Product Images Retrieval Based On Visual Phrases

With the fast-growing of online shopping services, there are millions even billions of commercial item images available on the Internet. The traditional keyword-based search engine has gradually cannot meet the needs of users. How to effectively leverage visual searching to find out the items of users’interests is an important yet challenging task. Nowadays, Bag-of-Visual-Words model is widely used in image retrieval. Combining BOW and inverted file which indexing database image set can greatly reduce the number of candidate image matching. In our method, we proposed a new visual feature-visual phrase which is more distinctive and includes spatial information between visual words. Combining the point feature-SIFT and the region feature-MSER, we define a visual word pair to be a visual phrase if two visual words co-occur frequently in same MSER regions. Besides similar global appearances (e.g. color, shape or pattern), most time users pay more attention to the local style of certain products, thus an ideal visual item search engine should support detailed precise search of similar images, which is the deficiency of current search systems and techniques. In this paper, this paper proposes two retrieval stages in our search engine-global matching stage and local matching stage. In global matching stage, we match two images using visual words and visual phrases in total image region. Then, an interactive manner is provided to users to accomplish a precise retrieval in local matching stage. In the second stage, color feature is also considered. Fusing the results of two stages, the final search result will return to users. The experimental evaluations show the promising results of our approach and system.