Mahtab Ahmed

Convolutional neural network training with artificial pattern for Bangla handwritten numeral recognition

Recognition of handwritten numerals has gained much interest in recent years due to its various application potentials. The progress of handwritten Bangla numeral is well behind Roman, Chinese and Arabic scripts although it is a major language in Indian subcontinent and is the first language of Bangladesh. Handwritten numeral classification is a high-dimensional complex task and existing methods use distinct feature extraction techniques and various classification tools in their recognition schemes. Recently, convolutional neural network (CNN) is found efficient for image classification with its distinct features. In this study, a CNN based method has been investigated for Bangla handwritten numeral recognition. A moderated pre-processing has been adopted to produce patterns from handwritten scan images. On the other hand, CNN has been trained with the patterns plus a number of artificial patterns. A simple rotation based approach is employed to generate artificial patterns. The proposed CNN with artificial pattern is shown to outperform other existing methods while tested on a popular Bangla benchmark handwritten dataset.

A rule based approach for NLP based query processing

Information is playing an important role in our lives. One of the major sources of information is databases. Databases and database technology are having major impact on the growing use of computers. In order to retrieve information from a database, one needs to formulate a query in such way that the computer will understand and produce the desired output. Generally, query processing is handled by the Structured Query Language (SQL). But the non IT people cannot be able to write SQL queries as they may not be aware of the SQL as well as structure of the database. So there is a need for non-expert users to query the databases in their natural language instead of working with the values of the attributes. This paper proposes an approach for accessing the database easily using natural language without having any knowledge about the query language. The approach is a rule based approach. The obvious advantage is that it makes a great promise for computer interfaces easier for the use of general people. Because of this, people will be able to communicate to the computer in their own language instead of learning a specialized language or commands. In order to test our approach in an actual computer environment, we have developed a prototype system. We obtained promising results using our system.

Classification of massive noisy image using auto-encoders and convolutional neural network

Image processing tasks has found a new dimension with the improvement of learning feature representation from images using deep networks. Most of the research works are conducted over pre-possessed image data in the lab. But, these methods fail in the real world scenario as most of the time the image required to classify is subject to noise and other disfigurement. For the last three decades, many researches has been conducted and numerus algorithms have been proposed with varying performances to classify noisy images. But in recent times, various autoencoders have outperformed all traditional methods for reconstructing native image from its noisy form and opened a new door for the research of noisy image classification. In this paper, we studied various auto encoders for reconstructing native images from noisy input images. We have applied convolutional neural network as classifier. Before classification task we have rectified noisy images using denoising autoencoder, convolutional denoising autoencoder and finally a hybrid of them as proposed in this paper. The proposed methods are evaluated by experimenting over benchmark dataset adulterated with noises of different proportionate. This method has outperformed some other prominent methods achieving satisfying classification accuracy even when the image is too much noisy (50% noise is added with the image data).

Convolutional Neural Network Training incorporating Rotation-Based Generated Patterns and Handwritten Numeral Recognition of Major Indian Scripts

ABSTRACT Handwritten numeral recognition has gained much interest in recent times because of its diverse application potentials. Bangla and Hindi are the two major languages in Indian subcontinent and a large number of population in vast land scape uses Bangla and Devnagari numeral scripts of these two languages. Well-performed handwritten numeral recognition system for Bangla and Devnagari is challenging because of similar shaped numerals in both scripts; few numerals differ from their similar ones with a very few variation even in printed form. In this study, convolutional neural network (CNN) based two different methods have been investigated for better recognition of Bangla and Devnagari handwritten numerals. Both the methods use rotation-based generated patterns along with ordinary patterns to train CNN but in two different modes. In multiple CNN case, three different training sets (one with ordinary patterns and two with clockwise and anti-clockwise rotation-based generated patterns) are prepared; three different CNNs are trained individually with each of these training sets; and their decisions are combined for final system decision. On the other hand, in the case of single CNN, combination of above three training sets is used to train one CNN. A moderated pre-processing is also employed while generating patterns from the scanned images. The proposed methods have been tested on prominent benchmark handwritten numeral datasets and have achieved remarkable recognition accuracies. The achieved recognition accuracies are found better than reported recognition accuracies of prominent existing methods; and such outperformance mounted proposed methods as better recognition systems. Moreover, CNNs performance improvement due to use of generated patterns has also been clearly identified from the presented experimental results.

