Theory And Application Of Automatic Scale Selection For Bionic Eyes

Human vision system is able to recognize objects by using multi-scale analysis, and that account for the ability of understanding whole shapes and local details in the meanwhile. Thus it is important to study the mechanism of scale perception, and this research topic remains a hot but difficult point in the domain of biologic vision. At present, techniques of image analysis and recognition have been developed quickly, results in a higher demand of image interpretation, and it is more likely to fulfill the task of modeling image details for the high definition technology being able to supply high resolution images. All these factors are pushing the development of theory and application of local features, therefore more and more researchers pay their attention on this topic. Beginning with analysis of scale selecting mechanism in vision cortex of human beings, this paper innovatively proposed a series of approaches for automatic scale selection, for the purpose of selecting more meaningful scales for features and supporting high level image segmentation. The main contents are as follows:In chapter1, the related study background and significance are introduced. On the basis of referring to domestic and international associated documents, system components, key technologies and the current application research situation of human vision and local feature are summarized. Then the main problems and new challenges in the application and research of automatic scale selection for bionic eyes are analyzed, and the main work of this study is summarized.In chapter2, the relativity of human vision mechanism and scale space is analyzed and the basic theories as well as properties and limitation of Gaussian scale space are expounded. Finally, the fundamental principle and theories of automatic scale selection in Gaussian scale space are introduced.In chapter3, an improved method named EP-DoG is proposed to resolve the problem of feature shifting from a scale level to another scale level. By analyzing the behavior of feature point as the analysis scale changing, an extreme path will be searched to describe the shifting route of feature point, and the characteristic scale will be chosen if the feature reaches a max DoG response in this scale level. The comparative experimental results show that the proposed method gain better performance than DoG.In chapter4, an innovative method named SMM is proposed to select scale for corner points detected in scale space, and the scale level in which the shape of ellipse determined by second moment matrix of a corner point remains unchanged would be the characteristic scale for this corner point. Besides, another innovative method named GFT is also proposed to select scale for features detected without scale space, and the round area around the feature point which reaches a min GFT energy value will be regarded as feature area and the radius of this circle is considered characteristic scale. The comparative experimental results show that these two methods perform better than existing methods and are both effective and robust while facing changes like rotation, zooming and blur.In chapter5, a novel method is proposed to select edge scale automatically, and this method can be used to split complicated interleaving edges into more meaningful edge segments that can be classified easily according to their features. Firstly, edge extreme path in Gaussian scale space is searched to obtain characteristic scale for each edge point, through the way of calculating maximum distance that edge point travels from one scale level to adjacent one and analyzing the four forms of extreme path evolution. Then the interleaving edges are split into pieces with different scales according their scale histogram and combine the edge pieces connecting with each other through extreme path in scale space. Experimental results show that the proposed method is effective in edge segmentation based on characteristic scales and saliency of edge.In chapter6, a segmenting method based on matching of local feature with automatic scale is presented for splitting overlapped moving objects’region into pieces, and each piece corresponds to only one moving object. To gain better result of motion segmentation, two novel motion segmenting methods, which take advantage of both GMM and TD, are proposed. The GFT method is employed to select scale for motion feature points and feature descriptor is utilized to match features in two frames of video sequence, therefore the histogram of displacement between matched features is constructed to separate objects moving in different speed and direction. Experimental results show that the proposed method is able to separate overlapped moving objects effectively.In chapter7, the major work of the thesis is summarized. The conclusion and innovations of this thesis are introduced. Finally, the future development topics are presented in order to provide guidance for researchers, who are interested in such kind of projects.