【摘要】 为了解决尺度变化和训练样本有限给纹理识别带来的困难,提出了一种基于支持向量机和遗传算法的纹理识别新方法。该方法用小波变换各子带图像共生矩阵参数、分析窗口大小、像素均值和像素标准差等参数作为纹理特征,用多类支持向量机作为分类器。用遗传算法对纹理特征集进行了优化;用输出纠错码将二类支持向量机扩展到多类,提高了分类器的泛化能力。用包含有25类单色自然纹理的图像库进行识别试验,结果表明,该方法的识别错误率小于10%,得到了比传统的贝叶斯等方法更高的识别率和更好的推广性。更多还原

【Abstract】 <Abstrcat>A new method, based on support vector machines (SVMs) and genetic algorithm, is proposed for texture recognition. The main difficulty of texture recognition is the lack of effective tools to characterize different scales of textures. To alleviate the problem,wavelet co-occurrence parameters, window size, mean and standard deviation of different level discrete wavelet transform (DWT) images are used as texture features. In particular, we apply multi-class support vector machines for classification, using error-correcting output codes (ECOC) for the extension of binary SVMs. On the other hand, genetic algorithm is used for feature optimization to obtain more accurate rates. Compared to the conventional Bayes classifier, SVMs produce more accurate recognition rates on the Ponce texture database. The error rate of experiments is less than 10%.更多还原