Research On Content-based Product Image Classification

In recent years, with the rapid development of computer and the Internet technologies, e-commerce grow up gradually. Online shopping has been accepted by more and more people with its characteristic of convenient and quick, and at the same time more and more online shopping platform has come to the fore. E-commerce websites show products to consumers in the form of pictures, therefore image search is one of the key technologies in online shopping. But because of the growth of image data, the efficiency and accuracy of image retrieval are limited, we must find accurate and efficient management for large-scale product image database. Content-based image classification technology can manage product image in order, retrieving based on this will effectively improve efficiency and accuracy, this paper gives an exploratory research on content-based product image classification.Firstly, we introduce the research status of content-based image classification technology briefly, and then discuss the process of content-based image classification. This paper makes a detailed analysis of several kinds of typical color, texture and shape features, and proposes an improved global descriptor based on traditional image global descriptor, which combines with the characteristics of product image. This feature not only represents global characteristic of product image, but also describes local detail information. Combining with classification methods, experimental results show that our proposed method can effectively improve the accuracy of product image classification compared with other features, and also increases the accuracy of natural image classification significantly.To solve the problem that when the distributions of data from the source domain and the target domain are different and no labeled data in the target domain are available, data from the source domain and the target domain can't share the same classification model, thus it's hard to solve the classification task in target domain. This paper use the knowledge learnt from the source domain to help and speed up learning in the target domain through transfer learning, and implement a product image classification algorithm based on inductive transfer learning. This method can effectively solve the product image classification in the target domain which is lack of labeled data.For multi-class product image classification, this paper proposes an improved automatic feedback sparse representation algorithm for product image classification. Considering the success that sparse representation got in the field of computer vision and pattern recognition, and compared with the traditional machine learning method, sparse representation does not need training stage, and the calculation is more simple and fast, therefore we consider to apply it in content-based product image classification. Combining the characteristics of product image with the data demand of sparse representation, first we preprocess the product image to extract the object region and normalize the size. According to the correlation of image features, we feedback the initial classification result through feature matching automatically, thus forms an improved automatic feedback sparse representation algorithm for product image classification. Experimental results show that this algorithm performs well in product image classification.