Unsupervised statistical neural networks for model-based object recognition

The article explores modern neural network architectures for the automatic detection and recognition of marine surface objects and obstacles of given classes throughout the full image area, applicable for execution in real or near real time on an optoelectronic vision system to au-tomate and improve the safety of civil marine navigation. A formal statement of the problem of automatic detection of objects on images is given. The state-of-the-art algorithms for detecting objects in images based on use of artificial convolutional neural networks were reviewed, their comparison was made and a reasonable choice was made in favor of the most efficient neuralnetwork architecture in terms of computational complexity to recognition accuracy. The subject area is studied, as well as publicly available databases of surface objects suitable for use in the training of algorithms using artificial neural networks. The article concluded that there is insuffi-cient labeled data for training neural network algorithms, as a result of which the authors inde-pendently collected research images and video sequences, prepared and labeled the collected data containing surface marine objects and other obstacles that represent a navigation hazard for ships. Based on the selected neural network architecture, a new neural network algorithm for automatic full-frame detection and recognition of surface objects was developed, and an artificial neural network was trained using the prepared database of images of typical objects. The resulting algorithm was tested by the authors on a validation data set, the quality of its work was estimated using various metrics, and the algorithm’s performance was measured. Conclusions are made about the necessity to expand the collected database of images of typical marine objects, further steps are proposed to improve the accuracy of the developed software and algorithmic complex and its implementation to be used in a marine optoelectronic machine vision system for automa-tion and improving the safety of civil navigation.

Object recognition in video has seen giant strides in accuracy improvements in the last few years, a testament to the computational capacity of deep convolutional neural networks. However, this computational capacity of software-based neural networks coincides with high power consumption compared to that of some spiking neural networks (SNNs), up to 300,000 times more energy per synaptic event in IBM's TrueNorth chip, for example. SNNs are also well-suited to exploit the precise timing of event-driven image sensors, which transmit asynchronous "events" only when the luminance of a pixel changes above or below a threshold value. The combination of event-based imagers and SNNs becomes a straightforward way to achieve low power consumption in object recognition tasks. This thesis compares different linear classifiers for two low-power, hardware-friendly, spiking, unsupervised neural network architectures, SSLCA and HFirst, in response to asynchronous event-based data, and explores their ability to learn and recognize patterns from two event-based image datasets, N-MNIST and CIFAR10-DVS. By performing a grid search of important SNN and classifier hyperparameters, we also explore how to improve classification performance of these architectures. Results show that a softmax regression classifier exhibits modest accuracy gains (0.73%) over the next-best performing linear support vector machine (SVM), and considerably outperforms a single layer perceptron (by 5.28%) when classification performance is averaged over all datasets and spiking neural network architectures with varied hyperparameters. Min-max normalization of the inputs to the linear classifiers aides in classification accuracy, except in the case of the single layer perceptron classifier. We also see the highest reported classification accuracy for spiking convolutional networks on N-MNIST and CIFAR10-DVS, increasing this accuracy from 97.77% to 97.82%, and 29.67% to 31.76%, respectively. These findings are relevant for any system employing unsupervised SNNs to extract redundant features from event-driven data for recognition.

