Bishwajit Purkaystha

Bio: Bishwajit Purkaystha is an academic researcher from Shahjalal University of Science and Technology. The author has contributed to research in topics: Bengali & Recommender system. The author has an hindex of 5, co-authored 6 publications receiving 70 citations.

Bengali handwritten character recognition using deep convolutional neural network

Abstract: Handwritten character recognition is a nontrivial task as it seeks to recognize the correct class for user independent handwritten characters. This problem becomes even more challenging for a highly stylized, morphologically complex, and potentially juxtapositional characters comprising language like Bengali. As a result, the improvements over the years in Bengali character recognition are significantly less as compared to the other languages. In this paper, we propose a convolutional deep model to recognize Bengali handwritten characters. We first learnt a useful set of features by using kernels and local receptive fields, and then we have employed densely connected layers for the discrimination task. Our system has been tested on BanglaLekha-Isolated dataset. It achieves 98.66% accuracy on numerals (10 character classes), 94.99% accuracy on vowels (11 character classes), 91.60% accuracy on compound letters (20 character classes), 91.23% accuracy on alphabets (50 character classes), and 89.93% accuracy on almost all Bengali characters (80 character classes). Most of the errors incurred by our model in recognition task are due to extreme proximity in shapes among characters. A significant number of errors was caused by the mislabeled, irrecoverably distorted, and illegal data examples.

Bengali speech recognition: A double layered LSTM-RNN approach

Abstract: Speech recognition may be an intuitive process for humans, but it turns out to be intimidating to make computer automatically recognize speeches Although recent progresses in speech recognition have been very promising in other languages, Bengali lacks such progress There are very little research works published for Bengali speech recognizer In this paper, we have investigated long short term memory (LSTM), a recurrent neural network, approach to recognize individual Bengali words We divided each word into a number of frames each containing 13 mel-frequency cepstral coefficients (MFCC), providing us with a useful set of distinctive features We trained a deep LSTM model with the frames to recognize the most plausible phonemes The final layer of our deep model is a softmax layer having equal number of units to the number of phonemes We picked the most probable phonemes for each time frame Finally, we passed these phonemes through a filter where we got individual words as the output Our system achieves word detection error rate 132% and phoneme detection error rate 287% on Bangla-Real-Number audio dataset

Comprehending Real Numbers: Development of Bengali Real Number Speech Corpus.

Abstract: Speech recognition has received a less attention in Bengali literature due to the lack of a comprehensive dataset. In this paper, we describe the development process of the first comprehensive Bengali speech dataset on real numbers. It comprehends all the possible words that may arise in uttering any Bengali real number. The corpus has ten speakers from the different regions of Bengali native people. It comprises of more than two thousands of speech samples in a total duration of closed to four hours. We also provide a deep analysis of our corpus, highlight some of the notable features of it, and finally evaluate the performances of two of the notable Bengali speech recognizers on it.

Product recommendation: A deep learning factorization method using separate learners

Abstract: Exponential growth in information has made it totally unimaginable to manually find a relevant product in a quick time, entailing the need for a mechanical recommendation system which would remember the users and recommend most suitable items Most of the approaches for such machinery have been to first find similarity in users or in items, and then exploit these similarities to recommend the products These methods produce better results when demographic information about users and items are given to them In this paper, we propose a deep neural network model which does not require any information be given to it other than the rating triples We created spurious user profiles and item characteristics by using separate learner weights at the bottommost layer The weights in the upper layers took these information, created by the weights at bottommost layer, to produce a real valued rating Our model produced an RMSE 41824 on Jester 4-million datasti, and this shows our deep network is comparable to the state of the art models

Bangla Handwritten Character Recognition using Convolutional Neural Network with Data Augmentation

Abstract: This paper proposes a process of Handwritten Character Recognition to recognize and convert images of individual Bangla handwritten characters into electronically editable format, which will create opportunities for further research and can also have various practical applications. The dataset used in this experiment is the BanglaLekha-Isolated dataset [1]. Using Convolutional Neural Network, this model achieves 91.81% accuracy on the alphabets (50 character classes) on the base dataset, and after expanding the number of images to 200,000 using data augmentation, the accuracy achieved on the test set is 95.25%. The model was hosted on a web server for the ease of testing and interaction with the model. Furthermore, a comparison with other machine learning approaches is presented.

SCCGAN: Style and Characters Inpainting Based on CGAN

Abstract: With the development of deep learning technology, many deep learning methods have been applied to font recognition and generation. However, few studies focus on font inpainting problems. This paper is dedicated to repairing damaged fonts based on style to repair damaged fonts in a better way. In this paper, we propose a CGAN (Conditional Generative Adversarial Nets)-based font repair method. This paper uses the content accuracy and style similarity of the repaired image as an evaluation index to evaluate the accuracy of the restored style font. The font content proposed by the paper based on CGAN network repair style is similar with the correct content.

DeepNetDevanagari: a deep learning model for Devanagari ancient character recognition

Abstract: Devanagari script is the most widely used script in India and other Asian countries. There is a rich collection of ancient Devanagari manuscripts, which is a wealth of knowledge. To make these manuscripts available to people, efforts are being done to digitize these documents. Optical Character Recognition (OCR) plays an important role in recognizing these documents. Convolutional Neural Network (CNN) is a powerful model that is giving very promising results in the field of character recognition, pattern recognition etc. CNN has never been used for the recognition of the Devanagari ancient manuscripts. Our aim in the proposed work is to use the power of CNN for extracting the wealth of knowledge from Devanagari handwritten ancient manuscripts. In addition, we aim is to experiment with various design options like number of layes, stride size, number of filters, kenel size and different functions in various layers and to select the best of these. In this paper, the authors have proposed to use deep learning model as a feature extractor as well as a classifier for the recognition of 33 classes of basic characters of Devanagari ancient manuscripts. A dataset containing 5484 characters has been used for the experimental work. Various experiments show that the accuracy achieved using CNN as a feature extractor is better than other state-of-the-art techniques. The recognition accuracy of 93.73% has been achieved by using the model proposed in this paper for Devanagari ancient character recognition.

Deep learning techniques for rating prediction: a survey of the state-of-the-art

Abstract: With the growth of online information, varying personalization drifts and volatile behaviors of internet users, recommender systems are effective tools for information filtering to overcome the information overload problem. Recommender systems utilize rating prediction approaches i.e. predicting the rating that a user will give to a particular item, to generate ranked lists of items according to the preferences of each user in order to make personalized recommendations. Although previous recommendation systems are effective in creating attired recommendations, however, they still suffer from different types of challenges such as accuracy, scalability, cold-start, and data sparsity. In the last few years, deep learning has attained substantial interest in various research areas such as computer vision, speech recognition, and natural language processing. Deep learning based approaches are vigorous in not only performance improvement but also to feature representations learning from the scratch. The impact of deep learning is also prevalent, recently validating its efficacy on information retrieval and recommender systems research. In this study, a comprehensive review of deep learning-based rating prediction approaches is provided to help out new researchers interested in the subject. More concretely, the classification of deep learning-based recommendation/rating prediction models is provided and articulated along with an extensive summary of the state-of-the-art. Lastly, new trends are exposited with new perspectives pertaining to this novel and exciting development of the field.