基于LS-SVM的岩石细观图像分析方法探讨

DISCUSSION ON ANALYTICAL METHOD FOR LS-SVM BASED MESOSCOPIC ROCK IMAGES

为了解决岩石细观力学试验中图像处理过程复杂、质量不高及操作效率低等问题,将LS-SVM的分类方法与数字图像处理的阈值分割法相结合,提出了人机结合的岩石细观结构图像系统分析方法。该方法将图像分割问题转化为分类问题,通过对训练样本的学习,生成可将试验图像分类的LS-SVM分类机,从而提取岩石细观力学试验中得到的感兴趣区域的特征图像以及量化细观结构。对花岗岩图像进行处理,处理后的结果表明,该方法可以获得高质量的岩石细观图像处理结果,处理准确率达到96.82%。采用三步搜索法选取参数,能在保证图像处理质量的前提下提高参数选取速度;对训练样本进行稀疏化处理,可以提高分类效率,缩短分类时间;为了减小人为因素的影响,训练图像的选取应具有代表性,且在生成训练目标前需进行图像后处理。

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.