基于RPCA和视觉显著性的风机叶片表面缺陷检测

摘 要：针对风机叶片表面缺陷检测问题，提出了一种基于鲁棒主成分分析(RPCA)和视觉 显著性的表面缺陷检测方法。在 RPCA 的基础上，通过增加噪声项和考虑像素的空间关系，以 利于缺陷的分割，即通过 F 范数正则项抑制高斯噪声和光照不均，利用 Laplacian 正则项约束像 素的空间关系，以保持显著图中具有相似显著值且空间相邻超像素的局部一致性和不变性。首 先，对输入的风机叶片表面图像进行超像素分割和特征提取，得到图像的特征矩阵；然后，利 用改进的 RPCA 法得到稀疏矩阵，根据稀疏矩阵和视觉显著性方法计算出缺陷区域的显著图； 最后，优化显著图并采用自适应阈值分割实现缺陷的检测。通过实验仿真和对实验结果定性定 量分析，表明该方法具有较高的准确率。

Surface Defect Detection of Wind Turbine Blades Based on RPCA and Visual Saliency

Abstract: Aiming at the problem of surface defect detection of wind turbine blades, a method based on robust principal component analysis (RPCA) and visual saliency is proposed. Based on RPCA, the method adds the noise term and Laplacian regularization term to facilitate the segmentation of defect images, that is, suppressing Gaussian noise and uneven illumination by F-norm regularization term, and constraining the spatial relationship of pixels with Laplacian regularization term which can preserve invariance and the local consistency among the spatially adjacent sub-regions with similar saliency values in a saliency map. Firstly, superpixel segmentation and feature extraction are performed on the surface image of input wind turbine blades to obtain the feature matrix of the image. Then, the sparse matrix is obtained by the improved PRCA, and the saliency map of the defect region is calculated according to the sparse matrix and visual saliency method. Finally, the saliency map is optimized and the adaptive threshold segmentation algorithm is used to detect defects. Through experimental simulation, the experimental results are qualitatively and quantitatively analyzed, which indicates that the proposed method has high detection accuracy.