Key Technology Research For Region-based Image Retrieval

In order to retrieve the image from the mass image data on the Internet for users, content-based image retrieval (CBIR) has been a hotspot and difficulty. However, region-based image retrieval (RBIR) is a deeper development of the CBIR, which uses the segmented regions to form the regional feature vectors and allows the regions mutual matching. The technology has not only developed toward to the object level, expressing and indexing image at the level, but also reduced the gap between low-level visual features and high-level semantic features at some extent. Therefore, it has important research value and market application prospect in the image search field.Firstly, the development and technical advantages of RBIR are introduced. And some key techniques of image retrieval are discussed, including region segmentation, feature extraction, region matching, similarity measurement and so on. Secondly, the advantages and disadvantages of fuzzy C-means clustering (FCM) algorithm and integrated regional matching (IRM) algorithm are discussed. Based on above, the corresponding improvement schemes are proposed. Finally, an image retrieval system prototype is designed, using precision and recall as the evaluation criterion to contrast the improved algorithm and IRM algorithm. The innovations mainly focus on the following two aspects. Aiming at the problems of selecting an optimum m value difficultly and image segmentation with weak robustness, a new method is proposed. The threshold method is adopted to segment image at first, and then used the FCM algorithm to further segmenting. Based on above, the probability of a pixel belonging a cluster is defined the distance between them, so each pixel is containing both the color information and spatial information. The time complexity of traditional IRM algorithm is very large and it doesn’t consider the region area ratio factor. So the robustness of retrieval result is weak. Therefore, the improved algorithms about integrated regional distance and significance matrix are proposed. A threshold is set to all the regional distances, then the rest are normalized. It greatly reduces the algorithm complexity. What’s more, taking into account the impact of target and background area ratio on image matching, the concept of "effective distance" is put forward, making the region area feature weight increase and achieve the effect of "the most similar and the highest priority".The experimental results show that the color, texture and region area features are all fused in the improved algorithm, and the similar images are more concentrated, which not only are more consistent with human visual perception, but also improve the precision and recall obviously and have better performance then the other retrieval systems.