基于DCGAN和改进YOLOv5s的桥梁表面缺陷检测识别

针对人工检测桥梁表面缺陷存在精度低、速度慢和漏检率高等问题,该文提出了基于深度卷积生成式对抗网络(deep convolutional generative adversarial networks,DCGAN)和改进YOLOv5s的桥梁表面缺陷检测识别模型.首先,通过DCGAN网络对自主采集的桥梁表面缺陷图像进行数据增强,建立每类缺陷样本数量较为均衡的数据集; 其次,在YOLOv5s模型基础上嵌入CBAM注意力机制模块,使模型将注意力集中于缺陷区域,从而提升图像分类的准确率; 最后,为验证所提方法的适用性,将包含4类桥梁表面缺陷的数据集进行训练与测试.实验结果表明:该文提出的模型在桥梁表面缺陷检测上的准确率为92%,相比其他深度学习模型具有更高的检测精度和鲁棒性.

The Bridge Apparent Defects Detection Based on DCGAN and Improved YOLOv5s

In order to solve the problems of low accuracy,slow speed and high missed detection rate in current bridge surface defect detection methods,the bridge surface defect detection method based on deep convolutional generative adversarial networks(DCGAN)and improved YOLOv5s is proposed.recognition methods.Firstly,through the deep convolutional generative confrontation network,the self-collected bridge surface defect images are enhanced to establish a more balanced data set with rich sample features.Secondly,the CBAM attention mechanism module is embedded on the basis of the YOLOv5s model to learn the importance of each channel and spatial feature independently,thereby improving the performance of image classification.Finally,in order to verify the applicability of the proposed method,four data sets of bridge surface defects will be included for training and testing.Experimental results show that the accuracy of the proposed model in the detection of bridge surface defects is 92%,which has higher detection accuracy and robustness than other deep learning models.