基于改进YOLOv8的轻量化道路病害检测方法

doi:10.3778/j.issn.1002-8331.2402-0243

摘要/Abstract

摘要： 针对复杂场景下道路病害检测模型占用内存空间大、计算复杂度高和检测速度难以满足实时目标检测要求等问题，提出一种面向复杂自然场景的轻量级道路病害检测模型DGE-YOLO-P。将网络中的C2f融合可变形卷积设计C2f_DCNv3模块增强对物体形变的建模能力，并对输入特征信息进行降维处理，有效降低模型参数量和计算复杂度。设计GS-Decoupled head检测模块，降低检测头参数的同时实现全局信息的有效聚合。同时，设计E-Slide Loss权重函数，为困难样本分配更高权重，充分学习道路病害中的难样本数据，进一步提高模型检测精度。采用通道剪枝减少模型冗余通道，有效压缩模型体积并提高检测速度。实验结果表明，DGE-YOLO-P模型相较于YOLOv8n模型mAP提高2.4个百分点，而模型参数量、计算量和模型大小分别降低58.1%、66.7%和55.5%。检测速度FPS由34帧/s提高到51帧/s。

关键词: 道路病害检测, 复杂场景, YOLOv8n, 轻量化, 模型剪枝

Abstract: Aiming at the problems of large memory space occupation, high computational complexity, and difficult to meet the real-time target detection requirements of the road damage detection model in complex scenes, a lightweight road damage detection model DGE-YOLO-P is proposed for the complex natural scenes. Firstly, the C2f fusion deformable convolutional design C2f_DCNv3 module in the network is enhanced to enhance the modelling capability of object deformation and the input feature information is dimensionality reduced to effectively reduce the number of parameters and the computational complexity. The input feature information is dimensionality reduced to effectively reduce the number of model parameters and computational complexity. Then, the GS-Decoupled head detection module is designed to reduce the parameters of the detection head while realising the effective aggregation of global information. At the same time, the E-Slide Loss weight function is designed to assign higher weights to the difficult samples, fully learn the difficult sample data in road damage, and further improve the model detection accuracy. Finally, channel pruning is used to reduce the redundant channels of the model, which effectively compresses the model volume and improves the detection speed. The experimental results show that the mAP of the DGE-YOLO-P model is increased by 2.4?percentage points compared with the YOLOv8n model, while the number of model parameters, computational volume and model size are reduced by 58.1%, 66.7% and 55.5%, respectively. The detection speed FPS is increased from 34 frame/s to 51 frame/s.

Key words: road damage detection, complex scene, YOLOv8n, lightweight, model pruning

胥铁峰, 黄河, 张红民, 牛晓富. 基于改进YOLOv8的轻量化道路病害检测方法[J]. 计算机工程与应用, 2024, 60(14): 175-186.

XU Tiefeng, HUANG He, ZHANG Hongmin, NIU Xiaofu. Lightweight Road Damage Detection Method Based on Improved YOLOv8[J]. Computer Engineering and Applications, 2024, 60(14): 175-186.

