Improving scene image classification with multi-class SVMs

We aim to get higher accuracy of scene image classification than attainable with existing methods. We propose using multi-class SVMs (Support Vector Machines) to get this desired higher accuracy. In the full paper, we explain in much detail how to structure multi-SVMs. Here we give only a briefing. Our multi-class SVMs consist of a number of 1-v-1 classifiers and use low-level features such as representative colors and Gabor textures; we make use of relevant information in the two papers by J. Platt^[3], J.H. Friedman^[5] respectively to structure our multi-class SVMs. In our experiments, we used 448 scene images from http://www.project.-minerva. ex. ac. uk. In this case, multi-class SVMs became 7-class SVMs. These experiments show preliminarily: (1) that the accuracy of scene image classification can be raised from 50%-70% attainable with neural network method, which gives the best accuracy among existing methods, to 60%-80% attainable with our 7-class SVMs; (2) that both different kernel functions and different parameters in a particular kernel function give quite different results of classification.