D3F-DET：轻量化多尺度融合的路面缺陷检测算法

doi:10.3778/j.issn.1002-8331.2502-0097

摘要/Abstract

摘要： 针对当前道路缺陷检测算法检测精度低、漏检误检率高、泛化能力欠佳的问题，提出了一种改进的轻量化检测模型，模型包含三种核心模块：轻量化双分支动态融合网络、动态重排列自注意力机制和跨分支交叉特征融合金字塔架构。轻量化双分支动态融合网络通过在双分支主干间共享信息，减少参数量的同时实现了特征空间的高效融合；动态重排列自注意力机制提供动态范围内的自适应特征聚合，显著提高了对动态范围变化目标的检测精度；跨分支交叉特征融合金字塔架构采用双向特征交互，增强了全局语义信息和局部细节特征的融合能力。实验结果表明，该模型在NRDD-2024数据集上的mAP@0.5达到了77.2%，对比最新的实时目标检测器RT-DETR参数量降低了29.63%，计算量降低了40.35%，精度提升了2个百分点，热力图对比显示模型可以给予更合适的注意力，同时对于新数据集展现出良好的泛化能力，为嵌入式设备上的实时路面缺陷检测提供了高效可靠的解决方案。

关键词: 道路缺陷检测, 目标检测, 轻量化网络, 深度学习, RT-DETR

Abstract: To address the current issues of low detection accuracy, high false negative and false positive rates, and poor generalization ability in road defect detection algorithms, an improved lightweight detection model is proposed. The model consists of three core modules： a lightweight dual-branch dynamic fusion network, a dynamic rearranged self-attention mechanism, and a cross-branch cross-feature fusion pyramid architecture. The lightweight dual-branch dynamic fusion network efficiently fuses feature spaces while reducing the number of parameters by sharing information between the dual branches. The dynamic rearranged self-attention mechanism provides adaptive feature aggregation within a dynamic range, significantly improving detection accuracy for targets with varying ranges. The cross-branch cross-feature fusion pyramid architecture enhances the fusion of global semantic information and local detailed features through bidirectional feature interaction. Experimental results show that the model achieves a mAP@0.5 of 77.2% on the NRDD-2024 dataset. Compared to the latest real-time object detector RT-DETR, the model reduces the number of parameters by 29.63%, decreases computational cost by 40.35%, and improves accuracy by 2 percentage points. Heatmap comparisons demonstrate that the model can provide more appropriate attention, and it shows good generalization ability on new datasets, offering an efficient and reliable solution for real-time road defect detection on embedded devices.

Key words: road defect detection, target detection, lightweight networks, deep learning, RT-DETR

贾翔宇, 张永宏, 阚希, 朱灵龙, 李旭. D3F-DET：轻量化多尺度融合的路面缺陷检测算法[J]. 计算机工程与应用, 2025, 61(17): 159-170.

JIA Xiangyu, ZHANG Yonghong, KAN Xi, ZHU Linglong, LI Xu. D3F-DET： Lightweight and Multiscale Fusion Algorithm for Pavement Defect Detection[J]. Computer Engineering and Applications, 2025, 61(17): 159-170.

