This PDF file contains the front matter associated with SPIE Proceedings Volume 11400, including the Title page, Copyright information, and Table of Contents

Most of dehazing algorithms are based on the well known Koschmieder image restoration model working on color images. Nowadays, many autonomous underwater vehicles use marker detection in order to estimate their 3D pose according to the marked target. In this paper, we propose an adaptation of the Koschmieder for grayscale images which is more suitable for marker detection. This is done by enhancing the final energy (the radiance) of the image lost during the processing. We show that by multiplying the final radiance image by the shifted transmission of the Koschmieder model, we can enhance the gradient and the contrast of the image. We have implemented our Koschmieder adaptation into two methods from the literature and proven its robustness on an underwater dataset containing ArUco markers. The obtained results outperform the existing methods in terms of marker detection rate without degrading the pose estimation.

Radon transform, and more specifically the multiscale Discrete Radon Transform, is a valuable tool to find straight structures on images. But the straight lines that DRT finds are those traversing completely the image, and hence DRT is unable to detect non-straight structures. That limitation is normally adressed by evaluating Radon transform on non-overlapped subsquares of the image. In this work we will show that the initial partial transform stages of multiscale DRT contain enough information to detect non-straight edges. By stopping on initial stages of an already linearithmic transform that can be executed on integer arithmethics, the proposed method is found to be much faster than its alternatives. We will show its ability to process images at video acquisition rate on mobile phones.

The problem of clustering transportation networks has been studied in the static framework by considering traffic conditions at a given time. Nevertheless, it is important to underline that traffic is a strongly time-variant process and it needs to be studied in the spatio-temporal dimension. Considering the fact that the congestion is spatially correlated in adjacent roads and it propagates with different speeds, partitioning a transport network into homogeneous trajectories that evolve over time can be extremely useful in order to design a real-time traffic control schema which alleviate or postpone congestion. The paper proposes an evolutionary spectral clustering approach to partition a graph transport network into connected homogeneous trajectories that evolve over time. In order to choose the number of clusters automatically, we use the density peaks algorithm which is based on the idea that cluster centers are characterized by a higher density than their neighbors and by relatively large distance from trajectories with higher densities. The clusters are recognized and the outliers are excluded from the analysis. This method is proved to be efficient regardless the shape and the dimension of the data set. We perform experiments on real road speeds for Amsterdam city traffic network, our results show that the proposed evolutionary spectral clustering algorithm outperforms the static clustering algorithms in its efficiency and robustness.

Paramedics often have to make lifesaving decisions within a limited time in an ambulance. They sometimes ask the doctor for additional medical instructions, during which valuable time passes for the patient. This study aims to automatically fuse voice and text data to provide tailored situational awareness information to paramedics. To train and test speech recognition models, we built a bidirectional deep recurrent neural network (long short-term memory (LSTM)). Then we used convolutional neural networks on top of custom-trained word vectors for sentence-level classification tasks. Each sentence is automatically categorized into four classes, including patient status, medical history, treatment plan, and medication reminder. Subsequently, incident reports were automatically generated to extract keywords and assist paramedics and physicians in making decisions. The proposed system found that it could provide timely medication notifications based on unstructured voice and text data, which was not possible in paramedic emergencies at present. In addition, the automatic incident report generation provided by the proposed system improves the routine but error-prone tasks of paramedics and doctors, helping them focus on patient care.

To incorporate object locations in a multi-target detection model, we assume that a close duplicate cannot be learned by the model efficiently. So, we use a region-based approach which uses more object location compared to the ground truth locations to localize the targets. The proposed model is able to learn a similarity metric with respect to the ground truth locations which is robust (low false positives) enough for varying images conditions, small aerial target sizes and using few training samples. We report preliminary results on how transfer learning of meta-data a effects small aerial target localization accuracies. Quality ranking from Intersection-over-Union (IOU) in region segmentation models on the aerial ground truth data using pre-trained models from ImageNet, AlexNet, and CIFAR-10 and initialization with three aerial datasets such as the satellite imagery XView2.

