A CONNECTIONIST APPROACH TO DEVNAGRI CHARACTER RECOGNITION

Abstract

With computers encroaching into every activity of man today, long hours of manual labour are spent in typing-in the data, before it is processed by the computer, be it in the field of word processing, database creation or software development. The field of character recognition has a very important role to play on this front. This dissertation deals with the first step in alleviating this effort, that is, electronically reading and recognising handwritten characters directly from the manuscripts, thus removing the need of typing. The domain chosen for recognition is the Devnagri consonants ( ). To negotiate the variations and deformities that occur in normal hand-writing, a connectionist approach is adopted. The manuscript is first scanned with an image scanner in BMP file format. The BMP file, in addition to the image pixel data, contains a header, in which various information related to the scanned image resides. So the first job is to segregate the pixel values by detaching the header. The extracted pixel data is then scanned by the program for detecting the existence of characters. The characters found are then, one after another scaled to a 30 x 30 pixel matrix of is and Os, so that the recognition procedure is immune to the variations in the size of characters. The matrix data after suitable reduction and encoding Is fed to a feedforward,. multilayer, error backpropagation trained neural network for recognition. The output is then interpreted to get the recognised character. By training the network for different shapes and sizes of characters, the recognition efficiency is enhanced to produce a very reliable character recognition software. The program code is divided into various modules and a combination of C and C++ is used to score them in about 2500 J lines. The whole software package Is executed on the SCI's IRIX system.