Studies On Image Retrieval Based On Visual Attention

Content-Based Image Retrieval (CBIR) is a research hotspot in the field ofcomputer vision. The traditional content-based image retrieval extracts the low-levelfeatures (color, shape or texture pattern) for indexing and then retrieves the image.However, this kind of image retrieval method based on the global features has thesemantic gap between low-level features and high-level semantic information.Thus, this article introduces the visual attention mechanism to the traditionalimage retrieval system based on image content. In this article, the image retrievaltechnology based on visual attention is researched: make use of the visual attentionmodel to extract the regions which most draw human attention as salient regions andextract features in the salient regions for retrieval.The image retrieval method based on visual attention in this article makes up thesemantic gap to some extent and gets the satisfying search results in a series ofexperiments. The key research work is shown as follows:1）First, the CBIR and the research status of visual attention model are reviewedand the global framework and key technologies of the image retrieval system based onthe CBIR are analyzed. On this basis, this article indicates the disadvantage of thetraditional image retrieval system and introduces the visual attention to build theimage retrieval system based on it.2）Second, this article analyzes the defects of several classical saliency detectionmodels and proposes a new kind of saliency detection algorithm: region-level saliencydetection algorithm based on frequency tuning. It takes both the image frequencyinformation and the influence by the spatial relation of the different image regionsinto consideration. The experiment finally shows that the salient regions of inputimage can be effectively detected by this saliency region detection algorithm.3）Then, this article analyzes the structure and principle of the traditional imagesearch engine and brings the visual attention technology into the search engine todesign a kind of image search engine based on the visual attention which helpsimprove the efficiency and precision of image retrieval comparing to the traditionalimage search engine.4）Finally, the image retrieval system based on the visual attention is constructedand the system test and evaluation is made in two kinds of image database respectively. And this kind of image retrieval system can accurately detect the mainsalient regions of input images, which makes up the semantic gap to some extent. Theprecision and recall are both higher than the image retrieval method based on theglobal image features.