作者:孙凯，翟晓卉，赵吉福，孙艳玲，董贤光，王艳

Authors:SUN Kai,ZHAI Xiao-hui,ZHAO Ji-fu,SUN Yan-ling,DONG Xian-guang,WANG Yan
摘要:摘要:针对智能电表HPLC通信模块外观划痕检测的问题，提出了一种先进的像素级划痕检测体系结构Res-DU-Net，其特点是采用了深度卷积神经网络技术。首先，Res-DU-Net通过卷积、残差结构、空洞卷积、级联等操作实现了像素级的滑痕检测，形成了U形模型结构。其次，采用Adam算法对智能电表通信模块外观图像进行了训练和验证。最后，Res-DU-Net在学习率为0.001，损失率为0.020的情况下，准确率为0.985，召回率为0.987。实验结果证明，Res-DU-Net在像素级划痕检测方面比传统方法和完全卷积网络（FCN）及U-Net更加有优势。
Abstract:Abstract:Aiming at the problem of the appearance scratch detection of the smart meter HPLC communication module, an advanced pixel-level crack detection architecture Res-DU-Net is proposed, which is characterized by the use of deep convolutional neural network technology. First, Res-DU-Net achieves pixel-level slip detection through operations such as convolution, residual structure, cavity convolution, and cascade, forming a U-shaped model structure. Secondly, the Adam algorithm was used to train and verify the appearance image of the smart meter communication module. Finally, Res-DU-Net has an accuracy rate of 0.985 and a recall rate of 0.987 when the learning rate is 0.001 and the loss rate is 0.020. Experimental results prove that Res-DU-Net has more advantages than traditional methods, Fully Convolutional Network (FCN) and U-Net in pixel-level scratch detection.

PDF全文下载地址:

可免费Download/下载PDF全文