The Method Of Image Retrieval Based On Visual Attention Model And Gist Features

Compare with using words to describe something, an image seems more vivid obviously. With the development of multimedia technology, a large number of data is saved as an image. However, it is a big challenge that how to find the image required by the user from a database because an image contains rich information. So, image retrieval technique was born.According to the history, image retrieval is divided into text-based image retrieval and content-based image retrieval. From a user perspective, the results suit human sense returned by the content-based image retrieval more than the text-based image retrieval. Traditional content-based image retrieval uses low-level image features such as color, texture and shape as rationale. But because there is a significant gap that cannot be ignored between low-level features and high-level semantic, retrieval performance is not satisfactory based low-level features. Visual attention mechanism overcome the "semantic gap" in a certain extent, at the same time, the scene perception information contains semantic information which describes the spatial structure of objects in the image. So, if they are introduced into the system of image retrieval may get better results.In order to simulate the visual attention mechanisms and scene perception of human visual system, to overcome the "semantic gap" in a way, and to propose a feature can describe image content better, we establish a calculation model of visual attention and Gist, use a complementary calculation model imitate the visual attention and the scene perception. Add edge information for Itti visual attention, after obtained the visual attention information, manipulate, and further obtain Gist features. The resulting features include both visual attention information and scene information of an image, so can be used to describe an image better.After the retrieval method is identified, we should establish an image database for experiments. There should be a large number and different categories of images to test the robustness of the retrieval algorithm. Experiments retrieval a same image with several retrieval methods, compare the experimental results and summarize. The paper retrieval ten categories image for multiple times, for overcome the impact of individual examples to experimental performance. Experiments show that, the retrieval system proposed by this paper has excellent performance for retrieval images which have a single salient area and have no background or have a simple background. And for other complex images, the retrieval algorithm also can get better performance.This article implements and designs the experimental environment on the plat of Microsoft Visual Studio2010and Microsoft SQL Server2008software, which services for related research.