Product Image Classification Based On Fusion Of Multiple Features

As is known to all, the emergence of the computer network has produced a profound influence on the structure of the modern society, and the resulting e-commerce has also make profound changes on people’s life style. Increasing people’s shopping way has changing from the traditional entity shop to the shopping website. Accordingly, the commodity images in all kinds of shopping website had a sharp increase, which undoubtedly increased the difficulty for the user to pick the goods in vast amounts of products. Therefore the research of this kind of problem not only has theoretical significance, but also has important practical application value. In order realize the automatic classification for commodity images and promote the intelligent development of e-commerce, the researchers have proposed many image classification methods. Based on the certain research foundation, the author has mainly done the following three aspects:Firstly, according to the analysis and study of the characteristics of commodity image, this thesis proposed to preprocess the images before the image classification, which is detecting the salient region and then split the image. And it will contribute to feature extraction and obtain a more accurate feature vector.Secondly, in terms of the two important properties of color and pattern style, this thesis put forward to use HSV color histogram and color moment to describe the image color, and also SIFT(Scale Invariant Feature Transform) and PHOG(Pyramid Histogram of Oriented Gradients) to represent the pattern style. And in the aspect of design of multiple features fusion method, the ordinary integrated weighting method has improved in this thesis, the calculate method of weight was improved. The reliability of a certain feature of a sample in the class was defined as the weight of the feature.Finally, the simulation experiment was conducted by using SVM classifier, and the method was compared with the single feature method and other multi-feature fusion method respectively. And the comparison results proved that the improved method has higher classification accuracy. This thesis has contributed to the realization of product image automatic classification, and further prepare for the realization of intelligent electronic commerce.