Kernel Methods For Image Object Recognition

Image target recognition is one of the most attractive areas in the field of image understanding and computer vision. It has wide usability in commercial area, civil area and military area. The pattern analysis technology based on kernel methods is a new technology of good performance and strict theory, and has been used in image target recognition. This thesis studies the specialities of image target recognition and pattern analysis technology based on kernel methods, does some exploratory researches on the construction of image kernel functions and multi-class classifiers.Traditional image target recognition methods is studied first, and then the image distance is proposed, which can replace the vector distance in traditional RBF kernel function. Based on this, the image kernel function which can accept two-dimension image data as input data directly is proposed. This new kernel function is constructed based on the traditional RBF kenel function, and can be used just the same way. It involves the information of energy, luminance, position, and structure of the two images, and can compare them more accurate than traditional ways which are only based on vector distance.The redundant-blocking image kenel function is proposed in chapter three to increase the stability of image kernel, by changing blocking strategy and adding redundant information into consideration.In chapter four the discrete-blocking image kernel function is proposed which can increase the practicability while the ones proposed in chapter two and chapter three have strict demand for input data. This new kernel function selects the candidate blocks using ROI technology, compares the relationship of positions of them using two point-matching methods proposed in this thesis, and decreases the limitation of input image data. The experiments on COIL show that, the new kernel function can be used on the image objects which can have rotation, scaling and translation changes, even those are sheltered partly.A multi-class classifier with rejection based on SVDD is proposed in the last chapter, which is different from traditional multi-class classifiers in theory of recognition strategy. It aims at recognizing objects instead of distinguishing them, acts more close to human than traditional ways. A new measure based on generalization ability is also proposed, which can represent the more essential relationship between samples and categories. At last, the multi-class SVDD with multi-level structure based on this new measure is proposed, and can get excellent results in not only bi-class but also multi-class recognition experiments.