In standard neural network training, the gradients in the backward pass are determined by the forward pass.As a result, the two stages are coupled.This is how most neural networks are trained hitherto.Gradient modification in the backward pass has seldom been studied in the literature.In this paper we explore decoupled training, where we alter the gradients in the backward pass.We propose a simple yet powerful method called PowerGrad Transform (PGT), that alters the gradients before the weight update in the backward pass and significantly enhances the predictive performance of a convolutional neural network.PGT trains networks to arrive at a better optima at convergence.It is computationally efficient, and adds no additional cost to either memory or compute, but results in improved final accuracies on both the training and test datasets.Power-Grad Transform is easy to integrate into existing training routines, requiring just a few lines of code.With decoupled training, our method improves baseline accuracies for ResNet-50 by 0.73%, for SE-ResNet-50 by 0.66% and by more than 1.0% for the non-normalized ResNet-18 network on the ImageNet classification task.

The authors propose a method to recognise facial expressions based on graph signal processing (GSP) techniques. Facial expressions are characterised by local patterns in the facial regions such as eyes, lips etc. and interrelationships among them. A facial expression recognition algorithm needs to capture these variations in the facial regions at the local level and the interrelationships of these regions at the global level. Hence, in the authors’ opinion, GSP seems to be an appropriate tool for the purpose. In this study, a novel method is presented which makes use of graph signals to represent the facial regions. They leverage spectral graph wavelet transform to extract information for creating the feature descriptor. Here, different types of two‐channel and three‐channel filter banks have been used by setting the weights of their channels for finding the optimum performance of the recognition rate. Through simulation studies, it is observed that the use of Abspline filter bank provides the best result. The experimental investigations on the extended Cohn–Kanade (CK+) and the JAFFE datasets have been carried out and the results confirm the effectiveness of the proposed method in recognition rate improvement.

In 1918, S. Ramanujan defined a family of trigonometric sums now known as Ramanujan sums. In this letter, we define a class of operators based on the Ramanujan sums termed here as Ramanujan class of operators. We then prove that these operators possess properties of first derivative and with a particular shift, of second derivative also. Applications of Ramanujan class of operators for edge detection and noise level estimation are also demonstrated.

This paper presents anew, simple, and efficient strategy for locomotion classification using a force myography (FMG) system developed in-house. The FMG signals were acquired from eight calibrated force sensors. These sensors were wrapped around the thigh and captured volumetric changes in thigh muscles during locomotion. The data were collected from eight able-bodied subjects and two unilateral transfemoral amputees while walking in different terrains, i.e., level walk, ramp ascent, ramp descent, stair ascent, and stair descent (SD). The repeatability of FMG signals within all the five locomotion modes was expressed in terms of variance ratio (VR). For all locomotion modes, except SD, the VR ranges from 0.1 to 0.3, showing high repeatability within the locomotion modes for able-bodied and amputee subjects. Furthermore, these locomotion modes are characterized by computationally simplified time-domain features of FMG signals and classified using linear discriminant analysis (LDA). The subject-dependent LDA model gave an average classification accuracy of 99.5% (SD = 0.5) and 96.1% (SD = 1.9) for able-bodied and amputee subjects, respectively. These results show the potential of the FMG system in a real-time classification of locomotion modes for lower-limb prosthetic control.

A discrete-time fBm (dfBm) process BH(n) is a Gaussian, zero mean, non-stationary, statistically self similar random process with self-similarity index H (Hurst exponent). Signal modeling via these processes has been used in many engineering applications. In this paper, we have shown that the orthogonal matrix Q that diagonalizes the auto-covariance matrix of 1st order dfBm is close to the columns of DCT matrix.

In the network data, interarrival times are heavy-tail distributed. Weibull, Pareto and Lognormal are the best examples of heavy-tail distributions. These distributions give time-varying arrival rates. Also, renewal processes for these interarrivals are non-homogeneous Poisson processes. The network traffic is, thus, non-stationary. Many of the theoretical tools, such as equilibrium probabilities for Markov chains, matrix geometric solutions and Laplace transforms are not available for queue with time varying rates. Since no closed form expressions of Laplace transform of Weibull, Pareto and Lognormal distributions are available, the queueing analysis becomes complicated. In this paper we present queueing analysis of heavy-tail network traffic, and thus, time dependent queuing model Mt/G/¿ is analyzed for heavy-tail Pareto arrivals. In particular, we find analytical expressions for the time-dependent mean function (offered load), denoted here as m(t), which depends on the time-dependent arrival rate function ¿(t) and service time distribution.

In this paper, we have proposed a novel scheme for the extraction of textual areas of an image using globally matched wavelet filters. A clustering-based technique has been devised for estim ating globally matched wavelet filters using a collection of groundtruth images. We have extended our text extraction scheme for the segmentation of document images into text, background, and picture components (which include graphics and continuous tone images). Multiple, two-class Fisher classifiers have been used for this purpose. We also exploit contextual information by using a Markov random field formulation-based pixel labeling scheme for refinement of the segmentation results. Experimental results have established effectiveness of our approach.

In this paper we have proposed a novel scheme for locating text regions in an image. The method is based on multiresolution wavelet analysis. We used matched wavelets to capture textural characteristics of image regions. A clustering based approach has been proposed for estimating globally matched wavelets (GMWs) for a given collection of images. Using these GMWs, we generate feature vectors for segmentation and identification of text regions in an image. Our method, unlike most of the other methods, does not require any a priori information about the font, font size, scripts, geometric transformation, distortion or background texture. We have tested our method on various categories of images like license plates, posters, hand written documents and document images etc. The results show proposed method to be a robust, versatile and effective tool for text extraction from images.

We consider an antipodal signaling system using wavelet modulation with scale diversity over a frequency-selective fading channel. The channel introduces a random gain for each scale. These scale gains are estimated by sending a block of pilot symbols. Coherent demodulation is carried out along with maximal-ratio combining using the estimated scale gains. From the characteristic function of the decision variable, we analyze the error performance. For the case of large signal-to-noise ratio, we obtain a closed-form expression for the bit error probability. We also compare the error performance of wavelet modulation with that of frequency diversity.

A method based on genetic algorithm (GA), a simple clustering procedure for rule base generation, and weighted least squares estimation is proposed to construct a Takagi-Sugeno-Kang (TSK) fuzzy inference system directly from numerical data. The rule-base generation method takes the approach of independently clustering input and output spaces, respectively, and assigning a weight to each rule to capture the relation in input-output data. Genetic process learns the number of linguistic terms per variable and the certainty factors of the rules (indirectly the membership-function parameters of the premise part of the fuzzy rules), and the weighted least squares method is used to determine the consequent part of the fuzzy rules. Simulation results on forecasting the stock market and a benchmark case study are included.

This paper presents a new model-based document image segmentation scheme that uses XML-DTDs (eXtensible Markup Language Document Type Definitions). Given a document image, the algorithm has the ability to select the appropriate model. A new wavelet-based tool has been designed for distinguishing text from non-text regions and characterization of font sizes. Our model-based analysis scheme makes use of this tool for identifying the logical components of a document image.

The present paper aims to present an entirely new approach for the development of "exact" recursive least squares algorithms for ARMA filtering and modeling when the inputs (asssumed here to be "white") are not observable. The approach is heavily based on the recently proposed "predictor-space" representation of ARMA processes 131 and theu se of some new, moreg eneral projection operator update formulas, breifly summarized here.