Image Retrieval Based On Visual Attention Model

With the rapid development of the network technology and the popularity of themultimedia digital information acquisition equipment, large number of digital images,video, and other massive multimedia information are generated every day. In order toobtain the images users needed from the image database, content-based image retrievaltechnology is becoming a popular research. The underlying features of image (color,texture and shape) are typically extracted in traditional image retrieval system. But thereare some limitations in this global-feature based approach, it ignores that differentregions of the image have different attractiveness to human visual system. In this paper,the visual attention model is adopted to find the region of interest in image for imageretrieval. This paper has researched the visual attention model, saliency regionextraction, image feature extraction. The main contents are as follows:1、Visual attention model is the basis of the saliency region extraction of the imageand it determines the visual effect of the saliency region extraction. The flow chart ofSong’s visual attention model is adopted in this paper, including Gaussian pyramidgeneration of the basic feature channel, the feature map generation and the feature mapsfusion. However in Song’s model the fixed window is used for center-surroundcomputing in the stage of feature map generation, which causes the model unable todepict the inner part of the saliency region, especially for large-scale target area. Inorder to overcome this shortcoming, this paper improves Song’s model. The image isassigned different scale computing window according to its resolution in the imageGaussian pyramid sequences and the high resolution image will be assigned large scalewindow. The experiment shows that, compared to Song’s model, the modified modelcan better depict the inner part of the saliency region.2、In terms of saliency region extraction, as the saliency map can only describe thesaliency of the pixels but not the region, this paper combines the image segmentationalgorithm and the saliency map. Firstly, the regions of the image are obtained by thesegmentation algorithm; then the saliency of each region can be calculated with thesaliency map; finally, the saliency region of the image can be obtained by threeselecting rules.3、In terms of image feature extraction, the background area of the image is alsoused for feature extraction in global-feature based method, which can’t accurately describe the main content of the image and it’s also a waste of computing resources. Inorder to solve this problem, in this paper, only the saliency regions of the image areused for feature extraction and the background region is ignored. However, there arealso some pictures in which the main content and the background are doped. For thisreason, the modified BDIP-BVLC feature is used in this paper. The binary image of thesaliency region is acquired by the morphology operation applied to the saliency map. Inthe binary image, the partial block with the value of0will not participate in the featureextraction.The experiment results show that the proposed method is better than the global-feature based method in terms of precise and recall, and it also outperforms the methodbased on saliency weighted; furthermore, the modified BDIP-BVLC feature based onthe saliency region has higher precise than the feature based on the saliency region.