Weighted spherical soft assignment and optimization for image content representation and retrieval

With the development and popularity of digital multimedia, content-based image retrieval (CBIR) has been a hot research topic in retrieval field. However, the multimedia information, especially in the picture and the image, is increasing rapidly, which presents a huge challenge to image retrieval systems. In order to obtain accurate and creditable results, the image representation is the key point. In some previous publications, the vector of locally aggregated descriptors (VLAD) is a new technology which enhances the image representation. However, the use of hard assignment in VLAD does not consider code word uncertainty and plausibility. In this paper, the spherical soft assignment (SSA) is used to enhance the image representation. Since SSA does not consider the gap between low-level features and high-level semantics, relevance feedback has been applied in SSA. The first retrieval results are used as training samples. Users mark positive and negative cases and treat them as class labels of the training samples by SVM learning. Then a classifier is constructed that suits for user´s query intention and with which all the images in the image database can be classified with the classifier. If images are divided into positive class, the distance between each image and the hyper plane is calculated. Then we will remark weights which are used to retrieve again. Results show that it can improve the precision of retrieval results.