Fast training of SVM for color-based image segmentation

A novel method based on support vector machine (SVM) for color image segmentation is presented. Considering image segmentation is a two-class problem, a two-class SVM to classify pixels in color space is proposed. In order to speed up training and optimize parameters of SVM, the color quantization and sample selection approaches are presented to reduce the size of training set and to make the reduced set separable. Training with reduced and separable dataset, minimizing the number of support vectors is regard as an estimation criterion of generalization performance to kernel parameter optimization. The classifier can be trained on-line and implemented in real-time. The new algorithm has been used to color-based image segmentation, and it brings robust performance in practice.