Product Image Classification Based On Multi-Features Confusion

Since the success of e-commerce website development and the rapid popularization of network multimedia technology, online shopping has become a convenient, fast, inexpensive and fashionable way of shopping. However, the ensuing image data is exponentially growing, and effective management of such ultra-large-scale multimedia data as well as providing fast accurate retrieval service is a challenging task. Currently, search services of online shopping are still dependent on text-based search engines by tagging and associated merchandise basic information. Huge problem exists from lack of further annotation for situations that the user finds it difficult to accurately describe the style, pattern, shape and other unique attributes. Introducing content-based image automatic classification to relieve pressure on commodity image database management and improve retrieval efficiency of consumer goods is the current urgent needs of e-commerce.In this paper, based on merchandise images of online shopping, we build a product dataset with specific attributes manually labeled. Plenty of experiments concerning different characteristics of the product images are carried out for attributes classification. The main contents and contributions are as follows:Firstly, shopping users mostly concern about two important attributes:color and pattern. According to the color, texture and shape distributes analysis upon the original and rough set of online products images, we extract color moments and color histogram features of the goods images under HSV color space, and choose local binary patterns, gradient-based local binary patterns, binary gradient contours and histogram of oriented gradients to describe the texture and shape information, which can be the representation of pattern attribute. Experiments have proved the classifying performance of these features.Secondly, the paper describes the classifying detail of different underlying features on color and pattern attributes classification. Feature-level fusion is committed to construct composite feature vectors for the different characteristics of two attributes. Through experimental tests, classification performance changes within different feature combinations. It has shown by experimental results that the accuracy of product image classification has been partly improved.Lastly, although each feature has its unique classification performance, the associated utilization of different features and cooperated decisions of classifiers have not been taken into consideration. Since using specific kernel in different classification algorithms can facilitate features’ outstanding performance, we introduce the multiple kernel learning methods to improve classification decisions. Certain number of experiments is designed to make best use of color, texture and shape features which collaboratively indicate the product image attributes. By comparing several sets of experimental results, performance analysis of features has been conducted in the classification of multiple kernel learning.