Traffic Sign Recognition Based On Local Features

With the rapid development of computer technology and the popularity of the Internet, people can purchase their favorite commodities without going out of home. Many popular web sites (such as Baidu, Google, Taobao) are using the text-based search method to find millions of product images corresponding to the text. However, the site labeling of the goods only illustrates some basic information of commodities, such as commodity brand, price, style, which is difficult to reflect the complete features of commodities that the customers need. This thesis uses the kernel methods in image classification based on the content of the product images. Retrieving product images in the specified category, thus it greatly improves the accuracy of the user need of commodities. The specified category is based on the idea of the image classification.This thesis uses the pyramid histogram of words as image feature descriptor and utilizes three kinds of kernel methods for product image classification.(1)Kernel Fisher discriminant analysis, through nonlinear mapping, linear discriminant analysis is transformed into kernel Fisher discriminant analysis, and then being applied to classification of product images.(2)In the nonlinear case, this thesis applies the support vector machine to solve classification problem. Incorporate the different distance formulas with the radial basis kernel function in support vector machine, and form the generalized Gaussian kernel function to improve the image classification effect.(3)According to multi-feature (pyramid histogram of keywords (PHOW), pyramid histogram of oriented gradients (PHOG) and GIST descriptor) kernel fusion method for the image classification, and realize the simulation results of classification system of product image in MATLAB.These three kernel methods are tested under the same experimental parameter settings. In contrast to the various generalized Gaussian kernel, the experimental results show that the chi-square kernel function has a better average accuracy and generalization performance. These three kernel methods are all used chi-square kernel function, the average accuracy of multi-feature kernel fusion method is higher than the other two single kernel methods (kernel fisher discriminant analysis and support vector machine), and the average accuracy of support vector machine is higher than the kernel fisher discriminant analysis.