Image Retrieval Based On Higher-order Features Extraction

With the development of electronic commerce, online shopping has been incorporated into our daily life. Product retrieval is a key part of the online shopping, directly affecting the user experience. With image as input, searching out the required information from the internet has becoming more and more valuable in commercial application and scientific research.There have developed many methods for visual image search recently. Any image processing operations all need to extract features from images to represent it. To find changes which are not sensitive to visual characteristics is difficult. Local features such as SIFT are often used to represent images for its good invariance under several image transforms. However, the weak discrimination of a single feature results in a large number of mismatches between local features. In addition, in the classical Bag-of-Words model which is commonly used in image retrieval, the local feature is quantified to visual word, and the discrimination is further reduced. In addition, due to the strong 2D structural characteristic of image, geometrical relationship is very important information of image content. However, many existing methods ignore the spatial relationship between the features.The purpose of this paper is to find an effective way of the combination of features in local region. It enhances the discrimination of feature to improve the precision of image retrieval by means of finding local spatial relationships of visual words. We proposed an image presentation framework which based on higher order features in this paper. The framework adopts the AP clustering algorithm to group local features in image together to form the visual phrases to improve the discriminative ability of features, and then define a robust matching criterion on it in order to improve invariance ability. In this way, the higher order feature is more discriminative than single local feature, and with higher repeatability. By experimental verification, compared with traditional methods, the proposed method improves the precision of image retrieval effectively.