Research On Key Technologies For Character Recognition Base On Neural Networks

Optical character recognition as a method of text automatically entered is widely used in different industries such as banking, shipping, commerce and communication, etc. It can greatly improve the speed of access to information and reduce the strength of people’s mechanical work. Over the years, optical character recognition technology has become the goal of many researchers. Optical character recognition technology has many achievements, but there are still many problems. So it can’t meet some actual needs. Therefore, it is important to research on key technologies for character recognition, look for reasonable improvement programs and improve the performance of character recognition system.In this paper, some key technologies for batch character recognition of the scanning characters image are researched base on digital image processing technology and pattern recognition technology. Some work on the following aspects is done.(1) The performance of several binarization methods is compared. Aiming at the problem of uneven distribution for the target area and background area and the difficulties of separation, the method of local iteration Otsu algorithm is used to achieve image binaryzation. Aiming at the problem of skew correction complex process and poor adaptability, a small angle correction algorithm using fourier transform and a regional localization algorithm based on horizontal and vertical projection are proposed. Then redundant information is reduced. The accuracy of target segmentation is improved. According to layout characteristics of character images in the practical application, range scanning method is used for fast segmentation of multiple lines of text in order to achieve a single character extraction.(2) This paper studies the various methods of extraction for character feature. The method of extracting the external contour feature and penetration characteristics after character skeletonization to combine into a single character feature dataset is proposed to improve character recognition rate of system. In the recognition stage, the widely used BP neural network is studied. But the disadvantages, such as slow convergence rate and easy fall into local minimum restrict its further development and application. The improved BP algorithm with self-adaptive learning rate and additional momentum can effectively reduce the training time, speed up the convergence rate and inhibit the optimization algorithm into a local minimum. The improved algorithm is applied to the image character recognition system.(3) System of character recognition based on artificial neural networks is designed to verify the feasibility of the proposed algorithms and the practicality of the system.The actual test results show that: the proposed approaches are practical and effective, the recognition rate for the scanning character image with higher definition and smaller inclination is close to100%. At the same time, the algorithm in entire system is quite simple, complexity is low enough to meet the requirements of fast identification.