参考文献

[1] DEVI M P A, LATHA T, SULOCHANA C H. Iterative thresholding based image segmentation using 2D improved Otsu algorithm[C]//Proceedings of the Communication Technologies, 2015: 145-149.
[2] 徐欢, 李振璧, 姜媛媛, 等. 基于Open CV和改进Canny算子的路面裂缝检测[J]. 计算机工程与设计, 2014, 35(12): 4254-4258.
XU H, LI Z B, JIANG Y Y, et al. Pavement crack detection based on OpenCV and improved Canny operator[J]. Computer Engineering and Design, 2014, 35(12): 4254-4258.
[3] HU W B,WANG W D,AI C B,et al. Machine vision-based surface crack analysis for transportation infrastructure[J]. Automation in Construction, 2021, 132:103973.
[4] GIRSHICK R, DONAHUE J, DARRELL T,et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014: 582-587.
[5] GIRSHICK R. Fast R-CNN[C]//Proceedings of the IEEE International Conference on Computer Vision, 2015: 1440-1448.
[6] REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[7] HE K M,GKIOXARI G,DOLLáR P,et al.Mask R-CNN[C]//Proceedings of the IEEE International Conference on Computer Vision, 2017: 2980-2988.
[8] 孙朝云, 裴莉莉, 李伟, 等. 基于改进Faster R-CNN的路面灌封裂缝检测方法[J]. 华南理工大学学报 (自然科学版), 2020, 48(2): 84-93.
SUN Z Y, PEI L L, LI W, et al. Pavement grouting crack detection method based on improved Faster R-CNN[J]. Journal of South China University of Technology (Natural Science Edition), 2020, 48(2): 84-93.
[9] 牛慧余, 包腾飞, 李扬涛, 等. 基于改进Mask R-CNN的混凝土坝裂缝像素级检测方法[J]. 水利水电科技进展, 2023, 43(1): 87-92.
NIU H Y, BAO T F, LI Y T, et al. Pixel-level crack detection method of concrete dam based on improved mask R-CNN[J]. Advances in Science and Technology of Water Resources, 2023, 43(1): 87-92.
[10] LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multi-box detector[C]//Proceedings of the European Conference on Computer Vision, 2016: 21-37.
[11] REDMON J, FARHADI A. YOLOv3: an incremental improvement[EB/OL]. (2018-04-08)[2024-01-02]. https://arxiv.org/abs/1804.02767.
[12] BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4:optimal speed and accuracy of object detection[EB/OL].(2020-04-23)[2024-01-02]. http://arxiv.org/abs/2004.10934.
[13] WANG C Y, BOCHKOVSKIY A, LIAO H Y 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,2023: 7464-7475.
[14] 韩豫, 张萌, 李宇宏, 等. 基于深度学习和ArcMap的路面病害智能综合检测方法[J]. 江苏大学学报 (自然科学版), 2023, 44(4): 490-496.
HAN Y, ZHANG M, LI Y H, et al. Intelligent comprehensive detection method of pavement diseases based on deep learning and ArcMap[J]. Journal of Jiangsu University (Natural Science Edition), 2023, 44(4): 490-496.
[15] CUI X N, WANG Q C, DAI J P, et al. Intelligent recognition of erosion damage to concrete based on improved YOLO-v3[J].Materials Letters, 2021, 302: 130363.
[16] 安学刚, 党建武, 王阳萍, 等. 基于改进YOLOv4的无人机影像路面病害检测方法[J]. 无线电工程, 2023, 53(6): 1285-1294.
AN X G, DANG J W, WANG Y P, et al. UAV image pavement disease detection based on improved YOLOv4[J]. Radio Engineering, 2023, 53(6): 1285-1294.
[17] WU C G, YE M, ZHANG J L, et al. YOLO-LWNet: a lightweight road damage object detection network for mobile terminal devices[J]. Sensors, 2023, 23(6): 3268.
[18] 王海群, 王炳楠, 葛超. 重参数化YOLOv8路面病害检测算法[J]. 计算机工程与应用, 2024, 60(5): 191-199.
WANG H Q, WANG B N, GE C. Re-parameterized YOLOv8 pavement disease detection algorithm[J].Computer Engineering and Applications, 2024, 60(5): 191-199.
[19] 王春梅, 刘欢. YOLOv8-VSC: 一种轻量级的带钢表面缺陷检测算法[J]. 计算机科学与探索, 2024, 18(1): 151-160.
WANG C M, LIU H. YOLOv8-VSC: lightweight algorithm for strip surface defect detection[J]. Journal of Frontiers of Computer Science & Technology, 2024, 18(1): 151-160.
[20] WANG W H, DAI J F, CHEN Z, et al. InternImage: exploring large-scale vision foundation models with deformable convolutions[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023: 14408-14419.
[21] CHOLLET F. Xception: deep learning with depthwise separable convolutions[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017: 1800-1807.
[22] LI H L, LI J, WEI H B, et al. Slim-neck by GSConv: a better design paradigm of detector architectures for autonomous vehicles[EB/OL].(2022-06-06)[2023-09-15]. https://arxiv.org/abs/2206.02424.
[23] GE Z, LIU S T, WANG F, et al. YOLOX: exceeding YOLO series in 2021[EB/OL].(2021-08-06)[2023-10-06]. https://arxiv.org/abs/2107.08430.
[24] YU Z P, HUANG H B, CHEN W J, et al.YOLO-FaceV2: a scale and occlusion aware face detector[J]. arXiv:2208.02019, 2022.
[25] HAN S, MAO H Z, DALLY W J. Deep compression: compressing deep neural networks with pruning[J]. Fiber, 2015, 56(4): 3-7.
[26] ARYA D, MAEDA H, GHOSH S K, et al. RDD2022: a multi-national image dataset for automatic road damage detection[J]. arXiv:2209.08538, 2022.
[27] LV W Y, ZHAO Y A, XU S L, et al. DETRs beat YOLOs on real-time object detection[EB/OL]. (2023-07-16) [2023-12-17]. https://arxiv.org/abs/2304.08069.
[28] WANG C Y, YEH H I, LIAO H P. YOLOv9: learning what you want to learn using programmable gradient information[EB/OL]. (2024-02-21) [2024-03-25]. https://arxiv.org/abs/2402.13616.
[29] ZHU J Q, ZHONG J T, MA T, et al. Pavement distress detection using convolutional neural networks with images captured via UAV[J]. Automation in Construction, 2022, 133:103991.