基于深度学习的公路路面病害检测算法

Pavement Defect Detection Algorithm Based on Deep Learning

针对公路路面病害图像存在背景干扰多、病害信息弱、尺度差异大等问题,提出了一种基于深度学习的公路路面病害检测方法。以YOLOv4算法为基础,在检测网络中引入可变形卷积,并提出基于路径聚合网络(path aggregation network,PANet)的自适应空间特征融合结构,充分学习公路路面病害的细节特征,实现不同尺度特征信息的高效融合;采用平均准确率损失(average precision loss,AP-loss)函数作为分类损失函数,促使网络在训练过程中更加注重于正样本。实验表明,在公路路面病害检测中,改进YOLOv4算法的平均准确率达到了95.34%,每张图像的平均检测时间为0.071 s。与快速基于区域的卷积神经网络(faster region-based convolutional neural networks,Faster R-CNN)算法相比,所提出的算法在持有较高检测准确率的同时,减少了运算时间,可以满足公路路面病害检测的准确性与实时性需求。

Aiming at the problems of background interference,weak defect information and big difference in pavement defect images,a pavement defect detection algorithm based on deep learning was proposed.Based on YOLOv4 algorithm,deformable convolution was introduced into the detection network,and an adaptive spatial feature fusion structure based on path aggregation network(PANet)was proposed to fully learn the detailed features of pavement defect and realize the efficient fusion of feature information at different scales.Average precision loss(AP-loss)function was used as the classification loss function of the network,which made the network pay more attention to positive samples in the training process.Experiments show that the average accuracy of the proposed algorithm can reach 95.34%,and the average detection time of each image can reach 0.071 s.Compared with faster region-based convolutional neural networks(Faster R-CNN)algorithm,the proposed algorithm has a higher detection accuracy and reduces the operation time,which can meet the accuracy and real-time requirements of pavement defect detection.