This PDF file contains the front matter associated with SPIE Proceedings Volume 11064, including the Title Page, Copyright information, Table of Contents, Author and Conference Committee lists.

Space object observation by using small ground-based telescopes has become a common supplementary means for usual space object observation because of its flexible operation and convenient carrying. However, due to atmospheric turbulence, urban lighting and other interference factors, the low resolution, blurred contour and scarce texture of the small ground-based telescope space target observation image seriously affect the space target recognition. At the same time, there are few real-time data for space target observation, thus the conventional deep learning method which needs lots of offline trainning data cannot be applicated. Therefore, it is necessary to prepare typical space target reference image by means of three-dimensional model and simulation image. Generative adversarial network provides a simulation way to prepare training data offline. In the preparation of generation images of space target recognition for ground-based observation, the 3D model data of space target are projected in two dimensions by setting elevation angle, roll angle and spin angle. And the two-dimensional generation images of space target are obtained. Then, GAN is trained by using real-time observation data of some existing space targets and corresponding two-dimensional generation images as input, and simulation generation images of typical space targets are obtained by using two-dimensional datum map on trained GAN. Experiments show that the reference images simulated by GAN are more authentic than those generated by traditional methods, and have higher correct rate, accuracy and efficiency on recognition mission.

The sea background often fluctuates violently and has a low contrast with the target, which brings difficulties in detecting the infrared maritime targets. To solve this problem, the mixture Gaussian background modeling for sea background in the Fourier domain (FGMM) was proposed. First, the mixture Gaussian background model was constructed for the amplitude spectrum sequence at each frequency point. Second, the amplitude spectrum of the test frame was compared with the mixture Gaussian background model to separate the background and foreground frequency points. And the parameters of each Gaussian distribution were updated to adapt to the change of seawater. Also, the two features of the neighborhood amplitude spectrum contrast and the information entropy of local amplitude spectrum were fused into the mixture Gaussian background model to get the final detection results. Experimental results showed that the proposed method has good effects in suppressing the seawater and detecting the targets. Moreover, compared with the traditional spatial mixture Gaussian background modeling algorithm, its performance has been significantly improved.

A scale space-variant filter (SVF) is proposed on the basis of Harris arithmetic operators, which can smoothly isolate noise efficiently at the situation of keeping edge information of the image. Comparing SVF with Gaussian filter under step jump signal and initial image input, the result indicates that SVF is better than Gaussian filter. Using SVF to detect feature points of an image, the experiment shows that feature points detected from SVF output contain more edge information. Using 2D space limitations, Euclidian distance limitation and angle limitation, we can eliminate redundant feature points so that all the useful feature points are distributed in all regions of the image evenly. From the result of the examination for noise-contained image, we can draw the conclusions that the new robust feature point detector can get more accurate position of feature points and the distribution of the points is more rational than that of the points without those limitations.

Image segmentation plays a crucial role in many medical imaging applications and is an important but inherently difficult problem. The paper discuses the method that classify unsupervised image using a Kohonen self-organizing map neural network. This method exits two problems: training time of the network is too long and the classified result and quantity were bigger influenced by the noise of image. Two-dimensional Discrete Wavelet Transforms (DWT) decompose MRI image into the small size and denoise approximation images. Kohonen self-organizing map neural network is trained with approximation image, then trained neural network classify pixels of original image. Training time of the network was notability decrease and the classified quality influenced by the noise of image was notability reduce. The technique presented here has shown a very encouraging level of performance for the problem of segmentation in MRI image of the head.

In the on-orbit maintenance task of aircraft, it is often necessary to operate non-cooperative targets. Relative pose measurement of non-cooperative target is one of the key technologies. In this paper, a fast method for detecting the artificial characteristics of non-cooperative object in the images of CCD cameras in ultra-close range is presented. The detection target here is the ellipse formed by the projection of the circular structure on the imaging plane. The geometric parameters of the circular is fitted by binocular vision matching of the detected ellipse, and a invariant point feature on the non-cooperative target is recognized and selected, which is combined with circular normal to establish the coordinate system of non-cooperative target, then the relative pose of non-cooperative target can be calculated, and the algorithm for measuring the relative pose of non-cooperative target are formed. The test results show that the algorithm is accurate and effective, and meets the requirements of accurate measurement and positioning technology in on-orbit service.