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT REQUEST OF AUTHOR.] Camera traps are a popular tool to sample animal populations because they are noninvasive, detect a variety of species, and can record many thousands of animal detections per deployment. Cameras are typically set to take bursts of multiple images for each detection, and are deployed in arrays of dozens or hundreds of sites, often resulting in millions of images per study. The task of converting images to animal detection records from such large image collections is daunting, and made worse by situations that generate copious empty pictures from false triggers (e.g. camera malfunction or moving vegetation) or pictures of humans. We offer the first widely available computer vision tool for processing camera trap images. Our results show that the tool is accurate and results in substantial time savings for processing large image datasets, thus improving our ability to monitor wildlife across large scales with camera traps. In this dissertation, we have developed new image/video processing and computer vision algorithms for efficient and accurate object detection and sequence-level classiffication from natural scene camera-trap images. This work addresses the following five major tasks: (1) Human-animal detection. We develop a fast and accurate scheme for human-animal detection from highly cluttered camera-trap images using joint background modeling and deep learning classification. Specifically, first, We develop an effective background modeling and subtraction scheme to generate region proposals for the foreground objects. We then develop a cross-frame image patch verification to reduce the number of foreground object proposals. Finally, We perform complexity-accuracy analysis of deep convolutional neural networks (DCNN) to develop a fast deep learning classification scheme to classify these region proposals into three categories: human, animals, and background patches. The optimized DCNN is able to maintain high level of accuracy while reducing the computational complexity by 14 times. Our experimental results demonstrate that the proposed method outperforms existing methods on the camera-trap dataset. (2) Object segmentation from natural scene. We first design and train a fast DCNN for animal-human-background object classification, which is used to analyze the input image to generate multi-layer feature maps, representing the responses of different image regions to the animal-human-background classifier. From these feature maps, we construct the so-called deep objectness graph for accurate animal-human object segmentation with graph cut. The segmented object regions from each image in the sequence are then verfied and fused in the temporal domain using background modeling. Our experimental results demonstrate that our proposed method outperforms existing state-of-the-art methods on the camera-trap dataset with highly cluttered natural scenes. (3) DCNN domain background modeling. We replaced the background model with a new more efficient deep learning based model. The input frames are segmented into regions through the deep objectness graph then the region boundaries of the input frames are multiplied by each other to obtain the regions of movement patches. We construct the background representation using the temporal information of the co-located patches. We propose to fuse the subtraction and foreground/background pixel classiffcation of two representation : a) chromaticity and b) deep pixel information. (4) Sequence-level object classiffcation. We proposed a new method for sequence-level video recognition with application to animal species recognition from camera trap images. First, using background modeling and cross-frame patch verification, we developed a scheme to generate candidate object regions or object proposals in the spatiotemporal domain. Second, we develop a dynamic programming optimization approach to identify the best temporal subset of object proposals. Third, we aggregate and fuse the features of these selected object proposals for efficient sequence-level animal species classification.

Object recognition is fundamental to some high-level computer vision tasks such as image segmentation, object tracking and behavior analysis. The main objective of object recognition is to answer whether a specified object exists in a given image, so extracting representative features from images and training a right classifier become the key techniques in this area. In this paper, we use convolutional neural network model to learn features from RGB-D dataset which are then given to a linear SVM classifier to classify objects. As the number of images in RGB-D dataset is not big enough to retrain a deep neural network with high feature extraction accuracy, we fine-tune the caffe model which was trained on approximately 1.2 million RGB images from ImageNet database. While the framework of depth image is intrinsically different form RGB image, we transform the depth image into three channels and use the same method with the RGB image to extract features. We can achieve a classification accuracy of 91.35% which is much better than the state of the art.

Deep neural networks inspired by the connections among biological neurons have been highly effective in the advancement of computer vision. Recent research into recurrent neural networks has taken account of forward as well as recurrent connections in a neural network architecture. In this study, we proposed a model that mirrors the architecture of the human ventral pathway with separate layers representing brain regions connected using long-range recurrent links. CIFAR-10/100 datasets were used to assess the performance of object recognition using the proposed model and the results were compared with those obtained using state-of-the-art methods. We demonstrated that the classification accuracy increased as the number of recurrences increased. Our results suggest that the proposed neuroscience-inspired model can facilitate object recognition in computer vision and may help to elucidate neurological mechanisms in the human brain.

Introduction. The article considers the solution of the problem of detecting fires at technological facilities in automatic mode. To solve this problem, an approach was chosen to create a convolutional neural network capable of operating on a real-time video stream. Aims and Purposes. The aim of this work is to create a neural network capable of detecting flames and smoke in the image from CCTV cameras. The purposes of the work include: selection of the optimal architecture of the neural network in accordance with the latest research in this field; speeding up the chosen architecture using quantization and filter thinning techniques. Methods. Different architectures of convolutional neural networks performing the task of detecting objects in an image are considered. Their performance and quality of work are compared. The YOLOv5 architecture, its target function, training methods and ways of speeding up work are considered. Results and discu ssion. The paper shows the results of training a convolutional neural network of the YOLOv5 architecture for the task of flame and smoke detection, as well as how the results change when applying neural network acceleration methods. It was determined that the use of such acceleration methods such as quantization and filter cleaning can significantly increase up the speed of the neural network, while almost no loss in accuracy of operation. Conclusions. As a result of the conducted work, the architecture of the neural network was determined for performing the task of detecting a fire source. Based on the chosen architecture, a neural network was trained to detect flames and smoke in the image. The speed of its work allows to process video stream in real-time without using graphic accelerator.

