Search

Over the past few years, neural networks have exhibited remarkable results for various applications in machine learning and computer vision. Weight initialization is a significant step employed before training any neural network. The weights of a network are initialized and then adjusted repeatedly while training the network. This is done till the loss converges to a minimum value and an ideal weight matrix is obtained. Thus weight initialization directly drives the convergence of a network. Therefore, the selection of an appropriate weight initialization scheme becomes necessary for end-to-end training. An appropriate technique initializes the weights such that the training of the network is accelerated and the performance is improved. This paper discusses various advances in weight initialization for neural networks. The weight initialization techniques in the literature adopted for feed-forward neural network, convolutional neural network, recurrent neural network and long short term memory network have been discussed in this paper. These techniques are classified as (1) initialization techniques without pre-training, which are further classified into random initialization and data-driven initialization, (2) initialization techniques with pre-training. The different weight initialization and weight optimization techniques which select optimal weights for non-iterative training mechanism have also been discussed. We provide a close overview of different initialization schemes in these categories. This paper concludes with discussions on existing schemes and the future scope for research.

The propagation loss varies with underwater channel conditions which might considered to be random phenomena. Modeling propagation loss will become meaningful iff mathematical model includes parameters namely viz projector's transmit voltage response (TVR), hydrophone's open circuit receiving response (OCRR), directivity patterns of both, channel parameters such as salinity, temperature, pressure, enclosure boundary conditions along with placements of Tx & Rx nodes & their operating frequency. To best of our knowledge, existing simulators are unable to trace eigen rays for very short range i.e. less than 0.1 km and therefore they are not suitable for computation of such short-range propagation losses. We have made an attempt to overcome limitations of existing simulators wherein we proposed mathematical model SWARA which includes parameters as mentioned above to study very short-range propagation losses using plane wave theory. To validate simulated propagation loss, we conducted tank trials at UWAA Lab, CARE, IIT Delhi to investigate effects of placements of projector & hydrophone on occurrence of transmission loss. The Simulated results of SWARA mathematical model shows that simulated maximum transmission loss is -0.18 to 0.10 times experimental maximum transmission loss, whereas simulated minimum transmission loss is -0.36 to 0.19 times experimental minimum transmission loss for placements of projector (ITC 1042) and hydrophone (Keltron 8240000001) at depths varying from 0.3m-1.2m & range varying from 2m-3.2m in uw tank facility of 3.85m long 2.4 wide 2m deep for 30kHz chirp signal (10kHz bandwidth) under static channel conditions.

Automatic License Plate Recognition (ALPR) is a very widely used system in applications such as parking management, theft detection, traffic control and management etc. Most of the existing ALPR systems fail to showcase acceptable performance on real time images/video scenes. This work proposes and demonstrates implementation of a deep learning-based approach to locate license plates of four wheeler vehicles thereby enabling optical character recognition (OCR) to recognize the characters and numbers on the located plates in real time. The proposed system is decomposed into three sub-blocks viz. Vehicle image/video acquisition, License plate localization and OCR. A simple setup using a reasonable resolution webcam has been designed to capture images/videos of vehicles at some entry point. We propose to utilize Single Shot Detector (SSD) based Mobilenet V1 architecture to localize the license plates. The hyper parameters of this architecture are selected with rigorous experimentation so as to avoid over-fitting. We have compared performance of two OCRs viz. Tesseract OCR, Easy OCR and found the superiority of Easy OCR since it utilizes deep learning approach for character recognition. NVIDIA Jetson Nano and Raspberry Pi 3B hardware platforms have been used to implement the entire system. The parameters of these three sub-blocks have been optimized to yield real time performance of ALPR with acceptable accuracy. The proposed and implemented system on Jetson Nano allows processing of videos for ALPR having accuracy more than 95%.

