Research On DS-SVM For Handwritten Digit Recognition

In recent years, support vector machine (Support Vector Machine, SVM), as a new machine learning method, is developed rapidly. Studies have shown that the method can be used in many fields and have a strong performance promotion. SVM, applied in image recognition and gene mapping identification in signal processing, have won the best performance and show its advantages by far. The handwritten numeral recognition as the basic research in the field of pattern recognition, therefore, SVM classifiers in it also be given considerable recognition results.The paper focuses on the handwritten digits to conduct a comprehensive study. This system includes three parts:a digital image pre-processing, feature extraction and identification numbers. The pre-processing section, including a grayscale image, denoising, threshold transform and an image normalization, are compared to select the best algorithm to adapt to the experiment by a different algorithm. The main part of the image feature extraction using three different methods of feature extraction (evidence source) from three different angles, including the coarse grid feature extraction, projection feature extraction, contour feature extraction. Proposed based on support vector machine and DS evidence theory combining multiple features fusion algorithm in digital identification process. This method is a method on three different feature extraction, through improved one-against-one vote principle classification level SVM classifier rough classification, while generating the different corresponding BPA functions; then, using the modified combination rule of D-S evidence theory, to generating some conflict factor between the evidence sources and others, in order to modify the original basic belief function. Though the fusion of multiple features, come into a new-belief function. The new-belief function as input to the two level of SVM classifier, and carry out the final decision recognition.The experimental sample mainly statistics a large number of 0-9 handwritten digits from the marking system. Through the C++platform experiment shows that the recognition algorithm, proposed in this paper,can effective reduction error recognition rate and increase recognition rate due to single feature and SVM. In addition, refer to the relevant documents and experimental conclusion show that the BP neural network, template matching and Bayesian algorithm have low recognition rate and save a lot of time to identify, at the same time, the system have a better fault tolerance ability and stronger robustness.