多特征非迭代超像素分割算法

Multi-feature non-iterative superpixels segmentation algorithm

针对简单非迭代聚类(SNIC)算法不能较好贴合图像中目标边缘的缺陷,提出了一种多特征非迭代超像素分割(MNSS)算法。在特征提取上,利用高斯卷积在Lab颜色空间上对每个像素点求出其水平和垂直方向的颜色梯度特征,通过对像素点进行腐蚀和膨胀操作,得到像素点形态学轮廓特征,在不丢失梯度特征表示的同时,增强算法边缘命中率。基于SNIC算法非迭代聚类框架,依赖于像素点间的颜色、空间、颜色梯度、形态学轮廓特征的加权距离实现超像素分割。在BSDS500公开数据集上的实验结果表明,在生成相同超像素个数情况下,MNSS算法与主流的5种算法相比,在保证时间复杂度低的同时,有效提升了超像素分割精度。

Simple non-iterative clustering(SNIC)algorithm can't adhere well to object boundaries.To address this drawback,this paper proposes a multi-feature non-iterative superpixels segmentation(MNSS).In feature extraction,Gaussian convolution is used to obtain the horizontal and vertical color gradient features of each pixel in the lab color space;Then the morphological contour features of each pixel are obtained by erosion and dilation operations,which can enhance the edge hit rate of the algorithm without losing the representation of gradient features.And last,based on the non-iterative clustering framework of SNIC algorithm,superpixel segmentation is realized depending on the weighted distance of color,space,color gradient and morphological contour features between pixels.The experimental results on BSDS500 public dataset show that the proposed MNSS algorithm can effectively improve the segmentation accuracy of superpixels while ensuring low time complexity compared with the five mainstream algorithms under the condition of generating the same number of superpixels.