|
With the continuous development of computer, all areas of society is accelerating the process of information, in this process, the optical recognition technology (OCR) some work has become the protagonist. OCR technology in handwritten digit recognition technology has developed rapidly in recent years. Although the computer can not do all the proper identification, but its application is already very broad: for example the field of public security license plate recognition, the financial statements for the financial area, bank notes automatically input information, the postal zip code area of ??automatic sorting and so on. Other hand, this problem is also the classic pattern recognition problems, the research for the field of pattern recognition and other problems still have a great inspiration for the pursuit of a better recognition performance, including higher recognition rate and faster identification speed, the researchers are still in the process of identifying all aspects of in-depth research, we can see, the problem of handwritten digit recognition research also has important academic significance and practical value. Firstly, according to the bank real work negotiable instruments scanned into the computer grayscale images to the actual scene handwritten digits as the sample source, a set of handwritten digits fonts, including 50,000 words training font and 20000 word test font, entirely from the actual scene of the font, font true grayscale images. Secondly, feature extraction, this paper uses a gradient breakdown characteristics, such characteristics of the image gradient direction information sub-segments and fuzzy statistics, formation characteristics, the third of four classifiers to achieve this: 1) simple nearest neighbor classifiers (NNS), which improves the characteristics of nearest neighbor classifier metrics from sample to sample distance metric, into the sample to measure a sample are a subset of the simplex type (Simplex) Distance close to portray more accurately, the recognition rate improved significantly. 2) nearest neighbor classifier simplex (NNS) and image deformation matching algorithm (IDM) a combination of two classifiers, it NNS, IDM, respectively, as the first, two classification algorithms, based on the results of the first stage classification Reliability decide whether two matches, the two classifiers also gained a high recognition rate. 3) Based on the IDM ambiguity classifier and NNS classifier combination two classifiers for solving distinguish easily confused sample. 4) Based on Support Vector Machine (SVM) classifier, but also have a good performance. Fourth, the paper also carried out some recognition reliability studies to guide the segmentation and classification of string combination. Finally, this paper attempts a K-nearest neighbor algorithm GPU parallel optimization algorithm, to obtain better acceleration effect.
|