基于非下采样Shearlet变换的磁瓦表面缺陷检测

针对磁瓦表面缺陷对比度低、图像受不均匀背景和磨削纹理干扰大等问题，提出了一种基于非下采样Shearlet变换（non-subsampled shearlet transform，NSST）的磁瓦表面缺陷检测方法。首先，对磁瓦图像进行多尺度多方向NSST分解，得到一个低频子带图像和多个频率和方向变化的高频子带图像。然后，根据缺陷在高低频域表现出的不同特征进行针对性的处理，在低频子带中分别计算行均值线图像和列均值线图像，将列均值线图像沿行均值线图像扩展，构造基于均值的自适应阈值面对低频子带进行滤波，以去除不均匀背景；同时，利用同一分解尺度下各高频子带系数中微弱缺陷信号的方差较大，显著缺陷信号的能量较大，而噪声和背景干扰信号的方差和能量均较小的差异，构造基于Shearlet高频分解系数方差和能量的综合高频缺陷识别算子，滤除高频子带中的噪声和背景干扰。最后，对修正后的分解系数进行逆NSST重构，得到背景均匀，磨削纹理和噪声干扰得到充分抑制的高对比度图像，采用自适应阈值分割方法，提取出缺陷区域。实验结果表明，该方法的假阳性率、假阴性率和检测准确率分别达到8.8%，5.0%和93.1%，算法在MATLAB仿真平台中平均运行时间为0.629 s；相较于现有的磁瓦表面缺陷检测算法，该方法能够有效地去除不均匀背景、磨削纹理和噪声干扰，检测结果更加准确，鲁棒性更强。

Defect Detection in Magnetic Tile Images Based on Non-subsampled Shearlet Transform

A novel surface defect detection method based on non-subsampled shearlet transform (NSST) is proposed for detecting defects with random noise, low contrast, uneven background and regularly distributed grinding texture in magnetic tile surface images. The method first employed NSST to decompose the original magnetic tile defect image into one low frequency subband and a series of high frequency subbands with multiple frequency domains and shearing directions. Then, the defect signals were processed pertinently and separately according to the varied characteristics of defects in the low and high frequency domains. In the low frequency subband, the mean images of row and column were computed, and an adaptive threshold surface constructed by extending the column mean image along the row mean image was adopted to remove the uneven background. Meanwhile, as weak defect signals and significant defect signals respectively exhibited large variance and energy at the same decomposition scale in high frequency domains and both of them were small in noise and background interference pixels, a discriminator based on the variance and energy of high frequency Shearlet coefficients was proposed for the removal of noise and background interference in the high frequency subbands. Finally, a denoising image with high contrast and even background could be reconstructed by applying inverse NSST to the modified Shearlet coefficients, and the defects could be accurately extracted using an adaptive threshold segmentation method. Experimental results show that the false positive rate, the false negative rate and the accuracy rate of the proposed method are 8.8%, 5.0% and 93.1% respectively, and the average elapsed time is 0.629 s while the proposed method is performed with MATLAB. Moreover, as the proposed method is effective in eliminating the interferences of uneven background, grinding texture and random noise, it has obvious superiorities over the existing ones in terms of both accuracy and robustness.