Boshir Ahmed

A real time speaker identification using artificial neural network

Nowadays it is obvious that speakers can be identified from their voices. In this paper detail of speaker identification from the real-time system point of view is described. Firstly, it have been reviewed the well-known techniques used in speaker identification then the details of every step in identification process and explains the ideas, which leaded to these techniques. We start from the basic definitions used in DSP, then we move to the feature extraction step. Being widely used in pattern recognition tasks, neural networks have also been applied in speaker recognition. In this study, we developed a text-independent speaker identification system based on Back-propagation Neural Network (BPN). BPNs supply flexibility and straightforward design which make the system easily operable along with the successful classification results. In order to analyze the system in practice we made appropriate software and using real data we ran several tests. Empirical results show that proposed approach greatly improves identification speed in feature matching step. From the experiment it is found that the system correctly identify 96% of the speakers, using less then one second of test samples from each speaker.

Optical character recognition using back propagation neural network

This paper represents an Artificial Neural Network based approach for the recognition of English characters using feed forward neural network. Noise has been considered as one of the major issue that degrades the performance of character recognition system. Our feed forward network has one input, one hidden and one output layer. The entire recognition system is divided into two sections such as training and recognition section. Both sections include image acquisition, preprocessing and feature extraction. Training and recognition section also include training of the classifier and simulation of the classifier respectively. Preprocessing involves digitization, noise removal, binarization, line segmentation and character extraction. After character extraction, the extracted character matrix is normalized into 12×8 matrix. Then features are extracted from the normalized image matrix which is fed to the network. The network consists of 96 input neurons and 62 output neurons. We train our network by proposed training algorithm in a supervised manner and establish the network. Eventually, we have tested our trained network with more than 10 samples per character and gives 99% accuracy for numeric digits (0∼9), 97% accuracy for capital letters (A∼Z), 96% accuracy for small letters (a∼z) and 93% accuracy for alphanumeric characters by considering inter-class similarity measurement.

Land cover classification for satellite images based on normalization technique and Artificial Neural Network

The Satellite images and the extracted thematic maps provide higher-level information for the recognize, monitoring and management of natural resources. It is very difficult to identify land cover classification manually from a satellite image. The remotely-sensed images are invaluable sources of information for various investigations since they provide spatial and temporal information about the nature of earth surface materials and objects. This study aims to determine the level of contributions of multi-temporal and multi-sensor data together with their principal components for Artificial Neural Network classifiers. The suitability of Back Propagation Neural Network (BPNN) for classification of remote sensing images is explored in this paper. Automatic image classification is one of the challenging problems of recent year. BPN is self-adaptive dynamic system which is widely connected with the large amount of neurons. It can solve the regular problem arise from remote sensing images. This paper discusses about the BPNN method to improve the high resolution remote sensing image. The principle and learning algorithm of BPNN is analyzed and high resolution imagery of Beijing has been used. Back Propagation Neural Network classifies the remote sensing image into the classified image of their pattern recognition.

An adaptive background modeling based on modified running Gaussian average method

Detecting moving objects is one of the most important research interest at present in computer vision due to its wide range of applications in traffic surveillance, human motion analysis and object tracking. Some approaches such as Gaussian running average provides faster background subtraction for object detection. However, it considers a fixed threshold for the background subtraction, which limits its application. In this research, a modification of Gaussian average technique has been proposed with the aid of an adaptive threshold and learning rate for traffic surveillance. The proposed approach develops a background model dynamically by extracting the edge information of individual frame. The application of adaptive threshold and learning rate over Gaussian average makes the approach more robust and suitable for video surveillance applications. The proposed approach has been tested on the real-time traffic data captured on a busy street using fixed camera. With the proposed technique, the moving vehicles are detected more accurately with little noise. The experimental results presented at the end reflect the suitability of the approach.

Bangla optical character recognition through segmentation using curvature distance and multilayer perceptron algorithm

This paper presents an optical character recognition approach especially for Bangla offline printed characters. Separation of lines, words and individual characters are the main difficulties in printed Bangla character recognition due to different shapes of characters. Different techniques have been applied and performance is examined. It has been studied that no particular algorithm is found for efficient feature extraction. Feature selection is an essential step of optical character recognition. Accurate and distinguable feature plays an important role to leverage the performance of a classifier. A novel feature extraction scheme based on “Discrete Frechet Distance” and “Dynamic Time wrapping” is proposed. Probability of occurrence of the pixels of the given character of different font are calculated which are used to train Multilayer perceptron Neural Network. Experimental results show 100% accuracy for trained characters and overall 90–95% accuracy for all basic characters of the developed method on training and the tests sets respectively.