参考文献

[1] 姚恩建. 综合立体交通专刊[J]. 北京交通大学学报, 2024, 48(4): 3.
YAO E J. Comprehensive three-dimensional transportation special issue[J]. Journal of Beijing Jiaotong University, 2024, 48(4): 3.
[2] 徐志刚, 车艳丽, 李金龙, 等. 路面破损图像自动处理技术研究进展[J]. 交通运输工程学报, 2019, 19(1): 172-190.
XU Z G, CHE Y L, LI J L, et al. Research progress on automatic image processing technology for pavement distress[J]. Journal of Traffic and Transportation Engineering, 2019, 19(1): 172-190.
[3] SZEGEDY C, TOSHEV A, ERHAN D. Deep neural networks for object detection[C]//Proceedings of the 27th International Conference on Neural Information Processing Systems, 2013: 2553 - 2561.
[4] ZOU Z X, CHEN K Y, SHI Z W, et al. Object detection in 20 years: a survey[J]. Proceedings of the IEEE, 2023, 111(3): 257-276.
[5] MAJIDIFARD H, ADU-GYAMFI Y, BUTTLAR W G. Deep machine learning approach to develop a new asphalt pavement condition index[J]. Construction and Building Materials, 2020, 247: 118513.
[6] WANG Q L, MAO J C, ZHAI X, et al. Improvements of YOLOv3 for road damage detection[J]. Journal of Physics: Conference Series, 2021, 1903(1): 012008.
[7] GUO G, ZHANG Z. Road damage detection algorithm for improved YOLOv5[J]. Scientific Reports, 2022, 12(1): 15523.
[8] DUAN K W, XIE L X, QI H G, et al. Corner proposal network for anchor-free, two-stage object detection[C]//Proceedings of the European Conference on Computer Vision. Cham: Springer International Publishing, 2020: 399-416.
[9] 向宽, 李松松, 栾明慧, 等. 基于改进 Faster RCNN 的铝材表面缺陷检测方法[J]. 仪器仪表学报, 2021, 42(1): 191-198.
XIANG K, LI S S, LUAN M H, et al. Aluminum product surface defect detection method based on improved Faster RCNN[J]. Chinese Journal of Scientific Instrument, 2021, 42(1): 191-198.
[10] 伊欣同, 单亚峰. 基于改进Faster R-CNN的光伏电池内部缺陷检测[J]. 电子测量与仪器学报, 2021, 35(1): 40-47.
YI X T, SHAN Y F. Photovoltaic cell internal defect detection based on improved Faster R-CNN[J]. Journal of Electronic Measurement and Instrumentation, 2021, 35(1): 40-47.
[11] HOSANG J, BENENSON R, SCHIELE B. Learning non-maximum suppression[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 6469-6477.
[12] REZATOFIGHI H, TSOI N, GWAK J, et al. Generalized intersection over union: a metric and a loss for bounding box regression[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2019: 658-666.
[13] HAN K, XIAO A, WU E, et al. Transformer in transformer[J]. arXiv:2103.00112, 2021.
[14] CARION N, MASSA F, SYNNAEVE G, et al. End-to-end object detection with transformers[C]//Proceedings of the European Conference on Computer Vision. Cham: Springer International Publishing, 2020: 213-229.
[15] ZHAO Y A, LYU W Y, XU S L, et al. DETRs beat YOLOs on real-time object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2024: 16965-16974.
[16] LUO G, HUANG M, ZHOU Y, et al. Towards efficient visual adaption via structural re-parameterization[J]. arXiv:2302. 08106, 2023.
[17] TARG S, ALMEIDA D, LYMAN K. ResNet in ResNet: generalizing residual architectures[J]. arXiv:1603.08029, 2016.
[18] ZHANG H, LI F, LIU S L, et al. DINO: DETR with improved DeNoising anchor boxes for end-to-end object detection[J]. arXiv:2203.03605, 2022.
[19] YUE M, ZHANG L Q, HUANG J, et al. Lightweight and efficient tiny-object detection based on improved YOLOv8n for UAV aerial images[J]. Drones, 2024, 8(7): 276.
[20] JIN Y, TIAN X Y, ZHANG Z, et al. C2F: an effective coarse-to-fine network for video summarization[J]. Image and Vision Computing, 2024, 144: 104962.
[21] ZHANG P F, LO E, LU B T. High performance depthwise and pointwise convolutions on mobile devices[C]//Proceedings of the AAAI Conference on Artificial Intelligence, 2020: 6795-6802.
[22] LI H F, ZHAO J Z, LI J X, et al. Feature dynamic alignment and refinement for infrared-visible image fusion: translation robust fusion[J]. Information Fusion, 2023, 95: 26-41.
[23] SUN S Q, REN W Q, GAO X W, et al. Restoring images in adverse weather conditions via histogram transformer[C]//Proceedings of the European Conference on Computer Vision. Cham: Springer International Publishing, 2025: 111-129.
[24] KAYA ?, ?ODUR M Y. N-RDD2024: road damage and defects[EB/OL]. (2024-01-09) [2025-01-01]. https://data.mendeley.com/datasets/27c8pwsd6v/4.
[25] JIANG H, LEARNED-MILLER E. Face detection with the faster R-CNN[C]//Proceedings of the 12th IEEE International Conference on Automatic Face & Gesture Recognition. Piscataway: IEEE, 2017: 650-657.
[26] JOCHER G, CHAURASIA A, STOKEN A, et al. Ultralytics/YOLOv5: v7.0-YOLOv5 sota realtime instance segmentation[J]. Zenodo, 2022.
[27] LI C, LI L, JIANG H, et al. YOLOv6: a single-stage object detection framework for industrial applications[J]. arXiv:2209. 02976, 2022.
[28] WANG C Y, BOCHKOVSKIY A, LIAO H M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2023: 7464-7475.
[29] JOCHER G, CHAURASIA A, QIU J. Ultralytics YOLO (version 8.0.0)[computer software][EB/OL]. (2023-12-01) [2024-07-10]. https://docs.ultralytics.com/zh/models/yolov8.
[30] WANG C C, HE W, NIE Y, et al. Gold-YOLO: efficient object detector via gather-and-distribute mechanism[J]. arXiv: 2309.11331, 2023.
[31] WANG A, CHEN H, LIU L H, et al. YOLOv10: real-time end-to-end object detection[J]. arXiv:2405.14458, 2024.
[32] WANG Z, LI C, XU H, et al. Mamba YOLO: SSMs-Based YOLO for object detection[J]. arXiv:2406.05835, 2024.
[33] ARYA D, MAEDA H, GHOSH S K, et al. RDD2022: a multi‐national image dataset for automatic road damage dete-ction[J]. Geoscience Data Journal, 2024, 11(4): 846-862.