Support vector machines trained by linear programming: theory and application in image compression and data classification

This paper formulates the learning of support vector machines (SVM) as a linear programming problem. An SVM has the property that it chooses the minimum number of data points to use as the centres for the Gaussian kernel functions in order to approximate the training data within a given error. A linear programming (LP) based method is proposed for solving both regression and classification problem. Examples of function approximation and class separation illustrate the efficiency of the proposed method. In addition, the paper explores the possibility of using SVM with radial basis function kernels to compress an image. Our results show that image compression of around 20:1 is achievable while maintaining good image quality.