Research On Heigh-Level Semantic-Based Image Retrieval

At present, digital image as a representative of visual information increases in the number. How to effectively and rapidly retrieve user required images consistent with its semantics from mass digital information has become a pressing problem. By taking the static natural images as research object, this paper proposes a method of region of interested extraction based on visual attention mechanism and image segmentation,founds a model of image high-level semantic extraction based on support vector machine (SVM) and develops a Matlab-based image acquisition system and a high-level semantic image retrieval prototype system based on web. The system reduces the "Semantic Gap" between the image low-level visual features and users, provides an effective way to image understanding and image semantic acquisition.The main content and conclusion of this research are as follow:⑴In order to extract regions of interest of natural images in precise, an extraction method of regions of interest based on visual attention mechanism and K-means clustering is proposed on the basis of the Itti model. According to biological visual attention mechanisms and cerebral cortex cognitive processes, it extracts the image visual features by using second-order differential function of Gaussian local iteration to generate the comprehensive feature map and the saliency map of image. It adopts k-means clustering methods to divide image pixels into three categories and extract regions of interest of image according to color clustering index generate regional segmentation map and the combination of saliency map and regional segmentation map. Experimental results show that the proposed method is more close to the process of human visual attention which can extract different images background of target area and has certain antinoise ability.⑵On the basis of getting region of interest of images, it makes use of color moments, invariant moments and moment invariant three methods to extract image features such as colors, shapes, and texture low-level. Based on supporting vector machine theory, it designs the structure of semantic classifier, low-level features in the image as input, the corresponding semantic category as output, and the radial basis function as a kernel function. By setting up relevant parameters, training classification model, it can effectively realize image low-level features to high-level semantic mapping, and obtains image's high-level semantic information. Semantic classification experiment is practiced based on the SIMPLIcity image database. The results show that, this model can classify still natural images effectively with an average classification correct rate of 84.4%.⑶On the basis of Analysis of image content and research image high-level semantic access, this paper develops a high-level semantic image retrieval prototype web system. It shows the image semantics access for the whole process, achieves the semantic classification of the image effectively, makes users easier to find out the proper images by using semantic keywords or sample figure The system is tested and evaluated from three aspects including accuracy, response time and load capacity.