有雾场景下的道路路面病害检测算法

Recognition Algorithm of Road Pavement Disease in Foggy Scene

针对有雾场景下的道路病害数据集较少,导致基于深度学习的目标检测模型在有雾场景下的道路病害检测过程中出现的过拟合、低鲁棒性和低准确率等问题,提出一种基于AECR-YOLOv5融合机制的雾天路面病害检测算法。首先,根据道路路面病害的特性,设计一种以病害类别为判别依据的数据集扩充方法,并以此方法获得扩充训练集;同时,使用大气散射模型将清晰图与双边滤波处理后的深度图进行合成,从而增强合成雾图的可信性。然后,根据任务依赖不定性对去雾网络和检测网络的损失函数进行加权融合,获得耦合损失函数,促进两者的有效融合;最后,通过改变输入和网络结构进行实验,结果表明,输入为扩充雾图训练集且网络结构为AECR-YOLOv5融合模型时,其在测试集上的mAP均高于其他组合方式。

Aiming at the problems of over-fitting,poor robustness and low accuracy of the target detection model based on deep learning throughout the process of road disease detection in foggy scene,a foggy road disease detection algorithm based on AECR-YOLOv5 fusion mechanism is proposed.Firstly,according to the characteristics of road pavement diseases,a data set expan-sion method depend on disease category is designed,and the expanded training set is obtained by this method.Meanwhile,clear picture and the depth picture after bilateral filtering are synthesized on the basis of atmospheric scattering model,so as to enhance the credibility of the synthetic foggy picture.Then,according to the uncertainty of task dependence,the loss functions of defogging network and detection network are weighted and fused to obtain the coupling loss function and promote the effective fusion of them.Finally,experiments are carried out by changing the input and network structure.Results show that when the input is the extended foggy picture training set and the network structure is AECR-YOLOv5 fusion model,its mAP on the test set is higher than other combination methods.