Research On Shape-based Image Retrieval Algorithm

With the rapid development of the internet and multimedia techniques, the morewidespread use of image in all walks of life and the image multimedia informationsources are expanding. Therefore, digital image has become a kind of indispensableinformation resources. Recently, various large capacity image or video database hasshown a growth of geometric, how to utilize such resources fast and effectively hasbecome an urgent problem to be solved. Content-based image retrieval technologiesintegrated image engineering, digital image processing, artificial, intelligence and otherrelated discipline knowledge. Content-based image has become the mainstream ofimage retrieval technology research and the focus in the domestic and foreign research.Shape-based image retrieval is one of difficult important part. Shape is an importantway to describe high-level vision features, compared with the image’s inherent featurescolor, texture, spatial relationships; shape feature has great superiority, such as theobject’s shape does not change with changing in the surrounding environment and oftenclosely linked with the target object, etc. In this dissertation, lots of exploratory researchhas been done around some key techniques of shape-based image retrieval. Maincontents and contributions of this dissertation are as follows:(1) Several key techniques and methods of CBIR are deeply analyzed anddiscussed: color-based image retrieval, texture-based image retrieval, spatialrelationships-based image retrieval and shape-based image retrieval, especially forshape representation and shape matching methods in shape-based image retrieval.Inaddition, the evaluation criterions for image retrieval algorithms were also introducedcorrespondingly.(2)This chapter presents a novel approach for contour-based shape retrieval. First,the object contour points got by contour tracing and denoising, corners points andsampled points are come from the contour points which can represent the shape verywell, and two vector sets are obtained with leading vectors from each point to thecontour centroid point. Then, the polar histogram of the directions and distances in thevector sets is used to describe the shape. Finally, the similarity measurement criterion ofthe two contours is given combined with the polar histogram of the two feature points.Compared with the traditional algorithms, the experiment result shows that ouralgorithm is efficient and has better performance in the robustness to the scaling, rotation and translation.(3) This chapter presents a novel approach for edge-based shape retrieval forimproving the conventional algorithm of image retrieval based on edge orientationhistogram. First, we extracted the image edge by the algorithm of the wavelet transformbased on maximum model. Then, we got the coherence edge direction histogram whichdepends on the edge point’s directions and correlation between neighboring edges. Atlast, we counted the similarity by circular shift, and then applied it to image retrieval.Experimental results show that our algorithm has the robustness against illuminationchanges, position translation, scale and rotated image and a superior image retrievingperformance.