the experiment proves that the constructed intelligent prediction strategy for T-stage of rectal cancer has a good accuracy, which provides a good auxiliary result for the clinical treatment of rectal cancer.

Forests play an indispensable role in the terrestrial biosphere by maintaining the global carbon cycle, preserving species diversity, and regulating the climate environment. The dominant tree species mapping in Yunnan is a crucial tool for reflecting the distribution of forest resources in the region. Feature engineering is very important in traditional machine learning. It primarily involves two demands: firstly, when there are too many feature dimensions, feature selection is performed to speed up model training by filtering out unimportant features. Secondly, when there are fewer features or poor model training performance, feature construction can be attempted to enhance dimensionality by understanding the problem. In this study, the study area, covering 78% of the forest in western Yunnan (approximately 45,000 km²), utilized the GEE (Google Earth Engine) cloud computing platform to combine long time series of Sentinel-2 images, topographic data, and environmental data to obtain 100-dimensional features. Three feature filtering methods, including Lasso, relevance hierarchical clustering, and feature recursive elimination, were used to filter the features, and three machine learning classifiers, including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support vector machine (SVM), were used to obtain nine 10-meter-long map products. The results showed an overall accuracy of 76.4% for the 9 tree categories. Feature filtering improved the overall accuracy rate by 1% to 2% and significantly increased the efficiency of the machine learning process.

3D Point cloud, which is considered as the simplest and most efficient shape representation for 3D objects, has been widely used in various real-world applications such as virtual reality, autonomous driving and digital twin. Point cloud completion aims to predict the complete shape structure and recover faithful details given partial observation of an incomplete input. Unlike previous completion methods based on linear architectures, this paper presents a novel hierarchical architecture for point cloud completion which divides the completion process into several levels in a coarse-to-fine manner and significantly improves the network capacity for recovering local details. First, we exploit feature connections between encoded partial inputs and decoded recovery results at the same resolution by extracting multi-scale feature points, which can provide rich information for the following generation process. Second, in order to exploit the local geometric information and interpolate the extracted features points, we introduce cross-attention based generators into the decoding phase. The cross-attention based generator preserves produced structures from previous levels and incorporate the extracted feature points into each step of a progressive generation. Extensive experiments show that our method outperforms state-of-the-art completion approaches on popular PCN and ShapeNet55 datasets.

In order to intuitively and succinctly reflect information, such as the dosage form and components of traditional Tibetan medicine prescriptions, and the correlation information between formulas and formulas, between formulas and traditional Tibetan medicine, we have collected and collated a certain amount of traditional Tibetan medicine formula data in this article, and constructed a knowledge map of traditional Tibetan medicine formulas. Through the construction of this atlas, traditional Tibetan medicine workers can intuitively display various direct or indirect related information of the formula, so as to facilitate a deeper understanding of the hidden knowledge in the formula information.

Object detection based on unmanned aerial vehicle (UAV) images is very challenging. The multi-scale size and high density of objects in the UAV view bring great difficulties. To fully address this issue to unleash the potential of UAV applications, the YOLOv5-STD model is proposed. First, we add one more head to locate extremely small object detection by shallow image features; second, we use the attention mechanism to optimize the backbone by the transformer; third, we use SPD-Conv to avoid the loss of fine-grained image feature information. At the last, sufficient experiments on the dataset VisDrone 2022 have proven that the model has good performance, compared with the basic model, the improved model has an average improvement of about 7% in mAP@.5 metrics, and the ablation experiments have verified that its improvement skills have a positive effect on the model. This paper can help developers and researchers get a better experience in the analysis and processing of unmanned aerial vehicle images.

The emergence of 3D point cloud analysis has brought about new opportunities and challenges in various fields such as autonomous driving, digital twins, and virtual reality. Accurate segmentation is crucial to 3D point cloud analysis, but challenges arise due to the lack of topological information, complex shapes, and sparsity and unevenness in point sampling. To address these problems, a novel point cloud segmentation network called PCSNet (Point Cloud Segmentation Network) has been proposed. PCSNet combines global and local features to determine the overall shape and detailed local information, respectively, through an encoder-decoder architecture that incorporates multi-scale feature fusion. The encoder progressively extracts local center points, fuses local features, and models global features with the transformer to construct multi-scale topological and semantic information. The decoder then recovers the original point cloud and incorporates multi-scale features by upsampling for accurate segmentation. PCSNet outperforms state-of-the-art point cloud segmentation approaches on two widely used benchmark datasets (ShapeNetPart and S3DIS).

