为了解决传统金属表面质量检测技术的缺陷检测精度不高、缺陷检测识别率不高、缺陷分类不准确的难题，搭建了一套基于机器视觉的金属板材表面检测系统。基于偏微分方程，利用图像等照度线改进中值滤波算法，对图像进行预处理，显著地抑制了图像的噪声。利用最大类间方差算法（OTSU）自适应确定一图像双阈值，改进了Canny算法中高斯滤波器对图像的灰度分布特征提取，使其不受亮度和对比度的影响。最后，利用SIFT （Scale-invariant Feature Transform）算法提取缺陷特征点，提出一种BP （Back Propagation）神经网络和SVM（Support Vector Machine）向量机结合分类器的检测方法，缺陷检出率为92.68%，单幅图像检测仅需49.8 ms，该缺陷检测系统对金属板材表面缺陷能有效提取与识别，满足金属板材表面在线检测的要求。

In order to solve the problems of traditional metal surface quality inspection technology, such as the low precision of defect detection, the low recognition rate of defect detection, and the inaccurate classification of defects, a set of metal sheet surface inspection system based on machine vision was established. Based on the partial differential equation, the median filtering algorithm is improved by using the iso-illumination lines of the image, and the image is pre-processed to significantly suppress the image noise. Using the Maximum Inter-Class Variance Algorithm (OTSU) to adaptively determine the double threshold value of the image, the Gaussian filter in Canny algorithm was improved to extract the gray distribution features of the image, so that it was not affected by brightness and contrast. Finally, using SIFT (Scale-invariant feature transform) algorithm to extract defect feature points, a classifier detection method combining BP (Back Propagation) neural network and SVM (Support Vector Machine) is proposed. The defect detection rate is 92. 68%, and the single image detection only takes 49. 8 ms. The defect detection system effectively extracts and identifies the surface defects of the metal sheet, and meets the requirements of online detection of the surface of the metal sheet.