一种基于YOLOX优化的轻量级路面病害检测方法

Light pavement disease detection method based on optimized YOLOX

受限于小型嵌入式设备以及移动设备等有限的计算资源,大型网络模型难以部署在此类应用场景中.为了解决该问题,基于YOLOX提出一种高效的路面病害识别模型.首先,将YOLOX主干网络替换为优化后的GhostNet来减少网络计算参数,并参考基于卷积块注意力机制兼顾空间和通道方向上自适应调整信息的优势,构建DAM(Dimensional Attention Model)代替GhostBottleneck模块中的SE模块,从而充分利用有限的网络容量进行强化特征学习;其次,提出DFM(Deep Fusion Model)模块来改进PANet并以此对高低特征层进行深度融合,获取更加丰富的特征信息来提高检测能力;再次,采用Complete-IoU Loss来拟合更加准确的检测框位置,减少方向误判的同时提高了检测效率;最后,引入Image-Multitasking数据增强方法来强化目标图像任务性,提高了网络的泛化能力和鲁棒性.在RDD2020数据集上进行模型对比,实验表明,改进后的GhostNet-YOLOX网络的mAP达到84.05%,高于现有的YOLOX-s(即66.26%),模型参数量缩小至14.53 MB,小于YOLOX-s(即34.21 MB),同时实际检测视频的帧数达到了26 p·s^(-1),提高了5.88 p·s^(-1),检测实时性显著提高.

Building a large convolutional neural network to improve the model accuracy is an effective method.However,limited by small embedded devices and mobile devices limited computing resources,large network model is difficult to deploy in such application scenarios.To solve this problem,this paper presents an efficient pavement disease recognition model based on YOLOX,which has high detection accuracy even in the application scenarios with limited computational resources.First,the YOLOX backbone network is replaced with the optimized GhostNet to reduce the network computing parameters,and by referring to the advantages of convolutional block attention module and adaptive adjustment information in space and channel direction,the DAM(Dimensional Attention Model)is built to replace the SE module in the GhostBottleneck module,so as to make full use of the limited network capacity for reinforcement feature learning.Secondly,the DFM(Deep Fusion Model)module is proposed to improve the PANet and to deeply integrate the high and low feature layers to obtain richer feature information to improve the detection ability.Thirdly,the Complete-IoU Loss is introduced to fit a more accurate detection box position,reduce direction misjudgment and improve the detection efficiency.Finally,the Image-Multitasking data enhancement method is used to enhance the target image tasking,improving the generalization ability and robustness of the network.Model comparison was performed on the RDD2020 dataset,as shown by experiments,the improved GhostNet-YOLOX network achieved an mAP of 84.05%,higher than the existing YOLOX-s(66.26%),the number of model parameters is reduced to 14.53 MB,less than YOLOX-s(34.21 MB).Meanwhile,the number of frames of the actual detected video reached 26 p·s^(-1),Raised by 5.88 p·s^(-1),the real-time detection is significantly improved.