A words-of-interest model of sketch representation for image retrieval

In this paper we propose a method for sketch-based image retrieval. Sketch is a magical medium which is capable of conveying semantic messages for user. It’s in accordance with user’s cognitive psychology to retrieve images with sketch. In order to narrow down the semantic gap between the user and the images in database, we preprocess all the images into sketches by the coherent line drawing algorithm. During the process of sketches extraction, saliency maps are used to filter out the redundant background information, while preserve the important semantic information. We use a variant of Words-ofInterest model to retrieve relevant images for the user according to the query. Words-of-Interest (WoI) model is based on Bag-ofvisual Words (BoW) model, which has been proven successfully for information retrieval. Bag-of-Words ignores the spatial relationships among visual words, which are important for sketch representation. Our method takes advantage of the spatial information of the query to select words of interest. Experimental results demonstrate that our sketch-based retrieval method achieves a good tradeoff between retrieval accuracy and semantic representation of users’ query. Keywords— Image retrieval, Sketch representation, Bag-ofvisual Words model, Words-of-Interest model, Markov chain model