The Research Of The Handwritten Numeral Recognition Based On Feature Extraction And Classification

Handwritten numeral character recognition is a classic pattern recognitionproblem， which has great theoretical research value with broad application prospect.Handwritten numeral recognition(OCR) is a branch of the optical characterrecognition technology. The main issue of the research is how to use the computer toautomatically recognize handwritten Arabic numerals. The off-line handwrittencharacter recognition can be the hardest research in the OCR field, and handwrittennumeral recognition is one of the most popular issues among these research.A common approach to solve a handwritten numeral recognition problem is todivide the solution into the two parts of feature extraction and classification. Basedon this idea, the paper first introduces a number of methods to extract imagedescriptors, including SIFT, PCA, and the HOG method.SIFT local feature descriptors show extraordinary robust against mostdisturbances such as scaling, rotation, occlusion, perspective and illumination changesin scene of matching and object recognition. PCA refers to Principal ComponentAnalysis, whose basis idea is dimensionality reduction. It can effectively find themost important elements and structure from the original data. HOG descriptor baseson the idea that local properties can be described by gradient or the distribution ofedge orientation density, it is generally used as feature descriptor for target detection.Based on Spatial Pyramid Matching， this paper analyzes the theory of theLocality-constrained Linear Coding and Sparse Coding. Statistical model of sparsecoding exhibits a maximum-selection behavior which is the properties of simple cellsin the primary visual cortex. Sparse Coding combines the learning of a sparse codewith a local maximum operation which leads to features that allow for competitiverecognition performance. Based on the principle that locality is more essential than sparsity, Locality-constrained Linear Coding incorporates locality constraint togenerate the representation of the image for classification.This thesis uses SVM for the classification of feature vector.In recent years,SVM has been successfully applied in signal processing, image recognition, genemapping and so on. SVM is based on the the theory of VC dimension and principle ofstructural risk minimization.It is Generally applied to solve the pattern recognitionproblem in scene of small sample quantity, high dimensional and nonlinear situation.Finally, this thesis realizes the image descriptor extraction, coding andclassification in the experiment and analyses effects based on different experimentalcombinations. The result shows that SIFT descriptor combines Sparse Coding canachieve a outstanding result.