摘要
针对图割法计算代价高并难以应用于纹理分割的问题,提出了一种基于滤波器阵列和小波域图割的纹理分割算法.首先对图像进行多层小波分解;然后在子带图像中使用构建的滤波器阵列提取图像的纹理特征,采用texton直方图作为纹理的统计模型,并采用直方图差计算像素点间的纹理相似度;最后根据子带图像计算虚拟尺度图的权值矩阵,构建关联范围递增的多尺度图结构,并根据规范割准则计算纹理的分割.分割结果表明:该算法在获得稳定和准确的纹理分割的同时能够将原始规范割指数时间复杂度压缩为线性时间复杂度,并能够计算大尺寸的图像分割.
Aiming at high computation cost and poor result of texture segmentation, a texture segmentation approach based on filter bank and graph cut in the wavelet domain was presented, in which images were decomposed through wavelet transform. Then, the texture feature was derived from filter bank defined on the subhands of the wavelet decomposition in every level. The texture was modeled with texton histograms to statistic feature vector, and the pairwise texture similarity was computed by comparing windowed texton histograms. The weights of the virtual level graph were calculated by the subband images. Multi-scale graph with increasingly connection radius was presented to segment the texture image with normalized cut. Experimental results show that the algorithm can compress the exponential-time complexity of normalized cut to linear-time complexity and segment large image, and the texture segmentation results of the method are stability and accuracy.
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2011年第5期105-108,117,共5页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(60775001)
关键词
纹理分割
小波分解
多尺度
图
规范割
texture segmentation
wavelet decomposition
multi-scale
graph
normalized cut
