Adaptive clustering algorithm for optical character recognition

In analytical optical character recognition, effects of noise and overlapping character blocks constitute a major problem to feature extraction algorithms. This problem degrades the performance of the recognition stage. In this approach, an adaptive clustering algorithm and Hamming Distance computation are proposed to aid the extraction and recognition processes. Initially a line is picked from the segmented lines of a template. The line matrix is projected unto a characters' vector which is used to compute the average width of a character in that line. This information is passed to adaptive clustering analysis. This algorithm check for characters with abnormal width, compute the total binary large objects and return each object as a separate character. Hence any block of characters mistaken for a single character can be detected and re-segmented. The recognition stage adaptively switches from correlation computation to Hamming Distance whenever the former is doomed to fail or produce undesirable results. The approach works well and an appreciable performance improvement is recorded. 5 test images of different font types are used with total of 739 characters. 89.04% successful recognition rate was recorded.