Research On Chinese Character Feature Extraction And Recognition Technique

Text location and identification technology has always been an important study branch in the field of image processing and computer vision. With the development of media technology, lots of image texts filled with complex backgrounds emerge in more and more application circumstances. It's a great challenge to let computer understand existing Chinese character inside, and Chinese character recognition technology based on these types of image text is considered to be a new research direction.Traditional Optical Character Recognition software OCR (Optical Character Recognition) can not deal with such type of image. Meanwhile, conventional Chinese character structure feature and statistical feature have some limitations in character representation. In order to solve these problems, we innovatively propose to adopt local feature to describe. This paper first chooses and analyzes three typical feature detection methods named Harris,SIFT and MSER. Through contrast experiments, we see SIFT is the best. Then, we emphasize on local feature, analyzing SIFT algorithm. With the aspects of shape character and gray-level information of images, we also raise two new feature descriptors: (1) SSIFT (Shape SIFT) based on relatively global shape character and SIFT feature; (2) GSD (Gray Scale Difference) descriptor. The results show: (1) New algorithms overcome and weaken the above problems to a certain extent, and perform good invariance against rotation, scale change and backgrounds influence.This paper computes Chinese character recognition rate to measure representation ability of all the features, referring to image matching way. Via experimental data, this paper makes further efforts to propose a new course-to-fine matching strategy, which makes a little improvement.Geometric constraint method is another key part of research. This paper first gives a common problem that Chinese character has similar local structure inside, and analyzes the usage of geometric constraint. Then, this paper brings forward a method based on Mean-shift clustering algorithm and a fresh measure-criterion suitable for high dimensional vector. This method well settles the mismatching problem thanks to resemble local feature. Constraint on space position relationship finally improves recognition performance.The current research achievements largely enrich research ideas in Chinese character recognition and its applications, which have a certain contribution both in theoretical significance and application value.