Ship object detection has a wide range of applications in both military and civilian. In the military, as an important and classical object in the field of military, the quick and accurate detection of ship is closely related to the success or failure of the battle. Similarly, as survival tools in civilian use, whether the number and location of ships can be quickly and accurately identified has a strong impact on marine rescue. In recent years, various object detection methods have been applied to ship detection, which include traditional methods and methods based on deep learning. However, due to the complicated background of inshore ships, the detection process is easily interfered by the buildings on the shore. As a result, the effects of ship detection are not yet satisfactory and researchers are making efforts to further improve these methods. In this paper, we propose an inshore ship detection method based on improved Faster R-CNN. As a typical and benchmark framework in the detection field, Faster R-CNN is chosen as the base pipeline. On the basis of this method, Soft NMS and Focal Loss are added with consideration of the characteristics that sizes of ships and distances between them are all different. Experiments on various data sets validate the effectiveness of our improvement. The mean Average Precision on our own 200 ship dataset has increased by 6% to 80.4%. The mean Average Precision on the public HRSC2016 ship dataset has increased by 1.1% to 84.5%.

Pantograph carbon slide is an important device in power supply system of electric locomotive, the pantograph location is greatly significant for the geometric parameter measurement of the pantograph-catenary system. In order to enhance the adaptability of pantograph detection algorithm to the scene, and to reduce the false rate and missing rate of pantograph detection, this paper proposes a novel method based on the pantograph template for fast matching and horizontal edge detection projection in monocular infrared pantograph images. Firstly, the prior knowledge of the position of the pantograph and the catenary is combined with the template matching method to realize the rough location of the pantograph, and then the precise location of the pantograph by horizontal edge detection and horizontal unilateral projection. The experimental results show that this novel adaptive method realizes the non-contact detection and location of the pantograph effectively, and improve the efficiency significantly.

When the small targets are dim and with dark-spots interference in infrared images, the targets can not be accurately detected by multi-level filter. In order to solve this problem, the dark-spots filtering algorithm based on contrast suppression is added to the original multi-level filter for detecting small dim targets. First, the result of difference image based on the low-pass filtered image and the original image preserves the information of the targets and dark-spots at the same time. Second, the difference image and the original image are superimposed to remove dark-spots, according to the inhibitory factor determined by the contrast between dark-spots and background. Finally, the targets are enhanced by using Gamma correction. The experimental results demonstrate that the algorithm enhances the targets, removes dark-spots, and suppresses the background of false detection, so the multi-level filter adding the proposed method can effectively detect small dim targets.

Because the spatial resolution of remote sensing equipment is low, how to accurately obtain the true spectrum of the target to guide the stable tracking of the target is a complex problem. This paper proposes an air moving target tracking method based on spectral unmixing. Firstly the target abundance is obtained through the pretreatment of infrared image, secondly, accurate spectrum of target is acquired based on the linear combination relationship between the background spectrum and the measured spectrum, and then corrected by the atmosphere, that is, the object side is obtained. Accurate spectrum. Through the spectral feature comparison, the target tracking is guided according to the fingerprint information of the precise spectrum. A large number of field experiments prove that this method can accurately measure the weak targets in the remote sensing background, improve the stability and accuracy of tracking, and lay a foundation for the subsequent remote sensing target recognition.

When line structured light is used in 3D measuring with wide view field, the extraction of light stripe center becomes difficult because of the complex image background and non-uniform illumination. A line structured light center extraction method based on peak intensity is proposed in the paper. The peak intensity value of every pixel is calculated and the image is segmented according to the threshold related to the peak intensity, thus the structured light stripe center is extracted. Line structured light stripe images with wide view field have been processed by the proposed method. The experimental results indicate that our proposed method extracts the line structured light stripe center accurately when the background is complex, the illumination is non-uniform and the view field is wide.

