Abstract: A character recognition device and system:(1) constructs a database comprising pairs of similar-looking characters;(2) selects critical features distinguishing these characters from each other; and(3) performs a recognition process based on the selected critical features.In the present invention, sample data is automatically analyzed to construct a feature reference database for a typical character recognition device. If two characters are confused for a threshold number of times, this pair is considered to be similar-looking or confusing. The reference data already accumulated is statistically analyzed for each similar-looking pair to determine the importance of certain features in distinguishing the two characters. These features are further statistically analyzed so that only the most dominant features will be looked at and the less dominant or similar-looking features will be ignored.

Abstract: A preprocessing method for smoothing handwritten Chinese characters before subjecting the same to computerized character recognition and/or other applications. The method disclosed in the present invention is a run-based smoothing method which takes advantages of the unique characteristics of the building blocks of Chinese characters, which consist essentially of vertical, horizontal, and tilted strokes. Accordingly, the two-dimensional array of pixels representing a digitized character are first converted into a run information map which will be repeatedly scanned and subject to three main smoothing treatments: (1) empty run filling, (2) small triangular area filling, and (3) filled run deleting. A unique protocol is developed for effectuating each of the smoothing steps. After each treatment, the run information map is updated, and the treatment steps are repeated until no more empty run filling or filled run deleting can be found, or until a specified number of cycles are exceeded.

Abstract: In accordance with the present invention, a neural network comprising an array of neurons (i.e. processing nodes) interconnected by synapses (i.e. weighted transmission links) is utilized to carry out a probabilistic relaxation process. The inventive neural network is especially suited for carrying out a variety of image processing tasks such as thresholding.