Application Research On The Kernel Fisher Method And Its Component In Image Recognition

In recent years, with the continuous development and mature of the computer technologyand pattern recognition theory, technology of image recognition is becoming one of the hotissue of the field of artificial intelligence and pattern recognition and computer vision research,there are lots of applications in the fields of industry,transportation,medical,security,educationand so on.The image recognition major includes image acquisition, preprocess, feature extractionand analysis, classification decisions and recognition. Image recognition has developed manyalgorithms, the most common method still is based on statistical characteristic. Kernel FisherDiscrimination Analysis are widely used in image recognition for its the optimal non-linearhandling and good classification performance, the kernel function in KFDA is one of the keyfactors.The research of kernel function mainly includes the structure of kernel function and itsparameter selection, the latter is frequently complicated. Many scholars have made plentifulresearches in SVM,but still exists some faults or restrictions in effect. So we considerresolving the question form the theory, whether or not exist kernel function withoutcomplicated parameters selection, some specialty of Fisher kernel are just correspondent, thispaper analyze and do new research from this point. The major job is as follows:The first step, we analyze some traditional kernel parameter selection methods by test,summarize some advantages and disadvantages, and lead to the ability of resolve complicatedkernel parameter selection from the theory.The second step, to resolve the iteration time-consumed of EM algorithm between Fisherkernel construction, we propose Fisher kernel of face image which is small sample based onclasses information, it don’t need complicated parameter selection, and we obtain the result bymaximum similarity evaluate; After that we analyze the character of kernel in deep, and judgeit whether or not own nice classification performance from the angle of kernel correction, andanalyze which face image more influence the performance of the kernel; then we put thekernel into the KFDA,verify the proposed method by the test of face images of two classes;finally we combine the other’s idea and mixture kernels, and propose the method aboutmixture kernels based on Fisher kernel in KFDA,this method reduce the training time incontrast to traditional mixture kernels.The last step, aim at the practical demand of traffic environment in city, we propose anmethod about special vehicle recognition and tire segmentation in city, it is one type methodof Fisher kernel of small sample image based on classes information, and then include someimage processing to the distinguished images. This paper explains the realization principleand the development process in detail. We use the DLL component and realize it in VS.