针对人工分割木材表面缺陷的烦琐性和阈值分割算法对缺陷像素信息衡量的不稳定性,提出了一种基于邻接自适应谱聚类的木材表面缺陷分割算法.算法以简单线性迭代超像素(simple linear iterative cluster,SLIC)为基础,对缺陷图像进行预处理,融合木材缺陷的纹理特性和超像素块间的距离尺度,并采用邻接自适应谱聚类进行分割;缺陷分割初步完成后,通过变异系数衡量缺陷块中像素信息的离散程度进行再次分割,克服初次分割结果的过分割问题;考虑木材表面缺陷形态学上的封闭性,将 2 次分割图像进行合并,继而用邻接扫描法对次分割图形进行填充,最终对木材表面缺陷进行分割界定.考虑木材表面缺陷种类的多样性,选取了虫眼、死节、活节等缺陷图像进行分割对比试验,相较于OTSU阈值分割算法,本研究算法在单个和多个木材表面缺陷分割方面,类别平均像素准确度(mean pixel accuaracy,MPA)分别提升 4.69%,14.23%,平均交并比(mean intersection over union,mIoU)分别提升 33.27%,33.43%.本研究算法能够更加准确地将木材表面缺陷从复杂背景中分割出来,缺陷边缘轮廓的构建更接近于理想分割情况,且运行时间较短,对木材表面缺陷的分割具备较强的精确性与可行性.
Wood surface defect segmentation algorithm based on adjacency adaptive spectral clustering
The precise detection of wood surface defects can more effectively enhance the value of wood to meet the actu-al production needs,but the existing methods of detecting defects cannot meet the production requirements.Aiming at solving the tedious issues of manually segmenting wood surface defects and the instability of the threshold segmentation algorithm in measuring the defective pixel information,an algorithm for segmenting wood surface defects based on adja-cency adaptive spectral clustering was proposed.The algorithm was based on simple linear iterative cluster(SLIC),which pre-processes the defect images,divided them into sets of super-pixel blocks with similar internal information,then fused them with the texture characteristics of wood defects and the distance scale between super-pixel blocks and used adjacency adaptive spectral clustering for segmentation.After the initial segmentation was completed,the discrete degree of pixel information in the defective block was segmented again by measuring the coefficient of variation to over-come the over-segmentation problem of the initial segmentation result.In consideration of the morphological closure and completeness of the wood surface defects,the two segmented images were merged.Following that,the sub-segmented graphs were filled with the neighbor scanning method,and finally the wood surface defects were segmented and defined.Watershed segmentation algorithm,iterative thresholding algorithm,OTSU thresholding segmentation algorithm,and NJW spectral clustering were used as comparison algorithms for the proposed algorithm,and considering the diversity of defect types on the surface of wood,defect images such as insect eyes,dead knots and live knots were selected for seg-mentation comparison experiments.Compared with the watershed algorithm,iterative thresholding algorithm,NJW spectral clustering algorithm,the proposed algorithm in the segmentation of defects on the surface of a single piece of wood category mean pixel accuracy(MPA)were improved by 2.9%,17.18%and 1.2%,and the mean Intersection over Union(mIoU)were improved by 32.13%,39.71%and 13.32%,respectively.Compared with the OTSU threshold seg-mentation algorithm,the proposed algorithm increased the MPA by 4.69%and 14.23%for single and multiple wood sur-face defects segmentation,respectively.The mIoU was enhanced by 33.27%and 33.43%,correspondingly.The algo-rithm in this study can more accurately segment wood surface defects from the complex background,can segment images containing single or multiple defects,and adapted to the images of wood surface defects with different scales and color depths.The construction of defect edge contours is closer to the ideal segmentation situation,and the running time is shorter,which provides strong accuracy and feasibility for segmenting wood surface defects.
wood surface defectimage segmentationadjacency adaptive spectral clusteringsuper-pixelcoefficient of variation
万方数据
维普
