Interactive Image Retrieval Algorithm Based On Target Area Feature

With the popularity of Internet and the development of multimedia technology, digital imageplays an increasingly important role in people’s daily lives and work. However, as the scale ofdigital image database is growing at a shocking speed, how to retrieve the large number of imagesquickly and efficiently becomes a very important subject. Content-based image retrieval (CBIR)comes into people’s attention and quickly become a research hotspot. CBIR technique pays moreattention to characteristics of the image. Firstly, get the feature of query image. Secondly, comparethe feature with the other images’ corresponding features to get the similarity. At last, display theimage with higher similarity as the retrieval result to the user. But traditional CBIR typically facetwo problems: retrieval accuracy is not high and efficiency is too slow.In a view of the two problems, this thesis gives three improvements which concentrate on theInteractive image retrieval algorithm based on target area feature. Main works are as follows:(1) In traditional CBIR, people use global features to describe the image. But in fact imagealso contains redundant information, and users are more concerned about some local area of theimage. Therefore, this thesis focus on the image retrieval based on local target area and proposedtwo methods, one is corner detection with auto-adaptive threshold curvature enhancement, and theother one is target area extraction based on corner curvature. The algorithm judges the importanceof corners by curvature. Firstly, select the true corners through the auto-adaptive threshold.Meanwhile, enhance their curvature. Secondly, determine the image’s center of gravity by thecorners with larger curvature. At last, extract the target area by regarding the center of gravity asthe centroid. The image retrieval experimental results show that this algorithm can not only detectthe corners and extract the target area effectively, but also provide higher precision of imageretrieval compared with the traditional one. (2) In order to improve the efficiency of the search algorithm, this thesis adds cluster analysisalgorithm. Using FCM to classify the image database, get multiple cluster centers and centerimage of each class. Then compare the similarity of query image with each center image todetermine which category it belongs to. Finally compare the similarity of query image with allimages within the class. Experimental results show that the cluster analysis algorithm reduces theretrieval scope and improves the efficiency of the retrieval algorithm.(3) The reasons of image retrieval accuracy is not high except for the too much redundantinformation, but also due to the differences between the low-level image characteristic andhigh-level human semantic concepts. So this thesis adds relevance feedback to allow users labelcorrelation of each retrieval result. And then system adjusts the strategy according to the feedbackto guide a new retrieval. Experimental results show that the relevance feedback makes furtherimprovement on the precision and meets the demand of the users.