The Research Of Numerals Recognition Using Pseudo-Zernike Moments

With the rapid development of science and technology, especially for computer, signal processing and artificial intelligence, numerals recognition technology which is an important part of pattern recognition and computer intelligent interface technology has been widely used. Moment function is an important tool in image analysis and has been widely used in the field of computer vision, image processing and pattern recognition due to their invariant properties. This paper aims at the research of moment technique and its applications in image processing and recognition.Based on the former research work, first, this paper summarizes the current research status in character recognition, introduces the framework and the development of the character recognition system; and then, this paper discusses the definitions and properties of regular moments, orthogonal moments and other moments and invariants, compare and evaluate the performance of various moments in image representation, noise sensitivity, and information redundancy; this paper still discuss the features and properties of Pseudo-Zernike moments, they have rotation invariance, translation invariance, scale invariance and small noisy sensibility, they can be applied to reconstruct object directly and represent image with a minimal amount of information redundancy, can be used to describe motion pictures and so on; lastly, aiming at the problem of common algorithms of Pseudo-Zernike moments could only be used to deal with foursquare images, based on Chong's algorithm, this paper presents an improved algorithm based on mapping rectangular image into a unit circle which has better performance in both calculation time and recognition precision; also,this paper applies the improved algorithm to the recognition of numeral images and numerals in laser anti-counterfeit identifiers with the BP net. The results show that, in calculation time, the improved algorithm is only half of Chong's algorithm; in convergence of the BP net, with the 50 training samples, the Sum-Squared Network Error of the improved algorithm is 0.14465, lower than 0.14759 of Chong's algorithm; in recognition precision, with the 50 recognition samples, the accuracy of the improved algorithm is 90%,higher than 82% of Chong's algorithm. In conclusion, the improved algorithm of this paper is very effective to the recognition of numerals.