Background Modeling by Weightless Neural Networks

Weightless neural networks (WNNs) are an alternative pattern recognition technique where RAM nodes function as neurons. As both training and inference require mostly table lookups, few additions, and no multiplications, WNNs are suitable for high-performance and low-power embedded applications. This work introduces a novel approach to implement WiSARD, the leading WNN state-of-the-art architecture, completely eliminating memories and arithmetic circuits and utilizing only logic functions. The approach creates compressed minimized implementations by converting trained WNN nodes from lookup tables to logic functions. The proposed LogicWiSARD is implemented in FPGA and ASIC technologies to illustrate its suitability for edge inference. Experimental results show more than 80% reduction in energy consumption when the proposed LogicWiSARD model is compared with a multilayer perceptron network (MLP) of equivalent accuracy. Compared to previous work on FPGA implementations for WNNs, convolutional neural networks, and binary neural networks, the energy savings of LogicWiSARD range between 32.2% and 99.6%.

[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.

This study presents an innovative method utilizing weightless neural networks (WNNs) to identify and address various types of faults in air compressor modules. Random access memory (RAM) devices are harnessed by WNNs to emulate the functioning of neurons. The training process employs a versatile and effective algorithm aimed at generating reliable and accurate results. One notable benefit of employing WNNs is its ability to eliminate the necessity for network retraining and the generation of residuals. This feature makes WNN suitable for applications related to classification and pattern recognition. In this study, a specific type of air compressor, namely a single-acting single-stage reciprocating one, was chosen. Various potential faults like fluttering inlet and outlet valves, valve plate leakage, and check valve issues were taken into account. From the initial vibration data, statistical, histogram, and autoregressive moving average features were derived. For efficiency, the J48 decision tree algorithm was utilized to identify the pivotal features in this investigation. Following this, the features were divided into separate sets to evaluate the validation, training, and testing accuracies of the WNNs using the WiSARD classifier. Additionally, fine-tuning of hyperparameters was done to enhance classification accuracy while simultaneously reducing computational time. The results obtained demonstrate that, with the specified hyperparameter configurations, the WiSARD classifier attained an accuracy of 98.6667% for statistical features. The proposed method outperforms existing approaches, showing potential for real-time application in enhancing air compressor lifespan, reliability, and safety.

AI applications have recently driven the computer architecture industry towards novel and more efficient dedicated hardware accelerators and tools. Weightless Neural Networks (WNNs) is a class of Artificial Neural Networks (ANNs) often applied to pattern recognition problems. It uses a set of Random Access Memories (RAMs) as the main mechanism for training and classifying information regarding a given input pattern. Due to its memory-based architecture, it can be easily mapped onto hardware and also greatly accelerated by means of a dedicated Register Transfer-Level (RTL) architecture designed to enable multiple memory accesses in parallel. On the other hand, a straightforward WNN hardware implementation requires too much memory resources for both ASIC and FPGA variants. This work aims at designing and evaluating a Weightless Neural accelerator designed in High-Level Synthesis (HLS). Our WNN accelerator implements Hash Tables, instead of regular RAMs, to substantially reduce its memory requirements, so that it can be implemented in a fairly small-sized Xilinx FPGA. Performance, circuit-area and power consumption results show that our accelerator can efficiently learn and classify the MNIST dataset in about 8 times faster than the system's embedded ARM processor.

Physically unclonable functions (PUFs) have been explored as lightweight hardware primitives for the purpose of realizing robust security via strong authentication or secure key/ID generation. PUF harness manufacturing process variations for the purpose of generating binary keys or binary functions. An ideal strong PUF is a binary function that maps an m-bit input challenge to an uniquen-bit output response, making it attractive for authentication applications. Unfortunately, real strong PUF implementations suffer from reliability issues where the same challenge may produce different responses in the presence of noise. To overcome this problem, strong PUF leverages the availability of exponential number of challenge-response pairs (CRPs). A successful authentication event requires acquiring multiple CRPs and applying a threshold. In contrast, weak PUFs produce limited keys and are required to be highly reliable. Multiple techniques have been developed to achieve the necessary reliability. An additional prerequisite for strong PUFs is resilience against model-building attacks (cloning) by an adversary, who has observed a few CRPs, to prevent successful prediction of future CRPs. In this work, we first illustrate a strong PUF design that re-purposes a weightless neural network (WNN). Second, we showcase the robustness of WNN-based strong PUFs with respect to machine learning attacks, while providing desirable uniqueness and reliability metrics. Finally, we employ an initial entropy source of highly reliable weak PUF bits mapped to weightless neural networks (WNNs) for the purpose of creating a near-ideal strong PUF in terms of reliability. Our results show that it is possible to create highly reliable WNN–based strong PUFs with < 65 % ML accuracy by using as few as 32 initial reliable weak PUF bits.

