基于多尺度图谱和局部谱的目标提取算法

针对多尺度图谱算法不能有效提取含有较多纹理或包含差异较大区域的目标物体, 提出了一种结合图像平滑、多尺度图谱和局部谱的目标提取方法。首先对图像进行l0梯度最小化平滑处理, 锐化边缘的同时消除图像的部分纹理信息; 其次通过多尺度图谱方法对图像进行分割, 该算法结合了归一化割算法的高精确度和多尺度算法的高效率性; 最后结合局部谱理论, 将人工选取的种子区域作为约束条件, 进行有偏割向量估计, 通过最大类间方差法将该向量分割成目标和背景。实验表明, 该方法弥补了多尺度图谱算法的不足, 并能产生很好的目标提取效果。

Object extraction algorithm based on multi-scale spectral and local spectral

Aiming at the drawback of multi-scale spectral which could not extract object region that contained many difference areas or challenging textures. This paper proposed a novel scheme of object extraction. This new scheme was based on image smoothing, multi-scale spectral and local spectral. Firstly, it smoothed the image, sharpening major edges while eliminating part of textures by gradient minimization. Secondly, it used multi-scale spectral image segmentation algorithm to partition image, which combined high accuracy of normalized-cut algorithm and high efficiency of multi-scale algorithm. Finally, combining with local spectral theory, it selected seed region as an additional constraint, then characterized the optimal solution to new problem and showed that it could be interpreted as a locally-biased vector. This locally-biased vector was divided into object and background by maximum variance between clusters. Experimental results show that the proposed approach makes up for the shortage of multi-scale spectral and obtains more satisfactory results.