Recognition of chain-coded patches with statistical methods

Patch recognition with chain codes is used not only in the field of binary pattern matching but also in character recognition and binary patch recognition. Chain codes are often used in image processing. A chain code is the series of the direction codes of an objectʹs contour related to a starting point. This paper deals with images having only an outer contour. The disadvantages of the related known methods are the inefficiency of the rotation and scale invariance and their excessively noise-sensitive character. If we want a less noise-sensitive algorithm, we have to use statistical methods. If we introduce shape codes instead of chain codes, we can achieve rotation and starting point invariance. Regarding these codes as sample elements of random variables, they can be modified to become scale invariant and less noise sensitive, and we can apply modified χ2 or other tests on them. The algorithms can be used as prefilters. But, it is important to use some noise-filtering before using these methods, because the contour noise can spoil the efficiency of the methods to a large extent.