Object Extraction And System Implementation Towards Product Image Search

With the popularity of e-commerce websites, online shopping becomes an attractive, convenient and efficient shopping way. How to efficiently retrieval large scale product images becomes the hot topic in academia and industry. At present, Google, Alibaba etc. have provided visually similar product image search applications, but they can not achieve good performance because these search engines directly extract global vision features such as color, texture, and contour which will be influenced by the noises of cluttered background. In order to improve the search performance, we must wipe off the cluttered background and extract the real product object. The main purpose of this thesis is to solve the product object extraction problem, and two algorithms are proposed to improve retrieval performance. The main research work and contributions are as follows:Firstly, a principal object extraction algorithm is proposed to get products from images without models. To highlight the product for sale, a product is usually placed at a prominent position (that is, close to the middle of an image) to catch customers’ eyes though the background of product images is complex. In addition, the size of the major object is usually not too small. Based on these intuitions, the author uses the graph-based segmenation algorithm which can recognize the object edges and spatial information of product to find the principal object.Secondly, a multiple object extraction algorithm is proposed to gain objects from images with models. Though the models in image make it harder to find products, in another point of view, it’s an obvious clue to help us to find clothes. In this approach, first, get the rough area of cloth by using the skin and face information of models; second, use Gaussian mixture model to describe the statistical distribution of cloth (foreground) and background; in addition, the spatial information of product is taken into account to modify the models; then the refined models are used to gain cloth in images.Thirdly, a product image search system is implemented. Using the above two algrithms to wipe off the background and find products from the cluttered images, then extract the color and SIFT (Scale Invariant Feature Transform) features of the products. In front of the system, grab cut segmentation algorithm is used to interactive with users, and at the back of the system, the Euclidean distance and BoW (Bag of Words) are used to match color and SIFT features respectively. Finally, the experiments tell the algorithms proposed in this thesis can work efficiently, and on the other hand, they prove that the search strategy do improve the accuracy performance.