A NOVEL FEATURE EXTRACTION METHOD BASED ON GABOR TRANSFORM FOR CHARACTER RECOGNITION

摘要

In this paper, we firstly give an effective preprocessing method for character images with slant and distortion. Using minimal moment of inertia and rotation algorithm, we achieve rectification of slant which is the primary step of preprocessing. Then we present a novel and effective feature extraction method based on Gabor transform for character recognition. Different from other existing means, this feature extraction method computes ratios of maximum from character images' Gabor transform outputs at rows and columns respectively. The feature vector constructed by maximum ratios can exhibit desirable characteristics of local statistic and orientation selectivity. We test this kind of feature on 785 character images, which are from USPS and carry out the recognition work by a 3-layer BP neural network. Experiments indicate that this feature extraction method can achieve a recognition accuracy as high as 97.1% in character recognition.