Abstract：To deal with the problem of the complexity and low quality for the existing image processing methods in mesomechanical experiments of rocks，a man-computer method for image processing was put forward on the basis of the least squares support vector machines(LS-SVM) and image segmentation. In the algorithm，the image segmentation is transformed into the classification of LS-SVM. Through the learning of the training samples，the LS-SVM classifier，which can identify the experimental images，was produced. The characteristic images in the interesting regions are obtained in order to quantify the microstructures. The experimental results from analyzing the images of granite show that the proposed method has high averagely accuracy of 96.82% in practical detection. By using the three-step search method to select the parameters of LS-SVM，the scouting speed was greatly improved on the premise of ensuring the quality of image processing. The treatment of sparseness is beneficial for improving the efficiency of classifying，and shortening time of the work. In order to reduce the influences raised from human factors，the selection of training image should be representative；and the image post-processing should be carried out before the generation of the training target.
