Research areas

The major areas of on-going research in the Unit include Document Processing, Pattern Recognition, Natural Language Processing, Information Retrieval and Digital Signal Processing.

The CVPR unit is proud of a pioneering development of an Optical Character Recognition (OCR) system for the combined script recognition of Bangla and Devnagari (Hindi), the most popular scripts of the Indian sub-continent. The system performs document tilt correction, script line, word and character segmentation, character recognition and finally, error correction. Its performance has been found to be about 98%. Work on Oriya OCR and OCR for Bangla hand-written character recognition is in progress.

A major area of interest of this unit is the general problem of digital document analysis. Novel algorithms for skew estimation and correction have been developed. A system for the automatic detection and processing of mathematical expressions from documents is being designed. The issues of automatic layout analysis of both printed and handwritten documents involving segmentation, information and feature extraction, data compression, processing and understanding line drawings are under study while a system for the automatic extraction of meta-information from document images has already been developed. A Bangla programming interface to the Turbo-C compiler has been implemented which has been found to work well for programs of moderate complexity. Another work in progress is language data compression, both from hypertext and images of documents in Indian scripts.

There have been pioneering developments in the area of Natural Language Processing of Indian languages. A powerful Bangla spell-checker for automatic detection and correction of spelling errors in computer stored documents has been developed and extended to the spell checking of Hindi documents. Among other important developments are a talking dictionary and talking editor for Bangla text, a morphological processor for Bangla and a computer based Bangla speech synthesizer. One of the novel objectives is to develop computer aids for the handicapped such as the deaf and dumb as well as the blind. Among other studies, statistical analysis of a large corpus of Bangla text, analysis of human spelling error patterns as well as analysis of onomatopoeic Bangla words have been carried out. Work on developing a tri-lingual computer dictionary (Hindi, Bangla, English) is in full swing.

In the area of bio-medical image processing, the processing of histo-pathological images acquired with a confocal laser scanning microscope for the recognition and grading of diseased cells has been studied at length. New algorithms for 3-D segmentation of these images and feature extraction have been proposed. Bayesian approaches are also being tested on MRI images for brain modelling.

Among soft computing tools, considerable research has been devoted to studying and developing applications of neural networks and genetic algorithms. A modified back propagation algorithm for better convergence and performance has been proposed and tested successfully for real-life problems like segmentation of natural textures and detection of linear structures in satellite images. New neural network models for the purpose of shape analysis in 2 and 3 dimensions have been developed. These models, viz. the dynamic self-organizing neural network and topology adaptive self-organizing neural network, overcome the inherent limitations of the existing models such as Kohonen's self-organizing feature map. A new self-organizing neural network has also been developed for the solution of a well known optimization problem, viz. finding the smallest circle enclosing a given set of planar points. Shape matching and shape detection in a point pattern, symmetry analysis and the k-centre problems are the issues being explored using genetic algorithms.

Some new clustering approaches are proposed where new definitions like nearest centroid neigbourhood have been brought in. The definition of neighbourhood is generalized to fuzzy sets and concepts like convex hull and circular hull have been extended to the digital domain both in crisp and fuzzy cases.

Work has recently started on automatic information retrieval of documents in Indian languages, in pictorial as well as hypertext form. Encouraging results have been obtained on a Bangla database of small size.


No results found

News Archive