The first case of the novel Coronavirus disease (COVID-19) in India was recorded on 30th January 2020 in Kerela and it has spread across all states in India. The prediction of the number of COVID-19 cases is important for government officials to plan various control strategies. This paper presents a weekly prediction of cumulative number of COVID-19 cases in India. A graded lockdown feature, which describes the status of lockdown, is derived and incorporated in the input dataset as one of the features. For prediction, this paper proposes a model which is a stacking of different deep neural networks which have recurrent connections. Vanishing gradients is a common issue with such networks with recurrent connections. Proper weight initialization of the network is one of the solutions to overcome the vanishing gradients problem. Hence, the weight distributions and convergence performance of some state-of-the-art weight initialization techniques have been analyzed in this paper. The proposed model is initialized with the technique which would aid to avoid the vanishing gradients problem and converge faster to a lower loss. This paper also provides a comparison of the proposed model for univariate and multivariate prediction with other prediction models such as statistical model - Auto-Regressive Integrated Moving Average (ARIMA), and deep learning architectures long short term memory (LSTM), bidirectional LSTM (bi-LSTM) and gated recurrent unit (GRU). The results demonstrate that the proposed model gives better prediction results than these models.

The WHO has recommended ‘frequent hand washing’ as means to curtail the spread of ‘Public Health Emergencies of International Concern.’ Improvement in the seven step hand wash compliance rate has been shown to reduce the spread of hospital acquired infections. Most of the hand hygiene compliance identification systems developed over the years have restricted their focus on tracking the movement of healthcare workers to and from the hand wash station. However, these systems have failed to detect if the seven step hand wash were performed or not. We proposed and implemented a computer vision and artificial intelligence based system to detect seven steps of the hand wash process. We used the Visual Geometry Group-16 (VGG-16) network combined with the Long Short Term Memory (LSTM) module as a classification system. We developed the hand wash database of 3000 videos to train and optimize the parameters of the VGG16-LSTM model. The optimized model detects different steps of handwash with high accuracy and near real time detection ability. This system will prove to be useful for improving hand wash compliance rate and to curb the spread of infectious diseases.

Convolutional neural networks have delivered exceptional performance in various areas of computer vision. There has been growing research to develop deeper architectures with the availability of large datasets. Training such deep networks on large datasets is a tedious process as it involves optimizing a loss function by updating the parameters of the network. Weight initialization is a vital step before training neural networks as the correct choice of network weights ensures that the optimization converges to global minima in the least time. The weight initialization strategies in the literature can be categorized as (1) Initialization without pre-training, and (2) Initialization with pre-training. This paper presents a comparative analysis of the convergence performance of some widely used weight initialization techniques in these categories. This analysis is based on the diversity insights measured in terms of mean standard deviation captured from the feature maps. The experimentation has been carried out by training the AlexNet and VGG16 network on CIFAR-10 and CIFAR-100 datasets. The experimentation results demonstrate that the He initialization technique, which shows the best convergence performance among the others considered for the study, leads the training process such that the diversity of feature maps increases with epochs for both AlexNet and VGG16 network.

The performance of Minimum Variance Distortionless Response (MVDR) beamformer can be improved by enhancing its ability to suppress interference and noise effectively. Generally, the number of microphones should be large in order to get greater interference suppression. However, this decreases the stability of a beamformer. To improve stability a diagonal loading factor is added in the noise covariance matrix. An increase in loading factor causes poor suppression and larger deviations in the values of the loading factor mislead the steering vector in another direction. As the exact noise covariance matrix is unknown, it is estimated from the interference and noise. In this paper, the issue of estimation error is addressed. The amount of diagonal loading is estimated adaptively by considering actual snapshots of the input signal. An additional constraint on the diagonal loading is proposed, which improves the robustness and stability of the MVDR beamformer. A tradeoff between the stability and diagonal loading factor is investigated. The effect of adaptive diagonal loading on output SINR and stability is analyzed.

