Recognition of cardslip image is difficult task due to its intrinsic poor quality. The problem is worsen when digits are often printed on the underlying form lines. A method of recognizing cardslip images is proposed in which image enhancement, intelligent digit extraction and overlapped line processing are focused.
In the proposed method, the region of interests (ROI) are extracted by matching the histogram of horizontally projected black pixels of a form model with that of the form to process. Then, several advanced image enhancement techniques such as local thresholding and filtering, is processed to get enhanced image quality. The process is performed only on the ROI`s in order to reduce time to process. Then individual digit images are extracted by placing a fixed size window at a right place. The location of the window is determined by trial and error analysis by recognition score within the constraints on neighboring boxes. A simple Bayesian classifier is used for the digit recognizer, but it makes provision for handling form lines which often interfere proper recognition. Since we may know where the line passing by the form analysis, simply the line passing regions are treated as a don`t-care zone.
From an experiment using 57 cardslip images, 56 cardslips are processed successfully and all of the 1626 digits from 56 cardslips of successful processing are correctly recognized. This result is quite satisfiable although we may need more comprehensive evaluation and works to speed up the processing time.