Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (13): 305-316.DOI: 10.3778/j.issn.1002-8331.2301-0098

• Engineering and Applications • Previous Articles     Next Articles

Disease Detection of Asphalt Pavement Based on Improved YOLOv7

NI Changshuang, LI Lin, LUO Wenting, QIN Yong, YANG Zhen, FU Youhua   

  1. 1.School of Communications and Civil Engineering, Fujian Agriculture and Forestry University, Fuzhou 350100, China
    2.School of Transportation Engineering, Nanjing Tech University, Nanjing 211816, China
    3.State Key Laboratory of Rail Transit Control and Safety, Beijing Jiaotong University, Beijing 100084, China
    4.School of Traffic and Transportation, Beijing Jiaotong University, Beijing 100084, China
  • Online:2023-07-01 Published:2023-07-01

改进YOLOv7的沥青路面病害检测

倪昌双,李林,罗文婷,秦勇,杨振,傅幼华   

  1. 1.福建农林大学,交通与土木工程学院,福州 350100
    2.南京工业大学,交通运输工程学院,南京 211816
    3.北京交通大学 轨道交通控制与安全国家重点实验室,北京 100084
    4.北京交通大学 交通运输学院,北京 100084

Abstract: Aiming at the problems of low detection accuracy and inaccurate positioning in the detection of asphalt pavement diseases by traditional convolution network, an asphalt pavement disease detection algorithm based on improved YOLOv7 is proposed. Firstly, according to the imaging characteristics of the laser image, the combined filter-three histogram equalization algorithm is used to weaken the background environment interference. The K-means++clustering algorithm is used to set the initial anchor frame to speed up the convergence of the model. Then, the multi-head self-attention mechanism is combined with the maximum pooling layer to replace part of the convolution layer in the model backbone framework to improve the ability of the convolution network to learn the global features of the target. The funnel activation function F-ReLU is used as the activation function in the backbone network to expand the receptive field of the convolution layer. Finally, the A-SIOU loss function is used to optimize the model boundary box regression to accelerate the model convergence and improve the training accuracy. The experimental results show that the average precision, precision and recall rate of the improved detection algorithm for disease detection are 7.7, 9.4 and 5.8?percentage points higher than that of the original network, with better recognition accuracy. In practical engineering applications, the calculation deviation of the pavement condition index of each section is less than 1%, which is of great significance for promoting the intelligent detection of pavement diseases.

Key words: pavement diseases detection, deep learning, dimensional laser image, multi-head self-attention mechanism, loss function, activation function

摘要: 针对传统卷积网络对沥青路面病害检测时存在的检测精度低、定位不准等问题,提出一种基于改进YOLOv7的沥青路面病害检测算法。针对激光图像的成像特征,使用组合滤波-三直方图均衡化算法弱化背景环境干扰;使用K-means++聚类算法进行初始锚框设置来加快模型收敛速度;然后将多头自注意力机制与最大池化层结合代替模型主干框架中部分卷积层,提高卷积网络对于目标物全局特征学习能力;使用漏斗激活函数F-ReLU作为主干网络中的激活函数以扩大卷积层的感受野范围;最后使用A-SIOU损失函数优化模型边界框回归,加快模型收敛的同时提高训练精度。实验结果表明,改进后的检测算法对病害检测的平均精度均值、精确率和召回率相较原网络提升了7.7、9.4与5.8个百分点,具有较好的识别精度。在实际工程应用中,对各路段的路面状况指数的计算偏差均小于1%,对推进路面病害的智能化检测具有重要意义。

关键词: 路面病害检测, 深度学习, 激光图像, 多头自注意力机制, 损失函数, 激活函数