A simple and fast technique is presented in this paper for Indian license plate recognition system. Using OpenALPR’s software framework and RaspberryPi, a real-time Indian license plate recognition system could be implemented for some application specific purposes. HAAR and LBP features are extracted from the acquired vehicle images and subjected to training of cascade classifiers in order to localize license number plates. The validation process is used to optimally select number of stages of cascade classifiers. The extracted number plates are then utilized for character recognition. Cascade classifier with LBP features is suitable for localization of license plates with accuracy more than 98%. Whereas, the average number plate recognition accuracy is above 96% for images captured from front side. The proposed system has been prototyped using C++ and RaspberryPi 3 and experimental results have been shown for recognition of Indian license plates

In the last two decades, in the domain of video coding and compression, researchers have suggested several techniques for computation and time reduction for motion estimation (ME). We present a motion estimation algorithm for x265 video codec, based on a deterministic initial population in the genetic algorithm (GA). GA is known for its adaptive convergence, which is motivated by the biological process of survival of the fittest. The suggested scheme is targeted for the reduction of search points (SP) in a block matching motion estimation algorithm for integer-pel in B and P frames that are set to have three reference frames. The initial population constituted in our approach is a function of pre-encoded coding units at different spatial–temporal locations of the video frames and predefined hexagonal (HEX) locations. We propose a “deterministically starting” GA (GADet), toward deployment in x256 structure. In the framework of x265 code, GADet is found to offer reduction in SP at selected classes of videos considered for experimentation. To demonstrate the effectiveness of the proposed work, results have been compared with the block-based fast-full-search algorithm and the HEX search algorithm from the reference software. Traditional GA with a randomly constituted initial population, labeled as GAStc, is also implemented and an empirical comparison is carried out with GADet. The proposed GADet framework provides reduction in motion estimation time while rendering acceptable peak signal-to-noise ratio loss and an increase in a bit rate.

The parallel iterative algorithms are the major advancements in the field of computing. These algorithms lead to efficient usage of hardware as well as obtaining faster results. In this paper, we describe architecture to compute eigen values and eigen vectors of a matrix having dimensions up to 50 × 50 using cyclic Jacobi's Algorithm. Systolic array architecture is used to apply it to matrices of larger dimensions. We have implemented the architecture on FPGA Vertex-5 that takes about 8059 LUT slices out of 69120 slices for matrices of dimensions 50 × 50.

Textures can be broadly divided into two categories, namely, stochastic and deterministic. The stochastic textures are characterized by its statistical properties and do not have easily identifiable primitives. Even if one can extract such primitives, a placement rule description for such textures may be extremely complicated. One of the ways to describe and generate such textures is Simultaneous Autoregressive (SAR) Linear Prediction models. The major difficulty in utilizing this model is choosing proper neighborhood locations, within which pixels are considered interdependent. The work presented here emphasizes the automatic neighbor location selection based on sample correlation function. The algorithm is stretched to the possible extent with rigorous experimentation that proves the decorrelation phenomenon with residual image. Two new methods viz. ‘actual residual image’ and ‘uniform noise transformed to a noise with histogram matched to residual image’ are suggested to synthesize texture towards perceptual quality improvement.

Structural textures are characterized by repeating pattern called ‘Texton’ and placement rule — that determines the nature of periodicity. Based on periodicity, textures are classified as homogeneous — perfectly periodic and weakly homogeneous — quasi-perioidic. Both of these textures are assumed to be combination of Structural information, Illumination i.e. average brightness at different sites of the texture and Stochasticity to allow local variations. This paper deals with a top-down approach to extract structural information i.e. the grid, representative texton and illumination component from the original texture patch. Introduction of stochasticity makes the texture more natural and similar to original one. This technique does not produce verbatim copies in the synthesized texture. Since the representative texton generation, texton filling and stochasticity introduction are computationally heavier, the work presented in this paper claims the algorithmic development towards improvement in overall synthesis time.