This paper fits the shipping index and locates the influencing factors. On the model side, we conducted from three levels—the periodicity (rolling window), linearity (ARIMA), and nonlinearity (SVR) of the shipping index based on the rolling window autoregressive moving average-support vector regression (WINDOW-ARIMA-SVR) hybrid model. Regarding data, we took Tianjin Shipping Index (TSI) as the independent variable. The explanatory variables are 100 sets of Automatic Identification System (AIS) data and macroeconomic indicators from four provinces adjacent to Tianjin Port. First, the results show that the optimal rolling window length is four months based on the Mutual Information (MI) criterion. Second, the proposed model has the best fitting accuracy for TSI compared with several mainstream models. Finally, the significant influence variables of the nonlinear part of TSI change dynamically with the window rolling, but three types of indicators, namely trade import/export volume, GDP, and the total number of ships arrival/departure in the four provinces adjacent to Tianjin Port are significant in the whole domain.

On the basis of current research on traffic flow prediction, the article proposes a traffic flow prediction method based on the fusion of K-Means algorithm and GRU. This method first uses K-means for clustering analysis of traffic flow and establishes a traffic flow pattern database, and then predicts traffic flow through GRU training. After simulation experiments, the MAPE and RMSE values of the traffic flow prediction method based on the fusion of K-Means and GRU are lower than those of traditional GRU, LSTM, KNN, SAES, and SVM, and the fitting effect is good. It is a reference traffic prediction method.

In multisensor data fusion, optimal weighted fusion is a better method. In order to solve the weight problem in multisensor data fusion process better, an iterative weight correction method based on the optimal weight distribution principle is proposed. First, the reference sequence is constructed. Secondly, the mean and standard deviation of the measured data of each sensor are determined according to the reference sequence. Finally, the weight value of each sensor in the data fusion process is determined according to the standard deviation.

Computers and humans differ in their logic and approach when it comes to recognizing emotions, especially when it comes to micro expressions. Using this difference, we make a new CAPTCHA based on a new facial expression recognition algorithm. This new CAPTCHA requires the user to find a similar expression to the given picture and rotate the picture to fit the background. After that, we counter-attacked the CAPTCHA, and the probability of being successfully attacked was less than 1%. We also conducted user tests based on this. CAPTCHA is superior to others in terms of recognition success rate and recognition time and has a good application significance.

Detecting vehicles is a crucial task in autonomous driving systems for highways. However, current detection algorithms used in such scenarios may face issues such as overlooking or misidentifying small or faraway vehicles, as well as overall low detection accuracy in different situations. To tackle these challenges, this paper proposes a vehicle detection algorithm called CC-YOLO that specifically targets the highway driving scene. CC-YOLO is an improved version of YOLOv5s, leveraging the C2F module from YOLOv8 to replace the original C3 module in YOLOv5s’ Backbone and PAN-FPN. This replacement enhances the detection accuracy of vehicles while maintaining a lightweight model. In addition, the PAN-FPN is upgraded with a lightweight upsampling operator known as CARAFE to improve the recognition rate of small target vehicles. Finally, the detection head is enhanced with an ECA attention module to better learn vehicle features and improve the accuracy of vehicle detection. A vehicle detection dataset is constructed using real highway driving videos, and CC-YOLO is evaluated on this dataset. The experimental results indicate that, compared to YOLOv5s, CC-YOLO achieves an increased mAP of 6.8%, effectively improving vehicle detection accuracy in the highway driving scene.

PointNet++ is a simple but effective network designed for point cloud processing. However, the accuracy of PointNet++ has been surpassed by many other methods, like DGCNN and Point Cloud Transformer. These methods are way heavier compared to PointNet++, which is not favorable for the deployment of real-world products. In this paper, we propose a module called HD projection layers that was inspired by nonlinear kernels used in support vector machines. The HD projection layers project the features of the point cloud into a higher dimension, increasing the linear separability and therefore relieving the burden on the classifier. Equipped with HD projection layers, we extended PointNet++ into a new network, HD-PointNet, which also involves many other improvements and better training techniques. Experiments show that the accuracy of HD-PointNet is competitive against other modern methods while using fewer computation resources.

