Research And Implementation Of Handwritten Digit Recognition System Based On PHOG Description Operator

Handwritten digit recognition has a very wide range of applications, covering the postal code, bank notes, statistics, and other forms of data processing business. Handwritten digit recognition is a hot and difficult problem in pattern recognition, machine learning and other fields. In the actual application, handwritten digit recognition still suffer a great challenge. Given handwritten digital itself characteristics, such as smaller amount of digital information, larger digital writing difference, higher real-time requirements and so on. Therefore, the effectiveness of the handwritten digit recognition and quickness are the research focus.In this paper, the handwritten digit image feature extraction is studied, difficulties and the difficulties of handwritten digit recognition are analyzed to find out the corresponding strategies to resolve the difficulties. Focus on the HOG (Histogram of Oriented Gradient) principle analysis, implementation method, extraction process and performance analysis, and put forward the PHOG (Pyramid Histogram of Oriented Gradient) descriptor: for each handwritten digit image to build a three layers pyramid of HOG, in each layer to use different size of block to achieve statistics information of the local Histogram of Oriented Gradient, generating PHOG description operator. The PHOG operator with the characterization of the describing of the block size well, can represent the local representation of handwritten digits and shape information better, and the image gradient or edge direction distribution information, so the PHOG description operator can extract more robust features of handwritten digits.In addition, this paper uses SVM as the classifier, to choose a specific SVM model, determine the type of kernel function, through cross validation and graph search method, choose the optimal parameters C and ? for SVM model. Finally, the PHOG description operator vectors are classified, greatly improving the recognition accuracy and efficiency.Finally, this article apply the operator to handwritten digit recogniton on the MNIST handwritten digit database, through contrast experiment and the final analysis, the validity of the algorithm is verified.