Abstract
Various kinds of materials and methods that are used to print the characters on industrial parts often produce printed letters with differences in color, blurs and other characteristics. Those differences cause individual developments of recognition systems depending object characters. Therefore, the improvement of system productivity becomes a strong requirement.
To answer that requirement, we have developed a character recognition system using a neural network which functions automatic extraction of pattern features. This system uses a core network which is matched to the characteristics of the object characters, so that the system can reduce a leaning time. The core network is generated by a design module which contains knowledge on the applicable range of character recognition algorithms. The design module first selects an algorithm that is suitable for object characters and then outputs parameters that present the core network with functions equal to the selected algorithm.
This system was implemented on a workstation and an image processor, and used in the development of a reconsition system for the low-contrast laser marking characters that are hard to recognize with conventional binary image-processing method.
