Research Of Image Retrieval Based On Target Region Features

Content-based image retrieval technology has become a hot issue of domestic and foreign research field since it can retrieve the desired image from the large-scale image database effectively. It has important applications in medical image processing, security management, satellite remote sensing image processing and so on.There are two main methods of content-based image retrieval currently, one is based on the global features, and the other is based on the local features. The former method is simple, fast response but lacking high-level semantic information so it can not be well meet the user’s demand. For example, when user is only concerned with the local area of interest, the method based on the global features will be unable to highlight the significance of the interest area, resulting poor retrieval results. The method based on the local features overcomes the deficiencies of the former method because it understands and describes the images from the semantic level, more in line with the cognition of the human vision. So this paper focuses on a class of image retrieval methods based on the target area features, the main contents are as follows:A novel method for image retrieval based on feature points of contour is proposed. Firstly, mean shift algorithm is applied to segment an image into clusters and a main cluster is taken as a target object. Comparing with other similar methods, our method excludes the influence of the background image to the target image effectively. Then, the local maximum points of curvature of contour of the target are calculated and the object is described with its eigenvector comprised of these feature points. Finally, an effective matching mechanism based distance between the eigenvectors of a retrieve object and a retrieved object is used to carry out matching or recognition between two objects. Experimental results show that the proposed method has a higher recall and precision to the single target images. Furthermore, the proposed method also has greatly reduced computational time.Another image retrieval method based on user’s interest region is proposed. Firstly, an adaptive bandwidth mean shift tracking algorithm is applied to detect the user’s interest region. Secondly, the image of block difference of inverse probabilities is extracted by the block difference of inverse probabilities model that is based on the HSV color space. Thirdly, the salient points of the candidate region are extracted according to the block difference of inverse probabilities image. Finally, using color and space features of the salient points as the descriptor of the object to achieve matching among objects. The algorithm first detects the canditate target region, removing the interference of the background image and improving the accuracy of the characteristic expression of the candidates, then, the eigenvectors of the significant points are constructed by the interest region and the candidate region, in favor of the signature expression of the image and establishing the foundation of the further analysis and processing of the image. Experimental results based on the Corel image database show that the proposed algorithm has significant improvements on the performance indicators compared to the approaches that based on the global features.