中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊
重庆交通大学学报（自然科学版）

重庆交通大学学报（自然科学版） ›› 2026, Vol. 45 ›› Issue (6): 27-33.DOI: 10.3969/j.issn.1674-0696.2026.06.04

• 智慧交通基础设施 • 上一篇    

面向深度学习模型训练的RM-DCGAN路面病害图像增广方法

张联玖1，杨海建1，王冠2,3，祁翔4，赵宁雨2   

  1. (1. 中国港湾工程有限责任公司，北京 100027； 2. 重庆交通大学 土木工程学院，重庆 400074； 3. 重庆交大建设工程质量检测中心有限公司，重庆 400074； 4. 中交第一公路勘察设计研究院有限公司，陕西 西安710065)
  • 收稿日期:2025-11-12 修回日期:2026-04-20 发布日期:2026-07-10
  • 作者简介:张联玖(1980—)，男，四川宜宾人，高级工程师，主要从事道路施工与检测等方面的工作。E-mail：ljzhang@chec.bj.cn 通信作者：王冠（1993—），男，重庆人，高级工程师，博士研究生，主要从事道路病害方面的研究。E-mail：611250780015@mails.cqjtu.edu.cn
  • 基金资助:
    国家自然科学基金项目（5207083295）；重庆市自然科学基金项目（cstc2021jcyj-msxmX1011）;中国港湾工程有限责任公司重大研发项目（2024-重大-01）

RM-DCGAN Road Surface Disease Image Augmentation Method for Deep Learning Model Training

ZHANG Lianjiu1, YANG Haijian1, WANG Guan2,3, QI Xiang4, ZHAO Ningyu2   

  1. (1. State Key Laboratory of Safety and Resilience of Civil Engineering in Mountain Area, East China Jiaotong University, Nanchang 330013, Jiangxi, China;2. Jiangxi Provincial Key Laboratory of Comprehensive Stereoscopic Traffic Information Perception and Fusion, East China Jiaotong University, Nanchang 330013, Jiangxi, China)
  • Received:2025-11-12 Revised:2026-04-20 Published:2026-07-10

PDF

摘要： 针对因路面破损严重且训练样本匮乏导致深度学习模型特征学习与泛化性能不足，导致难以满足工程实际需求的问题，以尼日利亚路面病害（样本稀缺、破损程度大）为研究背景，提出一种改进型深度卷积生成对抗网络RM-DCGAN（residual and multi-attention DCGAN），用于生成高质量、多样化的路面病害图像以扩充训练数据集。该模型基于DCGAN的生成器与判别器基础结构，引入残差模块以缓解深层网络梯度消失问题，融合自注意力机制与通道注意力机制构建混合注意力模块以强化病害特征捕捉，采用含梯度惩罚项的Wasserstein距离作为损失函数，有效解决了传统DCGAN损失函数不稳定、生成图像质量欠佳的问题。基于尼日利亚道路自建数据集的实验验证表明：RM-DCGAN生成图像的IS值与FID值分别为5.44和183.68，较DCGAN分别提升110.9%、降低13.1%；利用该模型增广的数据集训练YOLOv5检测模型，对裂缝、坑槽等典型路面病害的平均检测精度（mAP）达92.31%，较传统数据增广方法提升了6.68%；RM-DCGAN数据增广方法可有效解决小样本路面病害检测中的训练数据不足问题，为小样本场景下路面病害检测提供了可靠的技术支撑。

关键词: 道路工程；数据增广；深度卷积生成对抗网络；路面病害检测；残差结构；注意力机制

Abstract: In response to the problem of insufficient feature learning and generalization performance of deep learning models as well as difficulty in meeting the practical needs of engineering, caused by severe road surface damage and a lack of training samples, the improved residual and multi-attention deep convolutional generative adversarial network (RM-DCGAN) was proposed in the research context of road surface diseases in Nigeria (with scarce samples and high degree of damage), to generate high-quality and diverse road surface disease images to expand the training dataset. The proposed model was based on the generator and discriminator foundation structure of DCGAN, introducing residual modules to alleviate the problem of gradient vanishing in deep networks, and integrating self-attention mechanism and channel attention mechanism to construct a hybrid attention module to enhance disease feature capture. Meanwhile, the Wasserstein distance with gradient penalty term served as the loss function was employed to effectively solve the problems of unstable loss function and poor generated image quality in traditional DCGAN. Experimental verification based on the self-constructed Nigerian pavement dataset demonstrates that the IS and FID values of the images generated by RM-DCGAN are 5.44 and 183.68, respectively, which are respectively 110.9% higher and 13.1% lower than those of DCGAN. The augmented dataset of the proposed model was used to train the YOLOv5 detection model, and a mean average precision (mAP) for typical road surface diseases such as cracks and potholes reaches 92.31%, which is 6.68% higher than traditional data augmentation methods. Research shows that the RM-DCGAN data augmentation method can effectively solve the problem of insufficient training data in small sample road surface disease detection, providing reliable technical support for road surface disease detection in small sample scenarios.

Key words: highway engineering; data augmentation; deep convolutional generative adversarial network; road surface disease detection; residual structure; attention mechanism

中图分类号: