Feedback-based architecture for reading courtesy amounts on checks

Abstract

The processing of bank checks is one application that continues to rely heavily on the movement of paper. Checks are currently read by human eyes and physically transported to the bank of the payer, involving significant time and cost. Since paper checks constitute a popular mechanism for noncash payments, and the volume of checks continues to be high, there is a significant interest in the banking industry for new approaches that can read paper checks automatically. We propose a new approach to read the numerical amount field on the check; this field is also called the courtesy amount field. In the case of check processing, the segmentation of unconstrained strings into individual digits is a challenging task because one must accommodate special cases involving connected or overlapping digits, broken digits, and digits physically connected to a piece of stroke that belongs to a neighboring digit. The described system involves three stages: the segmentation of the string into a series of individual characters, the normalization of each isolated character, and the recognition of each character based on a neural network classifier.

A system for processing handwritten bank checks automatically

“…The first implementation of the segmentation-and-recognition module solved the problem of multiple digits connected together, which is the most common problem found in real checks, but not the problem of broken digits. Subsequently, attempts were made to join characters together [70,71] if they were not recognized independently, but these approaches were not very efficient.…”

Handwritten Bank Check Recognition of Courtesy Amounts

“…19,22 In this paper, we show the level of improvement attained by combining multiple MLPs and also by combining one MLP with GRNN and ELM.…”

“…The ability of the neural network to automatically recognize multiple segments, delimiters or double zeros, instead of expecting a rejection in those cases, can improve the selection of the correct path and can also enhance the strategy in the feedback loop. 19,20 …”

“…The algorithm used by the research team is based on the idea that if a numeral is rotated through a series of slanted positions, it usually attains its minimum width when it is least slanted. 19 No real rotation is performed because the algorithm is slow and sometimes produces bad quality contours. Instead, all pixels in a row are moved to the left or the right, depending on the sign of the correction angle.…”

“…Some of our earlier approaches used a matrix size of 256 pixels (16x16). 19,22 Later the final size was changed to 117 pixels (13x9), while preserving the original aspect ratio. The size currently being used is 204 pixels (17x12); this provides slightly better quality and also preserves the original aspect ratio.…”