Many existing approaches to blackbox adversarial attacks follow attack strategies with predefined priori which are fixed throughout the process. As a result, they often require an excessive number of queries against the victim models to succeed. In this paper, we proposed a new attacking paradigm that better resembles real-world attacks in practical settings, where an agent (i.e., attacker) approaches the attack by taking actions (i.e., perturbations to the source image) through sequential interactions with the environment (i.e., the victim model) to achieve maximum rewards (i.e., the success of attack with the minimum number of queries). Naturally, as any action the agent chooses to take would alter the query image and change the state of the attack, the agent needs to adapt its policy accordingly along the trajectory instead of applying a predefined strategy unanimously. As an instantiation, we propose a “sequential query-based boundary blackbox attack” (SQBA), which learns a policy to adaptively select from a set of candidates attacking methods and then follow the selected method to apply one attack at each step. For demonstration, we restrict the candidate to subspace-based boundary attack methods. We show that the policy can be learned effectively with a variety of approaches, including imitation learning, policy optimization, and an ensemble of both. Extensive experiments on four benchmark datasets (MNIST, CIFAR-10, CelebA, and ImageNet) show that SQBA can significantly reduce the query complexity under different settings compared with baselines while keeping a 100% attack success rate. In addition, we find that the Reinforcement Learning agent as an ensemble of TRPO and BiLSTM performs the best among different agents.

Thorax disease classification is a complex problem due to lesions are scattered and different diseases are related. It will benefit the diagnosis to comprehensively analyze the context of the scattered lesions and multiple diseases. Most existing supervised methods use Convolutional Neural Networks (CNNs) to capture chest X-ray (CXR) features for classification, but usually ignore the context representation which is useful for multi-label classification tasks. In this paper, we propose a novel thorax disease classification network with joined CNN and Transformer (TC-CNNT) that learns global features and context representations simultaneously. Specifically, TC-CNNT includes a CNN branch to gradually extract global feature representations from low-level to high-level through convolutional filtering. Meanwhile, a transformer branch is designed to capture context-dependent features through the self-attention mechanism of the shift windows. In addition, the feature fusion and multi-loss strategy are applied to maximumly learn the complementary global and context representations. Finally, the TC-CNNT method is verified on the ChestX-ray14 dataset and compared with the state-of-the-art methods, the experimental results demonstrate its superior performance for thorax disease classification.

In order to detect and deal with floating objects on water surface in time and improve the supervision level of rivers and lakes, we propose a method for recognition of floating objects on water surface, based on Mask R-CNN algorithm. Firstly, we design a set of floating object label classification rules, and establish a real data sample set in the field of rivers and lakes. Then we propose a solution of floating objects recognition, which includes image capture service, AI analysis, and early warning service platform. We compare the floating object recognition method base on the Mask R-CNN model and the SIFT feature and conduct experiments with different feature extraction networks. The results show that the method is significantly better than the traditional SIFT method, the average accuracy is increased by 16.15%, the average recall rate increased by 13.75%, and the ResNet-based method is more capable of identifying irregular floating objects. This method is successfully applied to the river and lake supervision system, and the recognition accuracy of common targets is over 90%.

Battery electric vehicles are becoming more and more popular, and the number of power batteries scrapped is obviously increasing linearly. How to improve the utilization rate of retired power batteries has aroused extensive discussion among scholars from all walks of life. Based on the analysis of the retired power battery industry recycling chain, reverse logistics theory and reverse logistics network, this paper puts forward the reverse logistics recycling mode studied in this paper. According to the structure and layout characteristics of the retired power battery reverse logistics network, the location of each node in the reverse logistics network is systematically analyzed, and the mixed integer linear programming model is used to carry out the layout location and path planning of the logistics network under the retired power battery recycling mode. Finally, a specific numerical simulation case is selected and the model is analyzed and solved. Under the condition of sensitivity analysis of the objective function, the reverse logistics network optimization design scheme is obtained. This paper is very necessary to improve the utilization rate of used batteries, save resources, protect the environment and improve the academic system.

This paper gives out a multi-platform search method based on graphic search, and expounds the basic method of obtaining the initial search map based on the original information, communication information, and detection information. The search diagram initialization and the information update of detection and communication based on the target general and determination also are given out. Through simulation, the method is verified, and be able to implement the underwater collaborative search for the “PtoS” mode.

Automatic production system is a kind of artificially intelligent technology. To improve the intelligence of automatic production systems and easily implement simulation experiments, the solution properties of automatic production system scheduling problems under the JIT environment are researched. Firstly, to minimize total earliness/tardiness time, a mixed integer programming of the proposed problem is shown for simultaneously optimizing job scheduling and robotic move sequence. Secondly, to combine job scheduling and robot move sequence, the definition of robotic activity is applied, therefore, two-dimension scheduling problem can be transferred to one dimension scheduling problem. Thirdly, the robotic activity sequence represents the solution to the proposed problem. and solution definition is presented. Finally, the properties of the solution are discussed. The first property means the first two robotic activities and the last two robotic activities of the solution move the first job and the last job of job scheduling, respectively. That is to say, when job scheduling is given, the first two robotic activities and the last two robotic activities of the solution are determined. Thus the number of solutions is reduced from n(m+1)! to [n(m+1)-4]!. Solution space is reduced. The second property means the start moving time of the first robotic activity of each solution is non-negative. The third property and the fourth property show the relation of adjacent tanks. Robotic activity exchanges and job exchanges are discussed in the fifth property and the sixth property, respectively. According to these proposed properties, one solution can turn into another solution, and a theoretical foundation is provided for designing an optimization operator for solving automatic production system scheduling problem under just-in-time environment. The novel mode which optimizes the proposed problem will be developed for constructing simulation experiments.