Detection of particular meaningful objects in images is of great importance in many fields, including image retrieval or image quality analysis. In this contribution, eleven frequently observed objects (areas) in natural images are learned and detected. The presented algorithm is based on region merging and Bayesian decision theory. The main goal is not perfect recognition, since for that purpose it is necessary to use higher-level knowledge about the image content. Merging of segments proceeds only up to a reliable point, not to merge different categories. Unique merging criteria combine the values of probabilities attached to the segments for all the most likely categories, color and texture features and information about edges. Results are demonstrated on a few images and discussed.

The paper presented a novel discriminative model for efficient and effective recognition and simultaneous semantic segmentation of objects in images. The images are first segmented to give 'super-pixels'. Then the super-pixels are merged together and semantically labeled using a Condition Random Field (CRF) model. The use of a conditional random field allows us to incorporate different cues in a single unified model. The test on the standard dataset shows that compared with existing systems, the proposed system produces a detailed segmentation of a test image into coherent regions, with a semantic label associated with each region in the image.

The design of neural network models involves numerous complexities, including the determination of input vectors, choosing the number of hidden layers and their computational units, and specifying activation functions for the latter. The combinatoric possibilities are daunting, yet experience has yielded informal guidelines that can be useful. Alternatively, current research on genetic algorithms (GA) suggests that they might be of practical use as a formal method of determining ‘good’ architectures for neural networks. In this paper, we use a genetic algorithm to find effective architectures for backpropagation neural networks (BP). We compare the performance of heuristically designed BP networks with that of GA-designed BP networks. Our test domains are sets of problems having compensatory, conjunctive, and mixed-decision structures. The results of our experiment suggest that heuristic methods produce architectures that are simpler and yet perform comparatively well. © 1998 John Wiley & Sons, Ltd.

Classification task is one of the most common tasks in machine learning. This supervised learning problem consists in assigning each input to one of a finite number of discrete categories. Classification task appears naturally in numerous applications, such as medical image processing, speech recognition, maintenance systems, accident detection, autonomous driving etc.In the last decade methods of deep learning have proven to be extremely efficient in multiple machine learning problems, including classification. Whereas the neural network architecture might depend a lot on data type and restrictions posed by the nature of the problem (for example, real-time applications), the process of its training (i.e. finding model’s parameters) is almost always presented as loss function optimization problem.Cross-entropy is a loss function often used for multiclass classification problems, as it allows to achieve high accuracy results.Here we propose to use a generalized version of this loss based on Renyi divergence and entropy. We remark that in case of binary labels proposed generalization is reduced to cross-entropy, thus we work in the context of soft labels. Specifically, we consider a problem of image classification being solved by application of convolution neural networks with mixup regularizer. The latter expands the training set by taking convex combination of pairs of data samples and corresponding labels. Consequently, labels are no longer binary (corresponding to single class), but have a form of vector of probabilities. In such settings cross-entropy and proposed generalization with Renyi divergence and entropy are distinct, and their comparison makes sense.To measure effectiveness of the proposed loss function we consider image classification problem on benchmark CIFAR-10 dataset. This dataset consists of 60000 images belonging to 10 classes, where images are color and have the size of 32×32. Training set consists of 50000 images, and the test set contains 10000 images.For the convolution neural network, we follow [1] where the same classification task was studied with respect to different loss functions and consider the same neural network architecture in order to obtain comparable results.Experiments demonstrate superiority of the proposed method over cross-entropy for loss function parameter value α < 1. For parameter value α > 1 proposed method shows worse results than cross-entropy loss function. Finally, parameter value α = 1 corresponds to cross-entropy.

Abstract. This study discusses the problem of phytoplankton classification using computer vision methods and convolutional neural networks. We created a system for automatic object recognition consisting of two parts: analysis and primary processing of phytoplankton images and development of the neural network based on the obtained information about the images. We developed software that can detect particular objects in images from a light microscope. We trained a convolutional neural network in transfer learning and determined optimal parameters of this neural network and the optimal size of using dataset. To increase accuracy for these groups of classes, we created three neural networks with the same structure. The obtained accuracy in the classification of Baikal phytoplankton by these neural networks was up to 80%.