This proposal presents a novel use of Weightless Neural Networks (WNN) and Steinbuch Lernmatrix for pattern recognition and classification. High speed of learning, easy of implementation and flexibility given by WNN, combined with the learning capacity, recovery efficiency, noise immunity and fast processing shown by Steinbuch Lernmatrix are key factors considered on the pattern recognition exposed by the suggested model. For experimental purposes, the fundamental pattern sets are built and provided to the model under the learning phase. The additive, subtractive and mixed noises are applied to fundamental patterns to check out the response of the model during the recovery phase. Field Programmable Gate arrays are used in the implementation of such model, since it allows custom user-defined models to be embedded in a reconfigurable hardware platform, and provides block memories and dedicated multipliers suitable for the model.

Increasingly advanced image analysis in computer vision, allowing computers to interpret, identify, and analyze pictures with accuracy comparable to humans. The availability of data sources in decimal, hexadecimal, or binary forms enables researchers to take the initiative in applying their study findings. Decimal formats are typically used on traditional computers like desktops and minicomputers, whereas hexadecimal and binary formats were utilized on single-chip controllers. Weightless Neural Network is a method that can be implemented in a single chip controller. The aim of this research is to develop a facial recognition system, for eye and mouth identification, that works in a single chip controller or also called a microcontroller. The suggested method is a Weightless Neural Network with Immediate Scan approach for processing and identifying eye and nose patterns. The data will be handled in many memory locations that are specifically designed to handle massive volumes of data. The data is made up of primary face data sheets and face input data. The data sets utilized are (x,y) pixels, and frame sizes range from 90x90 pixels to 110x110 pixels. Each face shot will be processed by selecting the region of the eyes and nose and saving it as an image file. The eye and nose will identify the face frame. Next, the photos will be converted to binary format. A magazine matrix will be used to transmit binary data from a minicomputer to a microcontroller via serial connection. Based on a known pattern, the resultant similarity accuracy is 83,08% for the eye and 84,09% for the sternum. In contrast, the similarity percentage for an eye ranges from 70% to 85% for an undefined pattern.

The segmentation of moving objects in image sequence can be formulated as a background subtraction problem—the separation of objects from the background in each image frame. The background scene is learned and modeled. A pixelwise process is employed to classify each pixel as an object or background based on its similarity with the background model. The segmentation is challenging due to the occurrence of dynamic elements such as illumination change and background motions. We propose a framework for object segmentation with a novel feature for background representation and new mechanisms for updating the background model. A ternary pattern is employed to characterize the local texture. The pattern and photometric features are used for background modeling. The classification of pixel is performed based on the perceptual similarity between the current pixel and the background model. The segmented object is refined by taking into account the spatial consistency of the image feature. For the background model update, we propose two mechanisms that are able to adapt to abrupt background change and also merge new background elements into the model. We compare our framework with various background subtraction algorithms on video datasets.

