基于SLIC0融合纹理信息的超像素分割方法

SLIC0-based superpixel segmentation method with texture fusion

由于SLIC0算法在分割时仅考虑图像的颜色、亮度、空间位置特征,没有考虑纹理特征,当分割具有繁杂纹理的自然图像时,其分割的超像素无法精准地符合区域或目标的边界或外轮廓,因此提出基于SLIC0融合纹理信息的超像素分割算法——SLIC0-t。首先利用光谱分析描述图像中区域的纹理特性,然后在分割中融合能够准确反映图像中目标轮廓或区域边界的纹理特征;其次在分割过程中,进一步优化SLIC0围绕种子像素搜索近邻像素的搜索策略,采用以各个种子点为中心,在以预期超像素邻接距离为半径的圆盘内搜索的搜索策略;最后通过在公共图像库BSDS500上进行连续不同大小超像素的分割实验验证,结果表明:在边界召回率方面,SLIC0-t算法明显稳定优越于SLIC0算法;在欠分割错误率方面,其与SLIC0算法基本相当,处于可接受范围内。

SLIC0 （the simple linear iterative clustering （SLIC） version with 0 parameter） algorithm only considers color,brightness and space features,but no texture feature; when the natural image with clutter texture scenario is segmented,the superpixels segmented with SLIC0 do not accurately coincident with the boundaries of the areas or the contours of the objects.In order to solve this problem,this paper presents the SLIC0-t algorithm based on the SLIC0 with texture feature fusion.First,the area texture in the image is described with spectrum analysis,and then in the segmentation the texture features that are able to accurately reflect the contour of the objects or the area boundaries are fused.And in the process of segmentation,the search strategy of SLIC0 searching for the neighbor pixels around the seed pixel is further optimized.The search strategy is adopted,which performs the searching in the scope of a disk that takes the seed point as the center and the expected adjacency superpixel distance as the radius.Finally,the segmentation experiment verification for continuous superpixels with different sizes on the BSDS500 public image data sets was conducted.The results show that the proposed SLIC0-t algorithm is significantly superior to SLIC0 algorithm in the boundary recall rate,and has almost the same undersegmentation error rate as SLIC0 algorithm,which is in the acceptable range.