电子测量与仪器学报

武兵,田莹.基于多尺度特征融合的轻量化道路损伤检测算法[J].电子测量与仪器学报,2025,39(11):175-184

基于多尺度特征融合的轻量化道路损伤检测算法

Lightweight road damage detection algorithm basedon multi-scale feature fusion

DOI：

中文关键词: 道路损伤检测 YOLO11 多尺度特征增强通道注意力轻量化

英文关键词:road damage detection YOLO11 multi-scale feature enhancement channel attention light weight

基金项目:国家自然科学基金（62072086）、辽宁省教育厅高校基本科研项目（LJ242510146006）资助

作者	单位
武兵	辽宁科技大学计算机与软件工程学院鞍山114000
田莹	辽宁科技大学计算机与软件工程学院鞍山114000

Author	Institution
Wu Bing	School of Computer and Software Engineering, University of Science and Technology Liaoning, Anshan 114000, China
Tian Ying	School of Computer and Software Engineering, University of Science and Technology Liaoning, Anshan 114000, China

摘要点击次数: 590

全文下载次数: 395

中文摘要:

为了提高现阶段道路损伤检测方法在复杂环境下检测困难、细节纹理丢失严重、效率低等问题，提出了多尺度特征融合的轻量化YOLO算法（MSL-YOLO）。首先，在YOLO11n的基础上进行改进，针对损伤目标特征表达能力弱，设计特征融合通道注意力(feature fusion channel attention，FFCA)模块提高损伤信息的权重，加强特征信息的提取，减少冗余信息；为了在复杂环境下更好地捕捉不同尺寸的损伤目标，设计了一种多尺度特征增强(multi-scale feature enhancement，MSFE)模块提升模型的多尺度特征融合能力，进一步提高检测性能；为实现模型轻量化和检测实时化，在Neck部分引入了轻量级网络(lightweight network，LNet)来减轻模型的计算复杂度，方便模型的部署和应用。实验结果表明，在RDD2022道路裂缝数据集上，所提方法检测平均精度为52.5%，模型参数量为2.3×106，相较于YOLO11n算法平均精度提升了1.8%，参数量下降了11.5%。不仅能满足对道路损伤检测的高精度、高速度、轻量化的要求，且具有较强的鲁棒性和实时性。

英文摘要:

In order to improve the current road damage detection methods in complex environment detection difficulties, serious detail texture loss, low efficiency, a multi-scale feature fusion lightweight YOLO (MSL-YOLO) method is proposed. Firstly, based on the improvement of YOLO11n, the Feature fusion channel attention (FFCA) module is designed to improve the weight of damage information, strengthen the extraction of feature information, and reduce redundant information. In order to better capture damage targets of different sizes in complex environments, a multi-scale feature enhancement (MSFE) module is designed to enhance the multi-scale feature fusion capability of the model and further improve the detection performance. In order to realize the Lightweight model and real-time detection, lightweight network (LNet) is introduced in Neck to reduce the computational complexity of the model and facilitate the deployment and application of the model. The experimental results show that on the RDD2022 road crack dataset, the proposed method has an average detection accuracy of 52.5%, and the number of model parameters is 2.3×106, which is 1.8% higher than that of YOLO11n algorithm, and the number of parameters is 11.5% lower. It can not only meet the requirements of high precision, high speed and lightweight for road damage detection, but also has strong robustness and real-time.

查看全文查看/发表评论下载PDF阅读器