摘要
稀疏表示和字典学习在图像去噪、图像重建和模式识别等应用上取得了良好的效果,其利用稀疏系数和重构误差来作为模式分类的判别准则。稀疏表示纹理分割方法是将图像分割问题转换为像素点的分类问题。但通常稀疏表示分类方法是基于图像块特征,难以准确表征图像纹理信息。为了解决上述问题,提出基于Gabor特征的稀疏表示纹理分割方法。因为Gabor特征对图像纹理信息的鲁棒性,算法首先从每类纹理中选择一些像素点作为训练样本,计算其不同尺度和方向下的Gabor特征,将其作为初始化字典,通过判别性的字典学习算法(D-KSVD)更新字典,该字典学习算法在KSVD基础上使得字典更具有类别判别能力,最后以待分割图像的每个像素点作为测试样本,计算其Gabor特征。利用OMP算法得到测试样本在字典下的稀疏系数,根据稀疏系数得到类标签,进而对像素点进行分类,完成分割。通过在Brodatz纹理库上的实验结果表明,该方法有效提高了稀疏表示算法对纹理图像分割的正确率。
The method of sparse representation for texture segmentation is to convert the image segmentation into the pixel classification. Generally,the method of sparse representation classification is based on image block feature,which is difficult to accurately character the image’s texture information. To solve the above-mentioned problems,Gabor feature based sparse repre-sentation for texture segmentation is proposed in this paper,because Gabor feature is robustness to image texture. Firstly,some pixels are randomly select from each texture as training samples to calculate their Gabor features with different scales and orien-tations,and take these Gabor features as initialization dictionary. The dictionary is updated by discriminative dictionary learning （D-KSVD）algorithm. Based on KSVD,the algorithm makes the dictionary more discriminative. Finally,each pixel of the under segment image is taken as the test samples to calculate their Gabor features. The OMP algorithm is utilized to calculate the sparse coefficients to obtain the final class labels. The result of experiment on the Brodatz texture database shows that the pro-posed method can effectively improve the texture segmentation accuracy of sparse representation algorithm.
出处
《现代电子技术》
北大核心
2015年第10期73-77,共5页
Modern Electronics Technique
基金
国家自然科学基金(61273251)