Determining the best model for the neural network architecture and how to optimize the architecture with the metaheuristic Protis Approach is a subject of the study. A comprehensive investigation and utilization of metaheuristic methods are necessary. These methods aim to solve problems and adapt from the lifestyle of the amoeba protis. In this study, the proposed method modifies the life cycle of the amoeba, which consists of four phases: prophase, metaphase, anaphase, and telophase. These four phases are modified in the neural network architecture to optimize the appropriate number of hidden layers and produce an efficient architecture model. The results show that the protis approach optimized the neural network architecture, especially in generating hidden layers to improve the neural network model. Distinctive features of the results obtained are that the average range of degenerate neurons in the hidden layer is 0 to 35 neurons in each layer. The standard number of neurons makes it possible to solve the problem of determining the best model on the neural network architecture. The protis algorithm embedded in the protis recurrent neural network for categorical data measurements produces an average RMSE value, representing the difference between actual measurements and predictions, equal to 0.066. Consequently, the developed model surpasses the current classical neural network model in terms of performance. Regarding accuracy, the protis algorithm embedded in the neural network for categorical and time series data achieves an average precision of 0.952 and a recall of 0.950. The protis convolutional neural network achieves an accuracy of 95.9 %. Therefore, from the three tested datasets, the protis convolutional neural network exhibits the highest accuracy value

The study presents the results of the development and testing of deep learning neural network architectures, which demonstrate high accuracy rates in classifying neurophysiological data, in particular hemodynamic brain activation patterns obtained by functional near-infrared spectroscopy, during solving mathematical problems on spatial-numerical associations. The analyzed signal represents a multidimensional time series of oxyhemoglobin and deoxyhemoglobin dynamics. Taking the specificity of the fNIRS signal into account, a comparative analysis of 2 types of neural network architectures was carried out: (1) architectures based on recurrent neural networks: recurrent neural network with long short-term memory, recurrent neural network with long short-term memory with fully connected layers, bidirectional recurrent neural network with long short-term memory, convolutional recurrent neural network with long short-term memory; (2) architectures based on convolutional neural networks with 1D convolutions: convolutional neural network, fully convolutional neural network, residual neural network. Trained long short-term memory recurrent neural network architectures showed worse results in accuracy in comparison with 1D convolutional neural network architectures. Residual neural network (model_Resnet) demonstrated the highest accuracy rates in three experimental conditions more than 88% in detecting age-related differences in brain activation during spatial-numerical association tasks considering the individual characteristics of the respondents’ signal.

Object classification has been one of the main tasks in computer vision. With the fast development of deep learning, its performance in image classification and object recognition has presented dramatic improvements. In this paper, we propose a new deep convolutional neural network (CNN) architecture for robust object classification. The proposed model is fused with three traditional CNN approaches, Densenet201, Resnet50, and our proposed residual CNN. The fused network architecture allows parallel processing of the multiple networks for keeping the system sped up. A single shot deep convolution network is trained as an object detector to generate all possible candidates of different object classes. The output of each neural network is representing a single vote that is used in the classification process. 3-to-1 voting criteria are applied in the final classification decision between the candidate object classes. Several experiments were conducted to evaluate the performance of the proposed network. The experimental results show that the proposed approach has better performance than the networks used in the fusion process when they act individually. It also has lower miss rates when compared to several states of art methodologies.

Previous approaches for scene text detection have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with challenging scenarios, even when equipped with deep neural network models, because the overall performance is determined by the interplay of multiple stages and components in the pipelines. In this work, we propose a simple yet powerful pipeline that yields fast and accurate text detection in natural scenes. The pipeline directly predicts words or text lines of arbitrary orientations and quadrilateral shapes in full images, eliminating unnecessary intermediate steps (e.g., candidate aggregation and word partitioning), with a single neural network. The simplicity of our pipeline allows concentrating efforts on designing loss functions and neural network architecture. Experiments on standard datasets including ICDAR 2015, COCO-Text and MSRA-TD500 demonstrate that the proposed algorithm significantly outperforms state-of-the-art methods in terms of both accuracy and efficiency. On the ICDAR 2015 dataset, the proposed algorithm achieves an F-score of 0.7820 at 13.2fps at 720p resolution.