Implementation of the SVM neural network generalization function for image processing

Based on the statistical learning theory, Support Vector Machines is a novel neural network method for solving image classification problems. It has proven to obtain the optimal decision hyperplane and is also unaware of the dimensionality of the problem. The decision function is constructed with the support vectors obtained during the learning process. Each pixel bloc in the training database is processed as an input vector, the learning process finds out between input vectors those who will construct the solution (the support vectors), the weights and the threshold of the neural network. SVM does not need a test database and the solution depends entirely on the training database. The aim of our work is to exploit the regularities of the SVM decision function in an integrated vision system. The application of our vision system is object detection and localization. We use SVM classifier as the main module of the system. In order to reduce the classification computation time we are proposing a parallel implementation on an FPGA programmed with VHDL.