Semantic segmentation is one of the basic themes in computer vision. Its purpose is to assign semantic tags to each pixel of an image, which has been applied in many fields such as medical field, intelligent transportation and remote sensing image. In this paper, we use deep learning to solve the task of remote sensing semantic image segmentation. We propose an algorithm for semantic segmentation of the Attention Seg-Net network combined with SegNet and attention gate. Our proposed network can better segment vegetation, buildings, water bodies and roads in the test set of remote sensing images.

We apply the semantic segmentation method in deep network to high precision satellite image change detection, and propose a network framework to improve the detection performance.We directly processed the image after registration, without the steps of radiometric correction, and avoided the tedious steps of manual feature design by traditional methods.We tried to use Unet and Deeplab v3 model to divide the change area, and added the structure of jumping connection on the basis of Deeplab network, which made the edge of the detection graph more accurate and improved the performance of the network.The test results show that this method is effective for detecting the change of highprecision remote sensing images.

Robust infrared small target detection in infrared warning and defending system is a challenging task due to the low signal-to-clutter ratios and complex background. Motivated by human vision system, we proposed a scale adaptive patch-based contrast measure(SPCM) method for infrared small target detection. At the first stage, a patch-based contrast model is established for measuring small target scale response, whose highest response corresponding to the best estimated size of the target, at the same time, filtering out some non-target areas and leave potential candidates. At the Second stage, we calculate local patch contrast at candidate regions with the estimated target scale. Utilizing the just right scale, the patch-based contrast measure could effectively suppress background clutter and extract infrared small target in single image. Finally, an improved adaptive threshold method by using statistical information of candidate target is used to segment infrared small target. In order to verify the effectiveness of the proposed approach, we compared our method with several fixed-scale and multi-scale infrared small detection algorithm. Experimental results indicate that our method is not only able to effectively estimate the actual scale of the target, but also detect weak small target accurately in heterogeneous background with low and comparable false alarm ratio, while achieving three times faster runtime performance than multi-scale algorithm.

It is difficult to detect maneuvering dim target from clutter background. It is important to synthesize multiple features to improve the performance of dim target detection. In this paper, one dim target detection method based on space-time salient graph was proposed. Salient analysis was used both on space domain and time domain. One TCF improved filter was designed and used to obtain the time salient graph. And the AC salient filter for dim target was applied to obtain the space salient graph. One fusion strategy was designed to fuse the space salient graph and the time salient graph. The coarse-fine two stages strategy was designed to obtain target detection result. The space-time salient graph was segment to get candidate targets. And then the SCR filter was used to confirm the result of target detection. Experimental results show that proposal method improved the performance of maneuvering dim target detection from clutter background.

This paper analyzes the importance of real-time identification of key parts of non-cooperative targets in military applications, and proposes the shortcomings of current target detection algorithms, and a real-time localization algorithm for target points based on TMS320C6678 parallel processing. The algorithm firstly locates the approximate position of the target by down sampling and horizontal vertical projection of the original image, then locates the target key points by polar coordinate transformation and target edge curvature solution, and combines the multi-core of TMS320C6678 DSP to optimize the algorithm. Finally, the algorithm is parallel optimized on the TMS320C6678 multi-core DSP hardware platform, which can accurately locate the target parts within 5ms.

Detecting underground target is important for national defense and security. Using the temperature field simulation, we can obtain the simulation model of the underground target. The data pattern of simulation is different from the data pattern of infrared remote sensing (RS), but the two patterns have a mapping relationship. We transform the data pattern of simulation to the data pattern of infrared RS, and then compare the transformed simulation data with the actual acquired infrared RS data to find the difference, so as to detect the underground target. Most of mappings of simulation data and infrared RS data have no sufficient robustness, and the mapping function is susceptible to external environmental factors. Using pix2pix model, a mapping approach is proposed to transform the simulation data to the infrared RS data. To evaluate this method, we take Deshengkou area of Beijing for experiment. Experiment shows that this mapping method has better robustness and adaptability.

