컬러와 에지 정제에 기반한 잡음 불변 영상 검색

Abstract

Standard histograms are very efficient image features and they are widely used for content based image retrieval. They are insensitive to small changes in the image appearances. Due to the fact that they are not very unique and robust, many images of different appearances can have similar histograms. Hence, the main disadvantage of histograms is that they provide coarse characterization of an image. Color histograms, too, are widely used and suffer from the same problem. Color histograms are employed in systems as QBIC [1], Chabot [2] etc. In order to avoid these shortcomings of histograms, a modified scheme based on histogram refinement method [3] is presented. The histogram refinement method says that the pixels within a given bucket is to be split into classes based upon some local property, and these split histograms are then compared on bucket by bucket basis, just like normal histogram matching. The pixels within a bucket with same local property are compared. Therefore, the results are better than the normal histogram matching. Hence, not only the color features of the image are used but also spatial information is incorporated to refine the histogram. The results are obtained by testing the algorithm on a database of images provided in MPEG-7 data. The technique defined in this paper is based on histogram refinement, and is called as ‘color and edge refinement’. The proposed method splits the pixels in a given bucket into several classes similar to histogram refinement method. These classes are quantized and compared based upon their color and edge coherence vectors. Color refinement is based on histogram refinement method suggested by Pass and Zabih [3]. However, edge refinement involves a different mechanism. Pass and Zabih state that color histogram buckets are partitioned based on spatial coherence. A pixel is coherent if it is a part of some sizable similar colored region, otherwise it is incoherent. Therefore, in color refinement, the pixels are classified as coherent or incoherent within each color bucket. If a pixel is part of a large group of pixels of the same color which form at least one percent of the image then that pixel is a coherent pixel and that group is called the coherent group or coherent cluster. Otherwise it is incoherent pixel and the group is incoherent group or incoherent cluster. Two more properties are calculated next for each bin. Firstly, the numbers of clusters are found for each case, i.e., coherent and incoherent case in each of the bin. Secondly, the average size of clusters is computed. Hence, there are six values calculated for each bin: one each for percentage of coherent pixels and incoherent pixels, number of coherent clusters and incoherent clusters, average of coherent cluster and incoherent cluster. Edge refinement is also based on histogram refinement method. The buckets are formed based on edge direction in the pre-processing stage, and then, total number of on pixels is computed for each bucket. Sobel operator is used to compute the horizontal and vertical edges. Hence, two buckets are formed, which are further classified as coherent or incoherent. Four more values are calculated in each bucket. The numbers of clusters are found in each bucket by 8-neighborhood rule, and then the average of the cluster is computed in each bucket. Moreover, the number of straight edges and slanted edges is computed. Therefore, for each bin, there are five values: one each for total number of on pixels, number of clusters, average size of cluster, number of straight edges and number of slanted edges. Hence, we can conclude that color refinement method takes care of the color as well as the spatial relation feature. Moreover, edge refinement method implementation to the results of color refinement method presents significant improvement and provides more accurate and precise results as proved through rigorous experimentation. Furthermore, experiments have proved that the proposed algorithm provides better results than the equivalent methods.