Performance Evaluation of Different Image Sizes for Recognizing Offline Handwritten Gujarati Digits Using Neural Network Approach

This paper describes a system for recognizing offline handwritten Gujarati Digits using Neural Network. Data samples are collected from different writers on A4 sized paper. They are scanned using a flat bed scanner at a resolution of 150 dpi. Various pre-processing operations are performed on the digitized image. Random sized pre-processed image is normalized to uniform sized image. The pixel density is calculated as binary patterns and therefore a vector is created. These features are used to train and test the Neural Network. The recognition results are tested for 3 different sizes of images 7×5, 14×10 and 16×16 for the handwritten Gujarati digits. We have collected 3900 samples of handwritten Gujarati digits from various persons. Various pre-processing steps are applied on collected sample images before passing it to the neural network. Then the Neural Network has been trained and tested on these samples for the three different sizes. The results are evaluated for different image sizes of 7×5, 14×10 and 16×16. The overall recognition rates are 87.29%, 88.52% and 88.76% for different image size.