Fixed-point attack on the key parts of small aerial vehicles is an important means of UAV (Unmanned Aerial Vehicle) countermeasure. Because of the fast speed and flexible attitude of fixed-wing aircraft, the detection accuracy of key points of fixed-wing aircraft in infrared images is low and the speed is slow. This paper presents an improved detection and tracking algorithm based on SVM. Firstly, the detection module extracts the fixed-wing aircraft area by image segmentation, then extracts the characteristics of the fixed-wing aircraft, then uses SVM to judge the flight direction of the fixed-wing aircraft, and then locates the key points according to the direction. The experimental results show that the proposed detection algorithm can process 30 frames per second on the platform of DSP (TSM320C6678), and still achieve a high detection rate (<93%) with very high practical value.

The available high-resolution remote sensing images are growing exponentially in recent years due to the rapid development of remote sensing imaging. However, several problems still exist: 1) How to solve the difficulty caused by the scale and shape of object. 2) How to detect the object quickly and accurately. Inspired by the hierarchical visual perception mechanism, we propose a fusion method combining the low-level feature and high-level feature obtained by convolution neural networks to detect ship target. At the same time, we introduce deformable CNN layer into convolution neural networks to solve the diverse scale and shape of object. Finally, based on the visual attention mechanism, the object contextual information is integrated into the network. The experiment results show that our model can achieve good detection performance and the framework has good expansibility.

Location identification is a research hot spot in computer vision. For a scene with the building, the location recognition method in this paper can accurately detect the building and identify the location. The specific method is to firstly determine the local feature extraction method to obtain more stable local features under different conditions. Secondly, encode image features effectively, build Scene codebook, establish image index, and compare image similarity. Fast and large-scale image retrieval can be achieved in this way. Then, in order to filter out the error matching results and choose the best matching result, a matching algorithm based on local spatial consistency is proposed. The shape model voting method with small calculation amount is proposed to obtain the position of the building in the scene picture. The experimental results show that the method can more accurately identify the location of the building, and the building images show good robustness and distinguishability when they are transformed.

The extraction of Region of Interest (ROI) is an important information guarantee in the application of imaging matching guidance, which directly affects the acquisition probability and matching accuracy of the target. Image segmentation is an important method to extract the Region of Interest of the target. Based on image segmentation algorithm, histogram equalization and morphological filtering, this paper proposes an effective image processing method to extract the Region of Interest of the target. (1) A variety of image threshold segmentation methods are applied to the actual processing flow, and their segmentation performance is compared and analyzed. Some image segmentation methods are obtained, which are suitable for target region extraction in template image preparation and target potential region location in matching recognition. (2) Preliminary localization of visible remote sensing images is carried, using color information, to obtain local regions, then enhance the image using histogram equalization method, finally morphological filtering is used to remove the edge noise. (3) The Otsu method and Kittler minimum error method are processed in parallel, then the segmentation results are fused, and the evaluation indexes such as area constraint, similarity and contrast are filtered to obtain the target region .Tests have been done with visible image and infrared image in this paper. The result indicates that the effectiveness of the morphological filter is more obvious after histogram equalization for the original image. Besides, the Otsu method and Kittler minimum error method are processed in parallel, then the segmentation results are fused to get a more precise Region of Interest, thus ensuring the accuracy and timeliness of imaging matching guidance.

In recent years, deep convolutional neural networks (CNNs) have achieved great successes in object detection, however, feature extraction is still sensitive to scale variation. FPN is one of the majority strategies to deal with this problem. It uses a top-down pathway and lateral connection to combine high-level features with low-level features, and then generates robust features. However, in FPN, high-level features are still unable to capture the detail information, and this results in the inconsistent representations for the same objects with different scales. To solve this problem, we proposed a Feature Enhanced Module to get more robust features, which can help the networks to produce object localization with higher quality, i.e., without bells and whistles. The performance of the proposed method is shown by the experiments in which it achieves a 1.1 point AP50 gain and 2.3 point AP75 gain on the Pascal VOC dataset, comparing to the Faster RCNN with FPN.