The role of data analytics in industrial quality inspection has grown rapidly in the last few years. This has also instigated increasing interests in the development of analytical, data driven models based on machine learning techniques with vision systems in industrial quality inspection. The deep learning employs computational models which are composed of a series of transforming and processing layers to learn representations of data with multiple levels of abstraction. In this paper, we present a novel approach combining deep convolutional neural networks (CNN) and National Instruments (NI) vision system for image classification to detect cracks on product surfaces. To mimic cracks on product surfaces, 15 various samples with and without cracks were designed in a 3D computer-aided design (CAD) solid modeling software and constructed using a 3D printer. In data acquisition, these samples with dimensions of 10 x 10 x 1 cm were systematically imaged using CCD camera and NI vision system, and approximately 1000 images obtained. Acquired images were input to two CNN classifiers, AlexNet and GoogLeNet architectures. The CNN automatically learns a hierarchy of increasingly complex features and a related classifier directly from training data sets. For AlexNet architecture, the CNN employs 5 convolutional and 3 max pooling layers with 3 fully connected layers to perform the image classification. For GoogLeNet architecture, the CNN utilizes 22 layers with multiple inception modules and SoftMax classification layers. The preliminary results demonstrate the feasibility of using deep learning algorithms with NI vision system to perform the automated industrial quality inspection.

Satellite and aerial imaging is being one of the most attractive source of information for the governmental agencies and the commercial companies in recent times. The quality of the images is very important to pick the valuable information from the details especially for high value targets. Satellite images may have unwanted signals called as noise in addition to useful information for several reasons such as heat generated electrons, bad sensor, vibration and clouds. There are several image denoising algorithms to reduce the effects of noise over the image to see the details and gather meaningful information. Many of the traditional denoising methods can filter noise, but at the same time they make the image details fuzzy. This paper presents the convolutional neural network (CNN) based image denoising method that retains the image detail but removes the unwanted noise. The proposed method employs a residual learning strategy, meaning that the CNN network learns to estimate the residual image. A residual image is the difference between a pristine image and a distorted copy of the image, and contains information about the image distortion. An extensive experiments demonstrate that proposed CNN denoising model can not only exhibit high effectiveness in image denoising tasks, but also be efficiently implemented by benefiting from GPU computing. The proposed method is applied to both aerial and satellite imagery and effectiveness is measured using Peak Signal to Noise Ratio (PSNR), Structure Similarity Index Metric (SSIM), and Naturalness Image Quality Evaluator (NIQE), which is also called perceptual quality index.

We propose a new recognition method to extract effective information from receipts by integrating deep learning algorithms from computer vision and natural language processing. Our method consists of three parts. The first part provides effective areas for receipt detection. By removing noise and extracting the gradient of the receipt image, we determine the threshold to crop and reshape the useful receipt area. Detecting text from a receipt image is the second part, we modify and deploy the text detection algorithm connectionist text proposal network (CTPN) to locate the text region in the receipt. In the third part, we import the connectionist temporal classification with maximum entropy regularization as the loss function for updating the convolutional recurrent neural networks (CRNN) to recognize the text detection area, which converts the receipt from an image into the text. Based on our method, the effective information of a receipt can be integrated and utilized. We train and test our system using the data set published by scanned receipts optical character recognition and information extraction (SROIE). The results illustrate that our recognition system is able to identify receipt information quickly and accurately.

We propose a method of detecting and tracking occluded illegally parked vehicles. The method used a deep learning framework that can detect and track moving vehicles. To obtain the long term tracking of stationary vehicles the process must be capable of withstanding large changes in lighting, weather and large amounts of occlusion from passing vehicles. A modified dense SIFT descriptor algorithm has been developed. This compares the current frame with the background and removes objects in motion. The tracking of the occluded illegally parked vehicle is achieved by YOLO version 3 algorithm, combined with a predictive filter. For each illegally parked vehicle, the occluded portion is not used for feature point matching. Based on the matching result, the occluded illegal vehicle can be tracked. Our approach tested performance on a public database(i-LIDS) and the results indicate the method produces a very high accuracy compared to other published work.

Dementia progression is based on exploring models predicting longitudinal disease patterns and represents a challenging research field for neurodegenerative diseases. Conventional progression models are mainly continu- ous network diffusion models assuming a radial contact-based spreading. In dementia-affected brain networks, however, we observe atrophy in specific brain regions that do not assume omni-directional contact-based disease progression but favor a path-based progression relying on misfolded and aggregated proteins flowing from one region to another. Here, we propose a novel concept to biologically model disease progression based on an disease characterization matrix that comprises both the routing and the amount of traversing proteins. We compare the path-based spreading with the contact-based spreading mechanism for brain network graphs for structural MRI data for healthy controls, mild cognitive impairment and Alzheimer’s patients. As biomarkers we extract critical epidemic thresholds for both spreading mechanisms. The path-based spreading mechanism corroborates the clinical observations that disease spreading in dementia is persistent and thus increasing the transportation of misfolded and aggregated proteins as with disease evolution will lower the critical epidemic threshold.

