Sarfaraz Masood

Training neural network with zero weight initialization

We put forth a new paradigm for neural network training in which the initial weights to the network are set to zero. This is done in conjunction with random learning rate to achieve better results. To validate the work, the means test errors were calculated for the traditional approach and the newly proposed paradigm. These results suggest that this new paradigm can be used as an alternate approach to train the neural networks. This new paradigm gives lesser value for the mean test error for some problems than those generated using the traditional random initial weights initialization approach. These results suggest that this proposed paradigm is equivalent and even at times better than the traditional random initial weights initialization approach.

Novel approach for musical instrument identification using neural network

This work aims to solve the problem of musical instrument identification in monophonic audio samples. The instruments chosen for this work were piano, flute, violin, drums and guitar. The audio data were sampled into frames of fixed size & then MFCC and few other TIMBRAL features were extracted from them. These features were used for training and testing the network. But instead of selecting one frame as an input, a different method was used to create a training and testing inputs for the classifier. Several experiments were conducted to obtain the best possible network using different training algorithms, learning rates and number of epochs. The results obtained from the experiments suggest that the best obtained network is an efficient classifier of the musical instruments under consideration.

Real-Time Sign Language Gesture (Word) Recognition from Video Sequences Using CNN and RNN

There is a need of a method or an application that can recognize sign language gestures so that the communication is possible even if someone does not understand sign language. With this work, we intend to take a basic step in bridging this communication gap using Sign Language Recognition. Video sequences contain both the temporal and the spatial features. To train the model on spatial features, we have used inception model which is a deep convolutional neural network (CNN) and we have used recurrent neural network (RNN) to train the model on temporal features. Our dataset consists of Argentinean Sign Language (LSA) gestures, belonging to 46 gesture categories. The proposed model was able to achieve a high accuracy of 95.2% over a large set of images.

American Sign Language Character Recognition Using Convolution Neural Network

Communication is an important part of our lives. Deaf and dumb people being unable to speak and listen, experience a lot of problems while communicating with normal people. There are many ways by which people with these disabilities try to communicate. One of the most prominent ways is the use of sign language, i.e. hand gestures. It is necessary to develop an application for recognizing gestures and actions of sign language so that deaf and dumb people can communicate easily with even those who don’t understand sign language. The objective of this work is to take an elementary step in breaking the barrier in communication between the normal people and deaf and dumb people with the help of sign language. The image dataset in this work consists of 2524 ASL gestures which were used as input for the pre-trained VGG16 model. VGG16 is a vision model developed by the Vision Geometry Group from oxford. The accuracy of the model obtained using the Convolution Neural Network was about 96%.

Analysis of weight initialization methods for gradient descent with momentum

The back propagation algorithm using gradient descent with momentum is a commonly used training algorithm for the artificial neural networks. In this work, a set of experiments were conducted to obtain a detailed comparison of various known weight initialization methods. By doing so, the best suited weight initialization routines for the gradient descent approach with momentum was identified. Six problems of the functions approximation domain were selected for these experiments. Statistical metrics like one sided tailed t-test, the standard deviation of simulation error as well as its mean value were evaluated and used for the purpose of decision making. Results obtained from these experiments strongly advocate that the weight initialization method proposed by Nguyen and Widrow was the best suited technique while training the network by Gradient Descent with momentum approach.

Isolated word recognition using neural network

Isolated Word Recognition is the process of converting the spoken word into its corresponding text format. At present mainly Mel Frequency Cepstrum Coefficients (MFCC) is used as the feature extraction parameter i.e. the identifying features for the speech signal. Through this paper efforts have been made to determine the accuracy of an MFCC based system and also to build an isolated word recognizer based on word acoustic model that uses MFCC in combination with other features of speech such as Root Mean Square Energy, Length of the word and its Brightness. Using an artificial neural network as the classifier, the system was trained & tested for a set of spoken isolated words. The results obtained showed a high and an increased accuracy for the experiment in which along with MFCC other selected parameters were also involved against the experiment which only involved MFCC.

Genre classification of songs using neural network

The objective here is to eliminate the manual work of classifying genres of song in each song. With this startup work songs can be classified in real-time and proposed parallel architecture can be implemented on the multi-processing system as well. In this paper a set of features are obtained like beats/tempo, energy, loudness, speechiness, valence, danceability, acousticness, discrete wavelet transform etc., using Echonest libraries and are fed into the Parallel Multi-Layer Perceptron Network to obtain the genres of the song. The proposed scheme has an accuracy of 85% when used to classify two genres of songs that are Sufi and Classical.

Prediction of Human Ethnicity from Facial Images Using Neural Networks

This work attempts to solve the problem of ethnicity prediction of humans based on their facial features. Three major ethnicities were considered for this work: Mongolian, Caucasian and the Negro. A total of 447 image samples were collected from the FERET database. Several geometric features and color attributes were extracted from the image and used for classification problem. The accuracy of the model obtained using an MLP approach was 82.4% whereas the accuracy obtained by using a convolution neural network was a significant 98.6%.

Cryptanalysis of Image Cryptosystem Using Synchronized 4D Lorenz Stenflo Hyperchaotic Systems

Lately, a color image cryptosystem is suggested for secure wireless communication using 4D Lorenz Stenflo hyperchaotic systems. The proposition specified a nonlinear state feedback-based synchronization for master–slave Lorenz Stenflo chaotic systems. It presents seemingly successful application of synchronized chaotic systems for image encryption which is backed by simulations to assess the efficiency and stability of encryption. However, the image cryptosystem has the presence of certain loopholes. This paper aims to propose the cryptanalysis of this cryptosystem by exploiting existing vulnerabilities and loopholes. To prove that encryption algorithm is devoid of security, we mount the proposed attacks in the form of chosen-plaintext attack that recover the plaintext image from encrypted image without secret key. It is, therefore, shown through experimental simulations that the image cryptosystem is all insecure for use in practical applications of image-based secure wireless communication.

Dynamic 9 × 9 Substitution-Boxes Using Chaos-Based Heuristic Search

Large-sized substitution-boxes tend to provide high security and resistant to some attacks as compared to small sized S-Boxes. However, finding large and efficient n × n S-boxes (n > 8) is an open issue. Here, we propose to put forward a chaos-based heuristic search strategy to generate dynamic 9 × 9 S-boxes. The anticipated strategy has the ability to search optimized S-boxes as the generations are applied. As an instance, the S-box constructed by the proposed strategy is assembled and analyzed against the criterions such as bijectivity, nonlinearity, algebraic degree, differential probability, robustness to differential cryptanalysis, transparency order, etc. The simulation outcomes verify that chaos-based heuristic search strategy is streamlined and has proficiency of synthesizing cryptographically potent 9 × 9 dynamic substitution-boxes capable of exhibiting consistent performance lineaments.