Image Features Extraction Based On Similar Contents And Concepts

With the conclusions that image retrieval has made a rapid progress and a mass of theories, image feature extraction technology, the fundamental work of image retrieval, becomes more and more important. In accordance with the needs of image retrieval, this dissertation makes some useful research on the features of color, texture, interest points and spatial distributions. After that, these features are combined to form a new method to extract the common features among the images with the same contents and concepts. And the common features are the useful building blocks of semantic image retrieval.This dissertation can be divided into two parts, one of that is the feature extraction; the other is the common features expression.In the first part, the features of color, texture and interest point are covered.1) Color feature extraction: In this chapter, the gauss mixture vector quantization method is used to build the color histograms. The GMVQ method can capture spatial relationship among the pixel color intensities, and it is better than the traditional color histograms. So this method can provide a more appropriate way of exploiting the spatial difference between two images with the same color distribution. In this chapter, a lot of examples using gauss mixture vector quantization are given to validate the advantages. This method can also be helpful to extract the color features in image retrieval, and this result is also validated by 3 querry examples.2) Texture feature extraction: In this chapter, the Tamura method for texture features is adopted. Three features of Tamura method, i.e. coarseness, contrast, directionality, are the main subjects studied in this chapter, and the features are all represented by histograms respectively. And some improvements on feature expressions and similarity measurements are also introduced to get better applicability in the process to extract texture features. A self-adaptive histogram is built for the coarseness features in order to follow the size variations of the images. Meanwhile, nonlinearity logarithmically quantization method is introduced to express the contrast features, and this improvement can lead a better explanation of man's subjective view. After that, the simulation examples in this chapter also give the same results. The similarity measure is to combine the three corresponding texture features'distance with their weighted sum. The evaluation examples on image retrieval are given at the end of this chapter.3) Interest point feature extraction: In this chapter, the interest point detector is the Harris detector, and then it is combined with the local moment invariants to represent the interest point features. This feature shows the distinct advantages under the variations of light, viewpoint, and scale etc. And these advantages are also validated by some examples in the same time. The similarity measure method of interest points between two images are given according to simulation results. Some image retrieval examples are carried out in Matlab to show the performance of this method.In the second part, both the region similarity evaluation method and the common features extraction method are introduced. Region segmentation is the beginning of this process, and in this chapter a simple and efficient way of region segmentation based on color and texture are well explained. The spatial distribution's distance measure of different images is complicated in general, but a simple and coarse way can be attained by using a thumbnail image, and this way is studied in this chapter. When many kinds of features are in hand, some feature subsets, which can represent a specifical scene most appropriately, must be chosen to reduce the complexity of image retrieval. Through two application examples, the steps to extract the common features are listed in this chapter to show the main concept expression method. The common features, which are marked in the images with the same contents, can directly give the good effects and feasibility of this method.