In the MBZIRC 2020 competition, an Unmanned Aerial Vehicle (UAV) is required to intercept a moving balloon and put it into a specific location. The core of the task is to accurately identify the balloon’s centroid, which is also the purpose of this article. The process is composed of two sections: first identify the balloon candidate region based on Faster-RCNN, an end to end object detection algorithm, following a new method based on the color of balloon to extract the centroid finally. In terms of Faster-RCNN, images of balloon sample library are used to generate a number of target candidate regions by region proposal network(RPN), next the neural network is trained to generate a model, which can finally output the boundary box of the balloon, which we called candidate region. Next, in the candidate region, the process includes three parts: feature extraction, target segmentation and centroid marking. Improve the saturation to enhance the image, thus reducing the impact of reflection of sunlight. Then replace the color of the balloon to pure black, with the use of adaptive filtering to segment the balloon region preliminarily. Finally, to minimize the affections of noise, the largest connected region in the image is chosen to calculate its centroid position. Experimented with different backgrounds of images such as sky, grass, flowers and buildings, our method has gotten wonderful results, thus verifying the high accuracy of our method.

Particle filtering is a key technique for moving targets detection and tracking in the field of remote surveillance system and air defense systems. Moving targets can be tracked by particle filter without registration. However, standard particle filtering cannot suite for high-precision tracking and track small dim moving targets occupying a few pixels in image, having low signal-to-noise ratio (SNR) and always flicking. To solve this problem, an improved algorithm is proposed to achieve detection and tracking for small dim moving targets. In the new algorithm, the prediction process of particle filter is improved by a linear regression method. It is applicable to the sequential images where the moving targets become smaller and dimmer gradually. Small dim targets can be detected and tracked directly with low SNR and without registration. The trajectory of the moving target is learned automatically through the past state of the moving target, and the trajectory is used for generating the importance density function. The importance density function is used as the prior probability in particle filter to sample and update particles. Through continuously learning and updating the trajectory of the moving target, the tracking accuracy is improved. Experimental results show that the tracking accuracy of the moving targets is greatly improved, and small dim moving targets can be detected and tracked without registration.

Object detection is an important part of remote sensing image processing and analysis. Traditional object detection methods in remote sensing imagery encounter with tough challenges when detecting small objects such as aircrafts and automobiles, due to complex background clutter, small target size, variation of visual angle, etc. We propose a targets detection network to detect the aircrafts in large-format remote sensing imagery based on deep convolutional neural network. Our method utilizes the Feature Pyramid Network (FPN [1]) to extract and inosculate multi-scale convolutional features to model the characteristics of targets and background. Moreover, in order to reduce the computational complexity of convolutional neural network, we utilize MobileNet [2] as backbone network and propose a computational efficient region proposal structure. We collect and manually annotate a dataset for aircrafts detection in remote sensing imagery in order to evaluate the proposed method. We achieve an average precision (AP) of 0.91 on the dataset, which is superior to other state-of-the-art methods, while our model is still faster and more compact than other models.

xAerial object tracking technology is widely used in unmanned aerial vehicle, air traffic control and drone counters. The tracking algorithm using only a single feature is difficult to adapt to the complex tracking environment. How to achieve robust tracking of aerial objects under different interference conditions is the main problem studied in this paper. To solve this problem, this paper proposed a multi-feature fusion tracking algorithm based on confidence evaluation. At the same time, in order to solve the problem of large variation of aerial object scale, a parallel scale estimation strategy based on the Discriminative Scale Space Tracking algorithm is proposed in this paper. This method estimates the optimal width and height of the object respectively, which can estimate the scale variation of aerial object more stably. For the occlusion problem, an occlusion-aware part-based model is proposed in this paper. The part-based local model can deal with the occlusion problem effectively, while the global model is more suitable for dealing with the fast motion and motion blur of the object. Therefore, a tracking method based on multi-tracker relay is proposed in this paper. In this method, the tracking state is judged according to the confidence of the model. In the normal tracking state, the global model based on multi-feature fusion is used, and when occlusion interference exists, the tracking model is replaced by the parts-based local model. In this way, the algorithm can effectively deal with various tracking situations.

