针对带钢表面缺陷检测实时性要求高，采集的图像易受光照环境影响且缺陷特征弱等因素影响，提出一种基于谱残差视觉注意模型的带钢表面缺陷在线检测算法。首先，提出改进同态滤波方法对图像预处理，去除光照不均匀的影响，改善后续的分割结果。 然后，构建谱残差视觉注意模型，通过对数频谱曲线差分得到缺陷显著图像。最后，提出加权马氏距离方法对显著图像阈值化增强，并利用连通区域标记法，标记出原带钢图像的缺陷位置。对提出的算法进行了实验验证，结果显示: 该算法检测速度快，单幅图像平均检测耗时仅37.6 ms，满足带钢在线实时检测要求。在同一缺陷数据库与灰度投影法，多尺度Gabor边缘检测法和隐马尔可夫树模型法进行了性能对比，结果表明: 本文算法对带钢常见8类缺陷类型，平均检测率达到了95.3%，且漏检率和误检率较低，有效性高于对比算法。

As captured images for surface defect detection of a steel strip is vulnerable to lighting conditions, weaker defect characteristics and other factors, this paper proposes a new algorithm based on spectral residual visual attention mode to complete the strip surface defect detection in real time. Firstly, the homomorphic filtering method was proposed to preprocess the image to remove the influence of uneven illumination and to perfect the subsequent segmentation results. Then, a visual-attention model was constructed to obtain the defect saliency map by applying the subtraction to the logarithmic spectrum curve. Finally, the weighted Mahalanobis distance method was proposed to significantly enhance the saliency image thresholding. These locations of the defects in the original strip defect images were marked by using the connected-component labeling method. The proposed algorithm was verified by experiments. Experimental results show that the algorithm has a fast detection speed, and takes only 37.6 ms in the single image detection, meeting the requirements of the real-time detection. The comparative experiment with the gray projection method, multi-scale Gabor edge detection method and Markortree model was carried out in the same defect database, and the results show that average detection rate of the proposed algorithm reaches to 95.3% for 8 common types of defects. In the meantime,the missing rate and false positive rate are still low. These results validate the effectiveness of the algorithm.