Texture is a fundamental characteristic in many natural images that plays an important role in human visual perception and in turn provides information for image understanding and scene interpretation. The textured image can be modeled to describe, analyze and synthesize the texture. The model parameters capture the essential perceived qualities of texture. One of the important characteristics of texture data is the statistical dependence of the gray level at a lattice point on those of its neighbors. The spatial-interaction models characterize this statistical dependency by representing the intensity of a pixel, as a 2-D linear combination of the intensity of its neighbors and an additive noise. One way of specifying this interaction is simultaneous autoregressive (SAR) models. This is one of the most traditional methods used for modeling in the area of image processing. This paper presents the work done by the authors on parameter estimation and synthesis of textured images using Simultaneous Autoregressive (SAR) modeling. Different programs are developed in MATLAB to implement the parameter estimation and synthesis and are tested for their performance. The scope of this work includes the use of causal and noncausal methods for modeling and synthesizing natural textures. Simultaneous or spatial autoregressive models with causal and noncausal neighborhoods are used for parameter estimation and texture pattern generation. Parameter estimation is done by two different methods: The least square error (LSE) and maximum likelihood estimation (MLE). LSE method is preferred for causal models. MLE method is used for noncausal autoregressive models and it uses iterative algorithm. The synthesis procedure is based on generating a two dimensional autoregressive random field driven by a two dimensional zero mean white noise field with unit variance. Two different algorithms are used for synthesis of causal and noncausal AR models. Different image textures are synthesized using a given set of neighborhoods and parameters. Different patterns of synthetic images can be generated using various sets of parameters. A number of images from Brodatz album are tested for parameter estimation and synthesis. The synthesized image retains the pattern in the original image like vertical or horizontal streaks. An interactive graphical user interface (GUI) is developed using MATLAB that allows user to select one image from Brodatz album. The user can choose between causal or noncausal neighborhood and select number of elements in the neighborhood or choose any one set of neighborhood from different sets stored, and find out SAR model parameters by one of the methods of parameter estimation. The user can synthesize the image using these parameters. Both original and synthesized image are displayed side by side on the screen and the user can easily compare the two images. Thus the GUI offers an interactive platform for implementation of parameter estimation using different neighborhoods for the images from Brodatz album and synthesis of the image from these estimated parameters.

This paper deals with texture feature extraction operators, which comprise linear filtering eventually followed by post processing. Robust, rotation and scale invariant texture operators are important for digital image libraries and multimedia databases. A method of rotation and scale-invariant texture classification based on a log polar coordinate system is introduced. Texture is an important clue in region based segmentation of images. Here, we provide analysis and implementation of a set of distortion invariant texture operators viz circular Mellin features (CMF). The CMF represent the spectral decomposition of the image scene in the polar log coordinate system and are invariant to both scale and orientation of the target texture pattern. The image and CMF are correlated followed by magnitude detection based on thresholding. The CMF extractors have a functional form that is similar to Gabor functions; they have distortion invariant characteristics, unlike Gabor functions, which makes them more suitable for texture segmentation.

Structural textures are characterised by a repeating pattern called `Texton` and placement rule - that determines the nature of periodicity. Based on the periodicity, textures are classified as homogeneous - perfectly periodic - and weakly homogeneous - quasi-periodic. Both of these textures are assumed to be combination of structural information, illumination, that is, average brightness at different sites of the texture and stochasticity to allow local variations. A top-down approach extracts structural information, that is, the grid, representative texton and illumination component from the original texture patch and then the information is used to synthesise similar textures. Obviously, the technique does not produce ditto copies in the synthesised texture. The experimentation is carried out to improve quality of synthesised textures by incorporating multiple representative textons. Also a parameter, namely homogeneity co-efficient (HC), is suggested to compare the original texture patch and synthesised texture. The parameter captures variations in the textons, contents and sizes both, and thus can be used to compare the synthesis results. The suitability of the proposed synthesis approach and HC is verified by rigorous experimentation on weakly homogeneous artificial and standard structural textures. Efforts are also made to utilise the multi-core processing capability of the processor to improve the speed of analysis and synthesis phases.