Handwritten Bangla Numeral Recognition Using Deep Long Short Term Memory

Recognition of handwritten numerals has gained much interest in recent years due to its various application potentials. Bangla is a major language in Indian subcontinent and is the first language of Bangladesh, but unfortunately, study regarding handwritten Bangla numeral recognition (HBNR) is very few with respect to other major languages such as English, Roman etc. Some noteworthy research works have been conducted for recognition of Bangla handwritten numeral using artificial neural network (ANN) as ANN and its various updated models are found efficient for classification task. The aim of this study is to develop a better Bangla handwritten numeral recognition system and hence investigated deep architecture of Long Short Term Memory (LSTM) method. LSTM is a variant of recurrent neural networks (RNN) and is applied efficiently for image classification with its distinct features. The proposed HBNR-LSTM normalizes the written numeral images first and then employs two layers of LSTM to classify individual numerals. Unlike other methods, it does not employ any feature extraction technique. Benchmark dataset with 22000 hand written numerals with different shapes, sizes and variations are used in this study. The proposed method is shown satisfactory recognition accuracy and outperformed other prominent exiting methods.

Multiple Convolutional Neural Network Training for Bangla Handwritten Numeral Recognition

Recognition of handwritten numerals has gained much interest in recent years due to its various application potentials. The progress of handwritten Bangla numeral is well behind Roman, Chinese and Arabic scripts although it is a major language in Indian subcontinent and is the first language of Bangladesh. Handwritten numeral classification is a high-dimensional complex task and existing methods use distinct feature extraction techniques and various classification tools in their recognition schemes. Recently, convolutional neural network (CNN) is found efficient for image classification with its distinct features. In this study, three different CNNs with same architecture are trained with different training sets and combined their decisions for Bangla handwritten numeral recognition. One CNN is trained with ordinary training set prepared from handwritten scan images, and training sets for other two CNNs are prepared with fixed (positive and negative, respectively) rotational angles of original images. The proposed multiple CNN based approach is shown to outperform other existing methods while tested on a popular Bangla benchmark handwritten dataset.

Stacked auto encoder training incorporating printed text data for handwritten bangla numeral recognition

Recognition of handwritten numerals has gained much interest in recent years due to its various application potentials. Bangla is a major language in Indian subcontinent and is the first language of Bangladesh; but unfortunately, study regarding handwritten Bangla numeral recognition (HBNR) is very few with respect to other major languages such as English, Roman etc. Some noteworthy research works have been conducted for recognition of Bangla handwritten numeral using artificial neural network (ANN) as ANN and its various updated models are found to be efficient for classification task. The aim of this study is to develop a better HBNR system and hence investigated deep architecture of stacked auto encoder (SAE) incorporating printed text (SAEPT) method. SAE is a variant of neural networks (NNs) and is applied efficiently for hierarchical feature extraction from its input. The proposed SAEPT contains the encoding of handwritten numeral into printed form in the course of pre-training and finally initializing a multi-layer perceptron (MLP) using these pre-trained weights. Unlike other methods, it does not employ any feature extraction technique. Benchmark dataset with 22000 hand written numerals with different shapes, sizes and variations are used in this study. The proposed method is shown to outperform other prominent existing methods achieving satisfactory recognition accuracy.

Improvement of haar feature based face detection incorporating human skin color analysis

Face detection from a digital image or video stream is used often for various purposes. But sometimes a system detects an object or area as a face where there is no face at all. This paper presents a technique to reduce such wrong detection rate introducing human skin color (HSC) characteristic. The general property of human skin in RGB color space is that it possess R>G>B (i.e., red values are higher than green value and green value is higher than blue). In this study, such HSC property has been incorporated with the popular Haar feature based face detection (HFFD), to reduce wrong detection of faces. Proposed HFFD with HSC (HFFD-HSC) has been tested and compared with standard HFFD rigorously on a large number images with single and multiple faces. Experimental results identified the effectiveness of HSC incorporation in HFFD to improve its performance reducing wrong detection of faces.