Detection of sea surface targets in large-scale remote sensing images is one of the important research topics of ocean remote sensing technology. Ocean remote sensing images have the characteristics of wide format, strong interference and small target. This paper adopts the spinning target detection method, and proposes a ship detection model based on YOLO to output the real length, width and axial information. The model can accurately output the position, length and width and axial information of a ship target by predicting the minimum external rectangular area of the ship target, so as to realize multi-target detection and improve the detection performance significantly. To improve the recall rate of the target detection algorithm, this paper adopts the spinning target detection method, and proposes a ship detection model based on YOLO. Through redefining the representation of the rotation matrix and redesigning a new network loss function and the rotated IOU computing method, this model accurately outputs the real length, width and axial information, increases the output feature dimensions, and effectively raises the recall rate and speed of multi-target detection. Lastly, to improve the practicability of the algorithm on mobile devices, the model is processed in a lightweight way. Its parameters are significantly reduced while the detection accuracy is ensured.

The technology of automatic selecting landmark plays a significant role in aircraft navigation and ground information assurance. Compared to the normal object detection, it is quite difficult to describe and quantify the characteristics of a landmark due to its various status and no stable structure. This paper attempts to innovatively combine CNN with the technology of selecting landmark. The algorithm used in this paper uses a structurally stable adaptation region as a learning sample to train the CNN classification model. In the selection phase, remote sensing images were cut into pieces of patches, landmark of which was then recognized through the CNN classification model. Non-maxima suppression was used to filter out the low rate landmark and a correlation peak-based uniqueness analysis (the ratio of primary and secondary peaks and the highest sharpness of peak) was used to ensure landmark with no similarity pattern in the remote sensing image. The results indicate the effectiveness of proposed method for Selecting Remote Sensing Image Adaptation Structure.

Amid the maturity of machine learning, deep neural networks are gradually applied in the business sector rather than be restricted in the laboratory. However, its intellectual property protection encounters a significant challenge. In this paper, we aim at embedding a unique identity number (ID) to the deep neural network for model ownership verification. To this end, a scheme of generating DNN ID is proposed, which is the criterion for model ownership verification. After embedding, the model can complete the original performance and own a unique ID of this model as well. DNN ID can only be generated by the owner to check the model authorship. We evaluate this method on MNIST. Experiment results demonstrate that the DNN ID can accurately verify the ownership of our trained model.

When the model begins a new task, the challenge of naming the "catastrophic forgetting" limits the scalability of the deep learning network, which quickly forgets the learning capabilities it has. The fine-tuning method recommends that the original feature extraction be retained to extract the features of the new task and to achieve the purpose of learning the new class. However, this method degrades performance on previously learned tasks because the shared parameters change without new guidance for the original task-specific prediction parameters. This paper proposes general fine-tune method to reduce catastrophic forgetting in sequential task learning scenarios. The critical idea of the method is fine-tuning the parameters in each layer, unlike the traditional fine tuning only for the last layer. The experimental results show that the new method is superior to fine-tune, in the accuracy of the old task and the performance of the new task is better than that of the EWC. A distinct advantage is that old tasks do not limit the performance of new tasks but provide some support for new tasks.

This paper proposes an effective anti-occlusion target tracking algorithm based on template filtering.First,The location correlation filters are constructed to determine the target center position. In order to determine the target scale, a scale correlation filter is performed to sample multi-scale images surrounding the target region. Then,Peak-to-Sidelobe Ratio and average peak-to-correlation energy (APCE) are used to determine whether the target is occluded. When the target is occluded, the adaptive updating of the model is stopped.When the target is completely lost, the grid-based motion statistics algorithm is used to re-determine the position of the target.Experimental results demonstrate that our method can achieve a better tracking result than other methods.

