Accuracy improvement in odia zip code recognition technique

Abstract

Odia is a very popular language in India which is used by more than 45 million people worldwide, especially in the eastern region of India. The proposed recognition schemes for foreign languages such as Roman, Japanese, Chinese and Arabic can’t be applied directly for odia language because of the different structure of odia script. Hence, this report deals with the recognition of odia numerals with taking care of the varying style of handwriting. The main purpose is to apply the recognition scheme for zip code extraction and number plate recognition. Here, two methods “gradient and curvature method” and “box-method approach” are used to calculate the features of the preprocessed scanned image document. Features from both the methods are used to train the artificial neural network by taking a large no of samples from each numeral. Enough testing samples are used and results from both the features are compared. Principal component analysis has been applied to reduce the dimension of the feature vector so as to help further processing. The features from box-method of an unknown numeral are correlated with that of the standard numerals. While using neural networks, the average recognition accuracy using gradient and curvature features and box-method features are found to be 93.2 and 88.1 respectively.