基于改进U-Net的透明件划痕检测方法

为了满足透明件表面质量和市场竞争的需求，实现产品表面缺陷的自动化检测至关重要。本文针对透明件表面划痕快速检测问题，提出了一种基于改进U^2-Net的缺陷检测方法。首先，论述直接应用U^2-Net网络进行透明件表面划痕检测的数据集准备、网络搭建、损失函数、评估指标；其次，初始化网络进行训练，分析产生误检漏检及低效的原因；最后，优化损失函数，加入正则化技术，并给出在输入数据前加入Mosaic数据增强，在解码阶段融入深层可分离卷积以及加入Attention机制的改进方案。结果表明：本文提出的改进方案能够有效分割出不同情况下的划痕，准确率达到0.987，漏检率为0.006，并在检测速度上有19%的提升。可见改进U^2-Net的透明件划痕检测方法能够很好满足工业流水线准确检测缺陷的实际需求。

Scratch Detection Method of Transparent Parts Based on Improved U^2-Net

In order to meet the requirements of surface quality and market competition for transparent parts, it is of great importance to realize the automatic detection of surface defects in products. Aiming at the problems in rapid detection of surface scratch on transparent parts, this paper proposed a defect detection method based on improved U^2-net. Firstly, it discussed dataset preparation, network construction, loss function and evaluation index in direct application of U^2-net to perform detection of surface scratch in transparent parts; secondly, it initialized the network for training, and analyzed the causes of false and missing detection as well as for low efficiency; finally, it optimized the loss function, added the regularization technique, added Mosaic data enhancement before inputting data, integrated Depthwise Separable Convolution at the decode stage, and added the Attention mechanism. The results show that the improvement scheme proposed in this paper can effectively segment the scratches in different situations, the accuracy rate reaches 0.987, the missing detection rate is 0.006, and the detection speed increases by 19%. It can be seen that the scratch detection method for transparent parts based on improved U^2-net can fairly satisfy the actual demand of accurate defect detection in industrial production lines.