Research Of Image Classification And Annotation Based On SVM

With the rapid development of multimedia technology, image data is of explosion growth. How to obtain the target image quickly and manage so many images efficiently has become the research hotspot in the field of science and industrial production. As the key technology of processing huge amounts of image data, image classification and annotation can solve the disorder problem to a great extent, which has very profound value of research and application. Image annotation can be seen as a special classification, so the paper mainly studies image annotation based on classification method. Support Vector Machine (SVM) has a lot of unique advantages in tackling small sample, nonlinear and high dimensional Machine learning problem and is widely used in the field of image classification. Therefore, the paper focuses on image annotation based on SVM classifier, and conduct the studies around the single label and multi label problem. The research contents are as follows.On the aspect of images’single-label annotation based on SVM classifier, the paper proposed a new image classification algorithm based on Generalized Histogram Intersection Kernel and multi-features. Firstly, to the problem that single feature describes images partially and image classification accuracy is not high, we put forward the method of features combination. Secondly, having researched on the kernel of SVM classifier, we proposed a new kernel, which is named Generalized Histogram Intersection Kernel (GHIK). Finally, we conducted a comparative evaluation with other methods on a benchmark image dataset. The experimental results showed the proposed algorithm to be more accurate than other approaches.On the aspect of images’multi-label annotation based on SVM classifier, we proposed a new method to construct instances. In the paper, we first divided the image into blocks, then calculated the partitions’weights of human eye perception and use the value to distinguish the importance of the human visual perception. Last, we applied the MIML-SVM (Multi-Instance Multi-label SVM) algorithm to experiment on natural scene images. The experimental results showed that evaluation indexes were improved and proved the effectiveness of the proposed algorithm.