A test system with four cameras in the infrared and visual spectra is under development at FFI (The Norwegian Defence Research Establishment). The system may be mounted on a jet aircraft or may be used in a land-based version. It can be used for image acquisition or for testing of automatic target recognition (ATR) algorithms. The sensors on board generate large amounts of data, and the scene may be rather cluttered or include anomalies (e.g. sun glare). This means we need algorithms which are robust, fast, able to handle complex scenes, and data from up to four sensors simultaneously. Typically, estimates of mean and covariance are needed for the processing. However, the common maximum likelihood (ML) estimates are in general too sensitive towards outliers. Algorithms based on order statistics are known to be robust and reliable. However, they are computationally very heavy. But approximations to order statistics do exist. Median of medians is one example. This is a technique where an approximation of the median of a sequence is found by first dividing the sequence in subsequences, and then calculating median (of medians) recursively. This technique can be applied for estimating the mean as well as the standard deviation. In this paper we extend this method for estimating the covariance matrix and the mean vector, and discuss the strategy with respect to robustness and computational efficiency. Applications for use in image processing and pattern recognition are given.

Traditionally, homodyne and heterodyne are a combination of a frequency mixer followed by a low-pass filter. Mixing two signals of frequencies f₁ and f₂ generates signals of frequencies f₁ + f₂ and f₁- f₂ and their integer multiples. Both multiplication heterodyne and phase-sensitive detection have been demonstrated optically by using photorefractive, four-wave mixing (FWM). The multiplicative characteristic of FWM was used for mixing, and the response time of the photorefractive medium is used for low-pass filtering. If one of the input beams is both spatially and temporally modulated using a controlled oscillating membrane, depends on which model is heterodyned, one can generate an orthogonal set of Bessel bandpass filters. This scheme can be integrated part within parallel data acquisition systems for applications involved in nondestructive testing.

Real-time holography has been used in numerous analog computing applications; most important was in applications of real-time holography in four-wave-mixing for performing correlation and convolution. Correlation algorithms/devices are one of the important areas of pattern recognition. In controlling the beam ratio in four-wave-mixing for both correlation and convolution, it is possible to convert the matched filtering to inverse filtering correlation and convolution to deconvolution. We demonstrate that it is possible to optically differentiate images simply by deconvolving images with a step function’s Fourier Transforms. A perspective of image differentiation and its limitation, application with optical differential correlation, will be discussed in this paper.

Pattern detection is an active field in big data streams analytics with numerous ongoing challenges. Actually, due to the great velocity and variety of data, new patterns can appear and change over time. Existing state-of-the- art solutions consist in updating the pattern detection model regularly in order to integrate newly appeared and validated patterns. However, in several applications, such as security and defense, patterns can represent anomalies. Therefore, it becomes crucial to detect new patterns (i.e. new anomalies), as early as possible, in order to react at the right moment. Consequently, emergent pattern detection becomes a very challenging task. To tackle this challenge, we propose EPDA (Emergent Pattern Detection Algorithm): a new and validated algorithm for detecting emergent patterns in data streams. The originality of EPDA consists in exploiting frequent pattern mining techniques by proposing new statistical measures in order to estimate the evolution of emergent patterns over time. To perform this detection in a real-time, EPDA runs on the well-known Apache STORM distributed real-time computation system. To better fit our algorithm, we propose a new Apache STORM topology which is composed of one Spouts level and two Bolts levels. Experiments on a real data stream have shown the relevance of the proposed measures and the efficiency of our algorithm in a prediction task and in terms of execution time.

Performing underwater exploration with Autonomous Underwater Vehicles (AUV) requires low power and high resolution techniques. New computer vision techniques can be used for underwater image classification on embedded devices. These techniques must face machine resource constraints to offer high performance and low power consumption. This paper presents how to implement a Deep Neural Network (DNN) on Field Programmable Gate Array (FPGA) to perform underwater exploration with an AUV. We introduce tools and methodology to adapt the technology to the underwater context. This paper is part of a work to create an embedded system that can fit into an AUV to perform real time analysis of the underwater environment (using video camera as main sensor) with high autonomy and endurance. This will be achieved by overcoming underwater exploration challenges as : low power consumption, high classification performance, shortage of high-quality labeled data to train algorithm.

CPU performance is estimated from the execution of processes per unit time. The selection of the CPU scheduling algorithm in less time is a vital issue. In this paper, a novel approach has been proposed in which selection of an appropriate CPU scheduling algorithm is done through machine learning algorithms dynamically. The result of the proposed algorithm is shown in the experimental section. Through experimentation, it is found that a decision tree gives better results in terms of accuracy and computational time as compared to other machine learning algorithms.

