With the vigorous development of online shopping,the number of product images on the Internet is increasing at a geometric growth rate every day,and simple text search for products can no longer meet the needs of users.And far-reaching research significance.Different from the common image classification,commodity image has different characteristics and attributes from other types of images.Therefore,it is of great significance to provide a simple,efficient and targeted commodity image classification technology for commodity database management and reducing user retrieval time.Based on the current feature description method,the description ability of commodity image is insufficient.In this paper,according to the characteristics of commodity image,an improved algorithm is proposed based on the characteristics of the Bag of Visual Words model.The feature formation process of Bag of Visual Words includes four steps:1)local region extraction;2)feature region description;3)feature clustering;4)image feature descriptor formation.According to the characteristics of commodity images,this paper makes improvements in the three steps of Bag of Visual Words features:(1)In order to solve the problem of unreasonable distribution and small number of local regions extracted by traditional word packet feature model,a multi-scale local region extraction algorithm based on wavelet decomposition is proposed;(2)Aiming at the problem that the traditional bag feature model lacks color features in the description of feature regions,a fusion feature of surf descriptor and color vector angular histogram feature is proposed;(3)In order to solve the problem that the commonly used spatial representation methods do not combine the distribution characteristics of commodity images,a diagonal concentric moment spatial representation method is proposed in the process of image feature descriptor formation.Combining the characteristics of commodity images above to improve the features extracted by the Bag of Visual Words method,and using SVM multi-classifier for training and classification.The experiment uses 10 sub-data sets of the PI100 data set.The experimental results show that the classification effect reaches 89.73%under the combined effect of the three improvements,which is higher than other Bag of Visual Words models. |