When establishing a visualization platform for rotary kiln calcination systems, it was found that the abnormality of industrial sensors needed to be detected in order to achieve timely repair of sensors and data correction. This paper proposes a new LSTM-SRU neural network based on multi-scale time-series fusion for anomaly detection. Firstly, a multi-scale time-series feature fusion mechanism is established based on the data, such as long-term periodicity and short-term volatility. Then, based on the prediction error, the optimal threshold is used to detect potential abnormal data. The algorithm is validated on an actual rotary kiln sensor dataset and demonstrates superiority over baseline methods such as LSTM and GRU.

Aiming at the problems of multiple meanings, long lengths, and close connections with context information in the recognition of named entities in the field of Chinese tourism, as well as the complex composition of some entities, this paper proposes knowledge-enhanced tourism naming entity recognition method. Firstly, the knowledge-enhanced ERNIE pre-trained language model is utilized to obtain the semantic representation of tourism text. Secondly, the obtained word vectors perform feature learning and feature representation of local information on the input data through Convolutional Neural Networks (CNNs). Then, the Bi-directional Long Short-Term Memory (BiLSTM) is used to fully learn the forward and backward feature information of tourism text. At last, the tag decoding layer based on Conditional Random Fields (CRFs) is employed to address tag dependencies and produce the optimal sequence of tourism named entity tags. The effectiveness of the ERNIE-CNN-BiLSTM-CRF model is confirmed by experimental results, which involve a comparison with other models using the created tourism dataset.

To anticipate financial time series (FTS) more accurately and address issues with non-linearity, non-stationarity, and significant noise in financial data, neural networks and bionic algorithms are used. First, the FTS is mined for data using the Long Short-Term Memory (LSTM) neural network, which analyzes the time series’ features to predict how they will change over time. Second, the attention mechanism is used to enhance the LSTM neural network’s effectiveness in forecasting FTS. The experimental findings demonstrate that the forecasting model built has better outcomes for the Shanghai Stock Index and Shenzhen Stock Index in terms of results of fitting. And the error fluctuation range is modest, when compared to the forecasting impact of the RNN (Recurrent Neural Network) model and the LSTM model. Compared with commonly used forecasting algorithms, the forecasting errors of the designed forecasting model are reduced by 57.3%, 46.3%, 25.3%, and 59.4%, respectively. Therefore, the designed attention mechanism-improved LSTM model can forecast the FTS. The above results provide a reference for establishing a FTS forecasting model.

With the development of the Internet and advances in technology, financial transactions have become more convenient, but money laundering has also become more concealed, complex, and can be quickly adjusted according to regulatory policies, making traditional methods based on rule libraries and blacklists difficult to cope with the current Anti-Money Laundering (AML) challenges. To improve the efficiency and accuracy of AML work, various machine learning methods have been proposed. However, these methods are often difficult to be widely used due to their complex models and poor interpretability. To overcome the shortcomings of existing methods and adapt to the changing trends in money laundering while ensuring good interpretability, this paper proposes an AML method based on Hierarchical Risk Control Knowledge Graph (HRCKG). Firstly, a simulated AML dataset is generated based on a small amount of real data and money laundering patterns. Next, various transaction features are extracted to construct graded risk control indicators, which are used to automatically search for and filter risk control rules to form a rule library, and then construct an HRCKG. Finally, integrating rule-based reasoning and graph-based reasoning, this paper uses rule matching, node importance measurement, and community discovery algorithms to evaluate the money laundering risks of accounts, identify money laundering accounts, and discover money laundering groups. Through comparison with other machine learning methods and analysis of experimental results, the effectiveness and application value of the proposed method is demonstrated.

This paper proposes a low-latency method, based on FPGA, for edge extraction of hot-rolled strip. The method uses a global shutter camera module to capture strip images and detects the strip edge position by analyzing the intensity profile and the gradient of image gray values. The method has the advantages of low latency, strong robustness, and accurate results. The feasibility and effectiveness of the method are verified by FPGA simulation experiments, and a foundation is laid for the subsequent strip deviation automatic control algorithm.