Handwritten Character Recognition Feature Extraction And Classifier

With the continuous development of computer, the process of information in all fields of society is accelerating dramatically; meanwhile, the optical recognition technology (OCR) in this process has become the protagonist of some of the work. The handwritten numeral recognition is an important optical recognition technology branch. Although computers can not identify all correctly, they have been applied widely:for example, the recognition of public security license plate, financial sector, financial statements, business reports, banking information, automatic bill entry, automatic sorting of postal zip code areas and so on. On the other hand, it is also one of classic problem of pattern recognition, pattern recognition for its research and other issues are still a great inspiration for the pursuit of better recognition performance, including a higher recognition rate and a faster identification speed, the researchers still identify all aspects of the process of in-depth study. Obviously, the study of handwritten digit recognition also has important practical and theoretical academic significance.First of all, a set of handwritten digital character should be established according to the grayscale images which were scanned into the computer in actual work of bank notes in circulation and the sample source of handwritten digital sources in actual scene. It includes 50000 words of training data set and 20000 words of the test samples. The word stock from the actual scene completely, for real word grayscale image. Secondly, in the feature extraction, this paper uses a gradient-based sub-direction of the statistical features and fuzzy statistics to enrich the information of characters in gray-scale image edge, then be broken down to form the final character features. Thirdly, in the classification, this paper implements four classifiers:1) the nearest neighbor simplex classifier (NNS). It improves the distance measure way of the nearest neighbor simple classifier by measuring the distance from sample to sample into measuring the distance from the sample to the one's subsets so that the depict distance is more accurate, and raise the recognition ratio significantly.2) the secondary classifier consists of simple nearest neighbor simplex classifier (NNS) and the image deformation matching algorithm (IDM). It regards NNS as a first-level and IDM as a second-level classifier. That the reliability of the results of the first level classification decides whether the secondary match to be done or not. The secondary classifiers also achieved a high recognition rate.3) A classifier combined by IDM ambiguous with the NNS classifier to distinguish between easily confused samples.4) the classifier based on SVM. The classifier utilizes the support vector machine classification algorithm to classify. It also achieved good recognition effects. In addition, we also identified a number of reliability studies to guide the string segmentation and classifier combination. Finally, we try the parallel optimization of the K-nearest neighbor algorithm on GPU based on CUDA architecture, get good acceleration.