Spatial Weighting for Bag-of-Visual-Words and Its Application in Content-Based Image Retrieval

It is a challenging and important task to retrieve images from a large and highly varied image data set based on their visual contents. Problems like how to fill the semantic gap between image features and the user have attracted a lot of attention from the research community. Recently, the 'bag of visual words' approach exhibits very good performance in content-based image retrieval (CBIR). However, since the 'bag of visual words' approach represents an image as an unordered collection of local descriptors which only use the intensity information, the resulting model provides little insight about the spatial constitution and color information of the image. In this paper, we develop a novel image representation method which uses Gaussian mixture model (GMM) to provide spatial weighting for visual words and apply this method to facilitate content based image retrieval. Our approach is a simple and more efficient compared with the order-less 'bag of visual words' approach. In our method, firstly, we extract visual tokens from the image data set and cluster them into a lexicon of visual words. Then, we represent the spatial constitution of an image as a mixture of n Gaussians in the feature space and decompose the image into n regions. The spatial weighting scheme is achieved by weighting visual words according to the probability of each visual word belonging to each of the n regions in the image. The cosine similarity between spatial weighted visual word vectors is used as distance measurement between regions, while the image-level distance is obtained by averaging the pair-wise distances between regions. We compare the performance of our method with the traditional 'bag of visual words' and 'blobworld' approaches under the same image retrieval scenario. Experimental results demonstrate that the our method is able to tell images apart in the semantic level and improve the performance of CBIR.