Research And System Implementation Of Sketch-Based Image Retrieval

With the rapid development of internet and multimedia technology, there has been an explosive increase in multimedia information. How to retrieve the expected information among the huge multimedia information has become one of the most difficult problems. Content-based image retrieval (CBIR) is one of the main research fields in multimedia information retrieval, which takes the appropriate example image as the query and retrieve the most similar images to the query. The difference between content-based image retrieval and text-based image retrieval (TBIR) is CBIR can reduce the manual operation and provide a more objective way to retrieve.With the wide spread of Smartphone and tablet PC, the human-computer interaction mode has been changed a lot and touch has become the main input mode. People can draw the image sketch of the things in their mind on the touch screen. The sketch describes the outline information, which is more appropriate for human perception than texture and color information, and therefore, using sketch-based image retrieval method to retrieve images with similar shape to the query on large-scale datasets has attracted much attention.This paper introduces the background and significance of sketch-based image retrieval, and summarizes the current researches at home and abroad. According to the general procedure of sketch-based image retrieval, we introduce its preprocessing and feature extraction. In the preprocessing, Canny Detector and Berkeley Detector are compared. As for the global features of outline images, EHD, GIST and HOG are discussed. In the local features and retrieval structure of outline images, this paper introduces SHOG, IOCM and retrieval structure of local descriptors.Outline images do not include color information and light effect, so this paper combines the HOG principle with outline images and analyzes the HOG feature distribution characteristics of outline images. Among the various oriented gradients of cells in outline images, there is one gradient whose amplitude is much higher than others and it can indicate the rough tendency of outline, and therefore HOG principal component features are proposed. To improve the efficiency and accuracy, HOG principal component features are iteratively encoded into binary codes, and Hamming distance is adopted to measure the similarity. The experiment has demonstrated the superiority of the proposed method, including the higher precision rate and faster speed. In order to provide an interactive way for users, this paper utilizes WPF structure to design and realize the sketch-based image retrieval system. The system interface consists of system operation, drawing board, painting tools and result display. Users can generate the outline of the query either by drawing or copyingThis paper is supported by National Science Foundation of China (Grant No.70890083).