Vehicle identification is widely used in route planning, safety supervision and military reconnaissance. It is one of the research hotspots of space-based remote sensing applications. Traditional HOG, Gabor features and Hough transform and other manual design features are not suitable for modern city satellite data analysis. With the rapid development of CNN, object detection has made remarkable progress in accuracy and speed. However, in satellite map analysis, many targets are usually small and dense, which results in the accuracy of target detection often being half or even lower than the big target. Small targets have lower resolution, blurred images, and very rare information. After multi-layer convolution, it is difficult to extract effective information. In the satellite map data set we produced, the target vehicles are not only small but also very dense, and it is impossible to achieve high detection accuracy when using YOLO for training directly. In order to solve this problem, we propose a multi-feature fusion target detection method, which combines satellite image and electronic image to achieve the fusion of target vehicle and surrounding semantic information. We conducted a comparative experiment to demonstrate the applicability of multi-feature fusion methods in different detection models such as YOLO and R-CNN. By comparing with the traditional target detection model, the results show that the proposed method has higher detection accuracy.

During the movement of the target, we want to stabilize the target in the fixed region of the image. However, with inherent noise in this region, the information of the target on the image will be lost when the target enters this region. Besides, when there are other disturbances in the field of view, it will be more difficult to stabilize the target in the region with inherent noise. To solve this problem, we propose a tracking method to stabilize the target in the region of inherent noise in an image. We use multi-level filter detection method to detect the target in the image and obtain the position information of the target. Then according to the first 3 frames of images, the linear weighted sum value of the features and the motion trajectory of the target are estimated, and they are updated constantly in the process of tracking. The experimental results show that the tracking method has good performance in real-time, stability and anti-interference.

Detecting ship targets distributed in infrared images of clouds, waves and other complex disturbances in an unknown complex background, and determining the true ship target in the case where the target and the false alarm are very similar are widely used and challenging tasks. This paper proposes an infrared ship targets detection algorithm based on Bayesian theory and SVM combination. Bayesian theory can be used to estimate the probability of partial unknowns with incomplete information. Bayesian formula can be used to correct the probability of occurrence, and finally we can use the expected value and the modified probability to make the optimal decision. At the same time, support vector machine (SVM) is a novel small sample learning method with solid theoretical foundation. It does not involve probability measures and laws of large numbers, so it is different from existing statistical methods. Firstly, the paper introduces the infrared ship target detection method based on Bayesian theory. Then, the image is post-processed to remove redundant false alarm targets. Finally, the paper introduces the experimental data and the performance evaluation indicators of ship detection results, and compares with other ship detection methods to obtain experimental results.

In order to identify the threat types of small targets in complex sea and sky background according to operational rules. Using GA to modeling operational rules of the targets which include the action trajectory and the action speed. On the basis of crossover operator, mutation operator of GA to generate the original population. Calculating the optimal individual adaptive value, which take as the discriminating reference value. Comparison of adaptive function values of target actual motion trajectory and adaptive function of optimal motion trajectory value results to verify the validity of the modeling. Due to particularity of the battlefield ,less the priori knowledge of the attack target, can not achieve feature comparison in the cases to recognition. Decompose the operational rules into multi-function optimization problem which is GA skilled.

Aiming at the accuracy and speed of image change detection, an improved registration algorithm combining wavelet transform and SURF algorithm is proposed, and image change detection is completed by an image adaptive constraint threshold method. Firstly, the image is decomposed based on wavelet transform and the low-dimensional components are coarsely registered by SURF. Then the image is dimension reduced by PCA, and the obtained feature points are coarsely registered according to the bidirectional registration criterion. Then the RANSAC algorithm is used to select the exact one. The registration point is the least-squares fitting registration of the image. Finally, the image is detected by the adaptive constraint threshold method based on the mean ratio difference map based on the precise registration. The experimental results show that the accuracy and speed of the registration algorithm are better than those of SIFT and SURF. The detection method is better and the detection accuracy is improved.