Visual Words Refining Exploiting Spatial Co-occurrence Table

Bag of visual word (BOVW) model is widely used to represent the images in Content based Image Retrieval (CBIR). Spatial information is lost during the quantization from visual features to visual words in BOVW. A lot of researches have been committed in incorporating the spatial correlations of visual words into BOVW model. In this paper, exploiting the spatial co-occurrence of visual words, we build visual word co-occurrence table over the entire dataset and propose a hierarchical clustering approach to group visual words those usually co-occurrence into clusters as new visual words. Any two clusters are correlated via the calculation of the conditional probability of the multiple visual words in them. Utilizing the correlated clustering results, we succeed in refining the visual words and reducing the similar words' distinction in image ranking. Experimental results have demonstrated the effectiveness of the proposed scheme, without incurring any additional cost on the BOVW model.