Research On Image Classification Algorithm Based On Generalized Gaussian Combination Kernel

With the development of the Internet and multi-media technology, computer science is developing rapidly. Digital Image has become a part of everyone's life. With such background, the automatic recognition and understanding of digital has become a hot topic in computer science field. The purpose of image understanding is to recognize the meanings and contents of image through modeling and computation, and finally to understand the image like human does. The image understanding has a very wide range of applications in the fields of medical treatment, military affairs, and security. With the widening and deepening of research, image understanding is drawing more and more attention.This paper summarizes the research results of the image understanding field in China and abroad, especially the recent development in the field of multiple kernel learning in machine learning. This paper proposed the generalized Gaussian combination kernel for the first time. This method expands the parameters in the normal classifier. It can find out the hidden information of the semantic.After proposing the generalized Gaussian combination kernel, this paper uses the spatial pyramid and bag of visual words to organize the features. We describe this method in detail, especially how it can combine with this paper's kernel.In the experiment part, we put the proposed kernel in the completion algorithm to test it, and put it into the specific image classification problem. We use international Caltech-101 image datasets as the test data. We test computation speed, vocabulary size, kernel compare, and compare it with other state of art algorithms to verify the effectiveness of our algorithm.On the other part, we use the proposed generalized Gaussian combination kernel to erotic video classification and feature combination. In this paper's algorithm, we use two methods to detect erotic video. In the first method, we use three MPEG-7 color descriptors to extract features from frame series. Then we use the proposed kernel to combine the features and classify them. In the second method, we use skin detection and frame difference method to find out the active skin percentage in the video. Then we compare it with the threshold to get the result. Finally, we combine the result from these two methods together, to get the final result.