Research Of Classification Algorithm For Product Image Based On Content

The rapid popularization of Internet and continuous improvement of the infrastructure has brought a great development of domestic e-commerce.At the same time,the network shopping has become an important trend on a global scale.With the growing of size,products type and quantity in the virtual network has increased dramatically,how to show the abundant commodity information to customers becomes an important issue in the process of the intelligent electronic commerce.Product image is the main information carrier of goods on the Internet,so the automatic classification based on contents of the commodity image has important research value.It can effectively improve the overall effectiveness of the E-commerce market,such as quickly retrieve commodity information for both parties to the transaction,reasonable set strategy and personalized recommendation.This paper proposed an improvement classification algorithm of product image classification according to the characteristics of the commodity image,based on the current commodity image classification.The main work is as follows.(1)Self-adaptively separate images into foreground areas with much information and blank background areas according to characters of product images.Foreground could be separated into feature-areas with distinct local features and supplementary-feature-areas with gentle gray-scale variation.Extracting sparse SIFT in the feature-area and dense SIFT features in the supplementary-feature-area while ignoring the rest areas.To generate the final image feature,the sparse SIFT and dense SIFT are encoded by SCSPM and combined by the fusion function.With the SVM classification,the method proposed in this paper can get better results in product image classification.(2)For traditional SPM method although recorded the spatial location information of the image,but it can't reflect the classification ability of visual words in specific location.This paper proposes a SPM space weighted method based on entropy.The probability of different visual words in different categories may be different.According to entropy concept can be used to describe the classification ability of words in different areas.So calculating the word weight of visual words in different area of classified information can improve the discrimination performance of visual words.(3)For some of the commodity image sets with a single SVM classifier classification accuracy is difficult to further improving.This paper proposes a weaker SVM as weak learners for cascade AdaBoost classifier.In each round of training component classifier,samples would get weight which means each component classifier to its attention.By adjusting the weights,classifier will focus on the sample points which are more easily classified into the wrong sets.As a result,It could be get better classification performance on the commodity image sets.In this paper,MATLAB was used to simulate the process of commodity image classification.Experimental results show that this method can be effective for the product image classification and the average classification accuracy rate reached 87%.