Research On Road Damage Detection Model Based On Lightweight Network

In recent years,with the development of underlying hardware and high-performance computing platforms,deep learning algorithms have been widely used in the field of computer vision with the help of abundant computing power.It exhibits excellent performance on a variety of key tasks including image classification,object detection,semantic segmentation,and instance segmentation,making image processing technologies,especially object detection technologies,have achieved great development.However,in actual engineering applications,the object detection model based on deep learning has problems such as high complexity,large amount of parameters,and large computational overhead.There are high requirements of computational ability for terminal devices in the process of model training and inference,which leads to problems such as difficulty in deployment of models on low-performance computing platforms such as embedded or mobile devices,or poor real-time detection after deployment.In order to solve the problems of high complexity of the object detection model,large amount of parameters,and large computational overhead,the paper focuses on the lightweight network model of the road damage detection task,from the lightweight network structure design and model compression and acceleration algorithms.In terms of reducing the complexity of the object detection model,compressing the amount of model parameters,and accelerating model inference.The main research contents and work results are as follows:(1)Extensive research and evaluation are carried out on mainstream object detection algorithms and lightweight model algorithms,and the YOLOv4 algorithm is used to train on a self-made optimized dataset to obtain a road damage detection model.(2)For YOLOv4’s backbone feature extraction network CSPDarkNet53,there is a problem of high complexity,and the lightweight network MobileNet is introduced to improve the YOLOv4 network architecture.Aiming at the problem that the improved YOLOv4-MobileNet model produces a large loss of accuracy,it is proposed to introduce a lightweight CBAM attention mechanism to enhance the feature perception ability of the network,and reconstruct the basic structural units of Bottlenet.The improved YOLOv4-MobileNet+CBAM model has 29.3M parameters,and the real-time inference speed reaches 112.6FPS.The compression rate and acceleration rate of the model are both more than 2 times,while the overall accuracy loss of the model is only 1.0%.The results show that the lightweight network structure can reduce the model complexity while ensuring low accuracy loss.(3)Aiming at the problem of large parameter redundancy in the improved YOLOv4-MobileNet+CBAM model,a lightweight algorithm based on network pruning is introduced.Aiming at the problem that the Shortcut cross-layer connection structure in the model network is difficult to match the number of channels in each layer under a single pruning method,two optimization schemes,passive pruning and active pruning,are proposed,and different sparsity rates and pruning are selected.The branch rate parameter to achieve the best pruning effect.The pruned road surface damage target detection model Pruned_Model has only 5.9M parameters,and the inference speed reaches 324.8FPS,which is 5 times compressed compared to the original model,accelerated nearly 3 times,and the overall accuracy loss is only 2.1%.The results show that the pruning scheme is effective.