Lightweight Model of Pavement Damage Detection Based on YOLOV7 STBZ Architecture
YOLOV7 has problems such as parameter redundancy,low detection efficiency,and excessive model volume when used for pavement disease detection.To solve the above problems,a MobileOne network is used to replace the original backbone feature extraction network.The lightweight coordinate attention module Coordinate Attention(CA)is introduced into the network.The YOLOV7 Slim-Neck based on GSConv is proposed,and the model operation process is fused by the idea of multiple parameters.In training,Focal-EIoU Loss is used instead of CIoU Loss to solve the problem of unbalanced sample classification.The results show that the introduction of Focal-EIoU Loss can balance the inhomogeneity of samples and accelerate the convergence of the model.The MobileOne network improves the accuracy of the model while reducing the model parameters.The CA module strengthens the feature extraction ability of the model.Slim-Neck greatly reduces the model parameters while en-suring the learning ability.Compared with the traditional YOLOV7 architecture,the proposed model parameters are reduced by 78％,the image processing speed is three times of the original,the mAP performance index is in-creased by 1.71％,and the F score is increased by 0.022.The lightweight model is convenient for mobile plat-form deployment and has obvious engineering application advantages.
pavement engineeringpavement damagetarget detectionYOLOV7lightweightattention mecha-nismloss function
万方数据
维普