Detection of a moving human is challenging for real-time systems. Misdetection in high alert security areas may lead to heavy losses. This paper presents an optimized approach to avoid this misdetection in sensitive areas. Rotation invariant optimized correlation filters are used for detection of humans. Some pre-processing algorithms such as background subtraction and color space conversion have been linked to the correlation filters to minimize processing time and maximize the accuracy of target detection. The experimental tests of the proposed methodology validate that better accuracy can be achieved if the proposed optimized approach is utilized for moving human detection in real-time systems. In future work, the proposed approach will be extended to detect human activity at night and thermal imagery.

A new concept of targets recognition is presented which allow the computer resources redistribution and the computing processes reorganization depending on the characteristics of the targets. A target’s image complexity matrix is described. The necessary number of the target’s image pixels was determined dependent on the complexity matrix. The general purpose adaptive systems with fixed structure and reconfigurable architecture are described. The specialized opticalelectronic correlation multiprocessor functional distributed system is presented. The results of the investigation of the elaborated targets recognition systems are described.

The emergence of Generative Adversarial Network (GAN)-based single-image super-resolution (SISR) has allowed for finer textures in the super-resolved images, thus making them seem realistic to humans. However, GANbased models may depend on extensive high-quality data and are known to be very costly and unstable to train. On the other hand, Variational Autoencoders (VAEs) have inherent mathematical properties, and they are relatively cheap and stable to train; but VAEs produce blurry images that prevent them from being used for super-resolution. In this paper, we propose a first of its kind SISR method that takes advantage of a selfevaluating Variational Autoencoder (IntroVAE). Our network, called SRVAE, judges the quality of generated high-resolution (HR) images with the target images in an adversarial manner, which allows for high perceptual image generation. First, the encoder and the decoder of our introVAE-based method learn the manifold of HR images. In parallel, another encoder and decoder are simultaneously learning the reconstruction of the lowresolution (LR) images. Next, reconstructed LR images are fed to the encoder of the HR network to learn a mapping from LR images to corresponding HR versions. Using the encoder as a discriminator allows SRVAE to be a fast single-stream framework that performs super-resolution through generating photo-realistic images. Moreover, SRVAE has the same training stability and "nice" latent manifold structure as of VAEs, while playing a max-min adversarial game between the generator and the encoder like GANs. Our experiments show that our super-resolved images are comparable to the state-of-the-art GAN-based super-resolution.

The standardization of the Versatile Video Coding (VVC) is still carrying out. Its first version is planned to be finished by the end of 2020. VVC promises higher coding efficiency than the High Efficiency Video Coding (HEVC) due to the improvements of the different coding tools. The adaptive Multiple Transform (AMT) is one of the major enhancements of the transform module of the VVC where five transform types have been introduced. However, the encoder and decoder complexities have been increased dramatically. This paper discusses one solution for codec complexity alleviation by considering transform approximation. It focuses on the approximation of the DST of type VII. The proposed approximation consists in applying the Particle Swarm Optimization (PSO) algorithm to the target transform in order to get an approximate matrix with low computational complexity requirements. This algorithm is used to cope with traditional approximation techniques based on constrained optimization of the DST transform. The proposed transform has shown a significant computational complexity reduction with no quality degradation and a tolerable BD-rate loss.

This paper aims at comparing 4 top models for crowd counting and evaluating their highlights based on their performance. In DSNet, the distended convolution block network was proposed, where the distended layers are densely connected to each other in order to preserve information from continuously varied scales. Three blocks are cascaded and linked to dense residual connections to widen the range of levels covered by network and also a novel loss of consistency at multi-scale density level was introduced to improve performance. In SFANet, two foremost elements with VGG backbone CNN and two-way path multi-scale fusion networks were suggested for the front end feature extractor and back end to make density map in which one path highlights crowded regions present in images. The other direction is responsible for the fusion of multi-scale features and for the generation of the final high-quality high-density maps. In MANet (Multi-scale Attention Network), a new mechanism of soft attention was presented, which learns a series of masks and a level-conscious loss feature was introduced to regularize and direct the learning of different branches to specialize on a specific scale. In Bayesian Loss, a novel loss function was used to generate a density contribution model from the point annotations. We also analyzed the results of the 4 convolutional neural networks, extracted the pattern of convolutional neural network structure and found promising pathways for researchers in this fast-growing area.