Category Prediction For Commodity Image On E-commerce Platform

Recent years, as E-commerce platforms (e.g. Amazon, Taobao, etc.) mature, a growing number of Internet users chose to shop online. For better shopping experience, E-commerce companies treat images as an indispensable part of displaying commodities. In these platforms, large scale commodity images are uploaded by online shops. Due to the appearance of massive commodity images, commodity can be organized and accessed in a new and more effectively way. For example, users sometimes want to use images of one commodity to find similar commodities. But this is not a trivial task because there is tremendous difference between low level visual features and real world commodities. Therefore, how to predict the category of commodity based on its image has been an important research problem. In fact, the category prediction problem based on image classification problem. Compared with traditional image classification applications, category prediction problem has to deal with massive images and massive categories.In this paper, an integrate solution that is comprised of scalable algorithms and models are proposed to solve the problem.In this automation system of solving commodity category prediction problem, leveraging socially tagged images is a fundamental task. Because of the uncertainty of between images and their annotation category tags, irrelevant images are returned by image search engines. Irrelevant images filtering can be considered as a single category prediction problem in large scale data. To solve this problem, a single category prediction scheme based on graph cut and KNN algorithm is performed.To describe large scale commodity images, bag of visual words feature is extracted to solve massive categories prediction problem. However, naive clustering algorithms can’t meet the visual dictionary training situation which has large scale input visual interest points. Thus, a speed-up K-means algorithm is proposed in this paper. Triangle inequality is applied to reduce large number of redundant calculation. To get better initial centers of K-means algorithm, Hierarchical K-means algorithm is performed.To solve the prediction problem on massive categories, a novel scheme is proposed. SVM classifier based on χ2-RBF kernel is used for one-versus-one classification problem. A hierarchical prediction scheme which uses the idea of double-elimination is performed for the massive category prediction task.Based on these algorithms and models, our approach of commodity image category prediction have good time efficiency. So it can be used for the real world large scale commodity image data. And experiments on commodity images dataset shows that our image classification method can perform effectively.