Optimising memory usage in n-tuple neural networks

In this paper we propose a new face recognition method based on the weightless neural network system [1]. The algorithm uses 5-pixel n-tuples to map images, which passes through a ranking transform to obtain a binary n-tuple state. A digital neural network correlates the recurring states obtained from the current input pattern to those extracted from the test set. The data used in this paper is from the MIT-CBCL facial database [2], and the training data and testing data set each consist of 10 individual persons, with 100 examples of each subject. An error rate of 0.1% FAR and 0.1% FRR was achieved on data which was totally independent of the training set.KeywordsFace RecognitionFalse Acceptance RateFalse Rejection RateFace Recognition MethodIEEE Signal Processing SocietyThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Two different types of adaptive networks are considered for solving the centralized and distributed hy-pothesis testing problem. The performance of the two different types of networks is compared under differentperformance indices and training rules. It is shown that training rules based on the Neyman-Pearson criterion outperform error based training rules. Simulations are provided for data that are linearly and nonlinearly separable. I. INTRODUCTION The optimum Bayesian and Neyman-Pearson solution to the distributed decision fusion problem bearsstriking similarities to the structure of a neural network (NN), [28,29]. Moreover, NNs can, in principle learn arbitrary input-output mappings, provided that they are sufficiently smooth. These two facts motivate theuse of NNs for solving the centralized and distributed hypothesis testing problem. In selecting the properNN layout, one could argue that a perceptron-type NN can learn any input-output mapping, thus it can betrained to solve the hypothesis testing problem. However, the ability of a perceptron-type NN to learn an

A new background estimation and suppression algorithm was presented. In the algorithm, targets and observing noises were considered as mixed interferences of the image background. With this situation, image background was estimated adaptively and then background suppression was done in order to improve the signal-to-noise ratio (SNR) of targets. In this algorithm, firstly, a Zernike-facet model of image background was built up. Secondly, the total least squares (TLS) method was used to solve parameters of the model. Finally, background estimation and suppression were done using the model and its parameters. Simulations and several experiments demonstrating the effectiveness of this proposed algorithm were reported. And results show that this algorithm can be effective to estimate background in mixed noise environment and can preserve detail information of targets and improve SNR of targets. As a result, detecting probability and false probability will be improved in next process for automatic target detection and tracking.

Background estimation is one of the most challenging phases in extracting foreground objects from video sequences. In this paper we present a background modeling approach that uses the similarity of frames to extract background areas from the video sequence. We use a window over the frames history and compute the similarity between the selected frames of this window as a similarity window. The properties of similarity window depend on the characteristics of the scene and can be adjusted parametrically. Our primary results show that if proper parameters are chosen, this method can give a good approximation of the background model.

Automated traffic surveillance systems are widely used in intelligent transportation systems (ITS). However, the accuracy of video-based Vehicle Detection is heavily affected by complex environmental factors such as shadows, rain, illumination and glare. This paper introduces an approach to motion vehicle detection in complex condition over highway surveillance video. The framework is composed of two parts: background estimation and multi-feature extraction. A fast constrained Delaunay triangulation (CDT) algorithm based on constrained-edge priority is presented instead of complicated segmentation algorithms. We present a background block update modeling theory based on triangulation to estimate background self-adaptively. Consequently, we can get the difference between the current frame and the background model. After extracting features in triangular candidates, multi-feature eigenvector is created for each vehicle with Principal Component Analysis (PCA). We design a classifier to classify triangular candidate as a part of a real vehicle or not by Support Vector Machine (SVM). And then, a parallelogram is used to represent the vehicle's shape robustly. Finally, experiments using real video sequence are performed to verify the method proposed for complex environmental factors.

The ability to accurately predict natural backgrounds over a wide range of potential image acquisition conditions is key to wide-area change detection. The lack of robustness in pure image-based techniques has led to the development of site modeling and model-supported exploitation approaches. Instead of using multiple images to build a site model, we consider the possibility of using them directly to estimate the background of a new image for change detection. Our method uses a nonlinear mean-squared estimation technique to compute image backgrounds from multiple reference images. In comparing linear and nonlinear estimators, we find that although the performance of both improves as the number of reference images increases, the nonlinear estimator yields significantly better background estimates and is much less sensitive to registration errors between images than the linear estimator. We show that as the number of references increases, the pixels within a region become physically significant; i.e., they correspond to parts of the scene with similar physical properties. Synthetic aperture radar (SAR) background modeling and change detection results including a crosssensor change detection example in which SAR reference images are used to predict the background of an electro-optical (EO) image are presented. Significant performance gains over linear change detection (up to two orders of magnitude reduction in the false alarm rate) are demonstrated in each case over a 60-deg range of SAR sensor aspect angles.