Effective subspace detection based on cross cumulative residual entropy for hyperspectral image classification

Remote sensing hyperspectral images are blessings of technology through which the ground objects can be detected effectively with the cost of computer processing. For classification of hyperspectral images finding an effective subspace is very important to classify them efficiently. In recent years, many researchers have drawn their interest to extract data more effectively from hyperspectral dataset. In this research, an approach has been proposed to find the effective subspace by measuring the relevance of individual features through Cross Cumulative Residual Entropy from the Principal Component images. The Support Vector Machine has been used as the classifier for the assessment of the feature reduction performance. Experiment has been completed on real hyperspectral dataset and achieved 97% of accuracy which is better than the standard approaches studied.

Feature mining for effective subspace detection and classification of hyperspectral images

Hyperspectral image analysis is becoming an important field of research interest because of its wide range of applications in ground surface identification. New technology is being developed to capture hyperspectral images to cover more spectral bands and finer spectral resolution but also increases challenges to process those images for high correlation among data and both the spectral and spatial redundancy. This paper proposed a feature mining approach for the relevant feature selection as well as efficient classification of the hyperspectral dataset. Principal Component analysis and Mutual Information is two widely used feature reduction techniques utilized in conjunction for the feature reduction of the remote sensing data set. The kernel support vector machine classifier is used to assess the effectiveness of the detected subspace for classification. The proposed feature mining approach is able to achieve 99.3% classification accuracy on real hyperspectral data which higher than the standard approaches studied.

A deep domain adaption approach for object recognition using Multiple Model Consistency analysis

Domain adaption tends to transfer knowledge across domains following dissimilar distribution and where target domain has inadequate labelled samples. When knowledge is transferred from abundantly irrelevant sources negative transfer may occur resulting in poor classification of test samples. Deep learning research illustrates the semantic clustering as well as transferability of deep convolutional features for numerous tasks including domain adaption. Traditional clustering based domain adaption approaches are practical to handle negative transfer scenario. This paper presents a scheme that uses graph based consistency analysis of one supervised and another unsupervised model to effectively transfer knowledge using deep features. This approach uses local neighbourhood analysis to classify hard samples that are identified using consistency analysis of models. This method yields encouraging experimental results on benchmark domain adaption dataset compared to a single deep feature based supervised support vector machine classifier, demonstrating effective use of target domain data.

Neighbourhood consistency based deep domain adaption analysis for multi category object detection

Pattern classification in domains that follow dissimilar distribution and where target domain has insufficient labelled samples, requires transfer of knowledge across domains through a process called domain adaption. Deep learning research demonstrates the transferability of deep convolutional features that are activations of intermediate layers of convolutional neural networks for domain adaption. Traditional clustering based domain adaption approaches are practical to handle knowledge transfer scenario. This paper presents a scheme that uses local neighborhoods based consistency analysis of one supervised and another unsupervised model to effectively transfer knowledge using deep features. Contrasting conventional models this approach uses only two models to classify patterns except hard ones. Neighbourhood consistency analysis identifies the hard samples, and is classified using a third model. Experimental analysis has been carried out focusing change on category variation of different samples for train and test cases. The proposed approach yields encouraging experimental result on benchmark domain adaption dataset compared to a deep feature based single support vector machine classifier in terms of state of the art metrics demonstrating effective generalization of source domain information.

Multi-temporal FFT regression

Sequential data transmission regarding multi-temporal image analysis is mainly dependent upon prediction or forecasting. The transmission time can be substantially reduced by properly exploiting the temporal correlation. Multi-temporal images are often affected by sensor, and illumination variations, non-uniform attenuation, atmospheric absorption and other environmental effects which render system changes in them. Most of these changes are gradual and incremental. So a recent image can be predicted from a previous sequence of images if the amount of real land-cover change is limited. Regression based prediction is the most appropriate one in this case as it can quantify the relationships between images obtained by different measurement systems in different environments. FFT regression based temporal prediction is proposed in this paper whereby the least-squares minimization is conducted on the amplitude matrices of the readings via the FFT. For a given model, the value of squared coefficient of determination (R) is always increased beyond the value obtained by conventional regression which is a common quality measure of the chosen model.