Intelligent Character Recognition (ICR) is a specific form of optical character recognition (OCR) dealing mostly with handwritten texts. Due to their specificity, they are usually more adept in interpreting different styles and fonts of handwriting providing eventually higher recognition rates. Factors like language constructs, amount of research on ICR pertaining to the language, etc., essentially determines the amount of success achieved in its character recognition. This research mainly deals with the recognition of Gujarati Handwritten Characters. We have considered 34 consonants and 5 vowels; a total of 39 Gujarati Characters. The structure and lexicons of the language posed a challenge during the initial phase of segmentation; for that we have proposed new algorithm for segmentation. Our segmentation algorithm is able to address these concerns effectively. Different algorithms from different domains have been considered for comparative analysis like Transform Domain (DWT, DCT and DFT), from Spatial Domain; Geometric Method (Gradient feature), Structural method (Freeman chain code) and Statistical method (Zernike Moments). We have also proposed a new Combination of Structural and Statistical methods (Freeman chain code, Hu's invariant moment and center of mass) to extract feature vectors and it results into good amount of accuracy. These extracted feature vectors were further supplied as input into Support Vector Machines and their resulting accuracies were analyzed using 10 fold cross validation. SVM performs well on data sets that have many attributes and can also handle large number of classes.

This paper addresses the problem of recognizing handwritten numerals for Gujarati Language. Three methods are presented for feature extraction. One belongs to the spatial domain and other two belongs to the transform domain. In first technique, a new method has been proposed for spatial domain which is based on Freeman chain code. This method obtains the global direction by considering n x n neighbourhood and thus eliminates the noise which occurs due to local direction. In second and third method, 85 dimensional Fourier descriptors and Discrete Cosine Transform coefficients were computed and treated as feature vectors. Comparative analysis has been done for these three methods. These methods are tested with three different classifiers namely K-Nearest Neighbour, Support Vector Machine and Back Propagation NeuralNetwork. Experimental results were evaluated using 10 fold cross validation. The highest recognition rates obtained for full data set of 3000 digits are 85.67%, 93.60% and 93.00% using modified chain code, DFT and DCT respectively.

No abstract

No abstract