军事知识服务平台
RAFLS-DCGAN:一种基于小样本桥梁裂缝病害图像的数据扩充方法
RAFLS-DCGAN:A Data Augmentation Method for Small-Sample Bridge Crack Defect Images
摘要 本文提出一种基于改进深度卷积生成对抗网络(DCGAN)的桥梁裂缝图像扩充方法——基于残差网络(Residual network)、注意力机制(Attention mechanism)、全连接层(Fully-connected layer)、最小二乘损失(Least Squares loss)的深度卷积生成对抗网络(RAFLS-DCGAN).首先,将原始 DCGAN 中的二元交叉熵损失替换为最小二乘损失,并引入注意力机制以提升模型对裂缝特征的捕捉能力;其次,在网络中融入全连接层,并设计改进的残差模块应用于生成器与判别器,以增加网络深度并缓解模式崩溃问题,从而提取更深层次的图像特征,提升生成图像的真实性与多样性.实验结果表明,RAFLS-DCGAN 生成桥梁裂缝图像的 FID 值较 DCGAN、DCGAN+、LSGAN 和 SNGAN 分别降低 406.26、189.52、125.36 和 107.47;生成图像分类器评估(GICE)与真实图像分类器评估(RICE)较 DCGAN+分别提升 12.28 和 23.52 个百分点.利用 RAFLS-DCGAN 扩充的数据集进行裂缝识别,在 AlexNet、VGGNet 和 ResNet 上的准确率分别提升 17.01、25.85 和 25.93 个百分点;在YOLOv5、YOLOv8 和 YOLOv11 上的准确率分别提升 18.11、24.68 和 12.26 个百分点.
Abstract:
This paper proposes a bridge crack image augmentation method based on an improved Deep Convolutional Generative Adversarial Network(DCGAN)—namely RAFLS-DCGAN,which integrates residual networks,attention mechanisms,fully-connected layers,and least squares loss.First,the binary cross-entropy loss in the original DCGAN is replaced with least squares loss,and an attention mechanism is introduced to enhance the model's ability to capture crack features.Second,fully-connected layers are incorporated into the network,and an improved residual module is designed and applied to both the generator and discriminator,so as to increase network depth and mitigate mode collapse,thereby extracting deeper image features and improving the authenticity and diversity of generated images.Experimental results show that,for bridge crack image generation,RAFLS-DCGAN reduces the FID values by 406.26,189.52,125.36,and 107.47 compared to DCGAN,DCGAN+,LSGAN,and SNGAN,respectively.Its Generator Image Classification Evaluation(GICE)and Real Image Classification Evaluation(RICE)scores improve by 12.28 and 23.52 points over DCGAN+,respectively.When the dataset augmented by RAFLS-DCGAN is used for crack recognition,the accuracy improves by 17.01,25.85,and 25.93 points on AlexNet,VGGNet,and ResNet,respectively,and by 18.11,24.68,and 12.26 points on YOLOv5,YOLOv8,and YOLOv11,respectively.
年,卷(期): 2026, 65(5)
分类号: TP391.41
机标分类号: P642.2 TP391.41 S126
在线出版日期: 2026年11月4日
基金项目: 2022JQ-340:陕西省自然科学基础研究计划项目2026JC-YBMS-0033:陕西省自然科学基础研究计划项目2024GX-YBXM-278:陕西省重点研发计划24JR072:陕西省教育厅科学研究计划高校工程研究中心项目,