Abstract: Defect detection on the surface of sheet metal is the key to judging the rigidity and wear resistance of plate products.Aiming at the problems of poor effect,low accuracy and slow speed of traditional manual detection or general target detection,a defect detection method integrating an attention mechanism was proposed.A dual spatial and channel attention mechanism was introduced in the backbone network,which helped the network find the region of interest from deep features and improves the feature extraction capability of the model;then,deformable processing was added to the perceptual field of the detection head to cope with the deformation occurring in the actual image.The improved detection method was validated on the metal sheet surface defect detection datasets NEU-DET and GC10-DET.The results showed that the full category average accuracy was improved by 7% and 4.2% respectively,and the model could process the number of images up to 72 frames per second.In addition,the method also outperformed existing general-purpose detectors in terms of detection speed and full detection rate,which could better meet the needs of industrial scenarios.

Key words: defect detection, sheet metal, attentional mechanisms, deformable convolution

摘要： 金属板材表面的缺陷检测是判断板材产品刚性和耐磨性的关键。针对传统的人工检测或通用目标检测效果差、精度低、速度慢等问题,提出一种融合注意力机制的缺陷检测方法。首先,在主干网络中引入了空间和通道双路注意力机制,有助于网络从深层特征中找到感兴趣的区域,提高模型的特征提取能力;然后,对探测头的感受野增加可变形处理,以应对实际图像发生的形变。将改进后的检测方法在金属板材表面缺陷检测数据集NEU-DET和GC10-DET上进行验证。结果表明,全类别平均精度分别提升7%和4.2%,该模型每秒内处理图片数量的能力最高达72帧/s。另外,该方法在检测速度和查全率等指标上同样优于现有的通用检测器,能够较好地满足工业场景的需求。

关键词: 缺陷检测, 金属板材, 注意力机制, 可变形卷积