| Nowadays,transportation has become an important driving force for national economic development,of which highway transportation is an important component.With the rapid development of highway transportation in China,road surface diseases are becoming increasingly serious.If not repaired in a timely manner,it will cause greater harm and pose a threat to traffic safety.Road surface cracks are the most common manifestation of road surface diseases.Therefore,crack detection has become an important means of highway disease prevention and maintenance,which has great practical significance in highway maintenance.With the development of emerging technologies such as computer hardware and deep learning,crack detection has shifted from manual detection to processing using computer image technology,saving a lot of manpower and material resources and improving detection efficiency.Pavement crack detection is an image detection method that targets pavement cracks.Early object detection used traditional algorithms for detection,but the drawback was that it required manual extraction of image features and reconstruction and expression.With the development of deep learning,the emergence of R-CNN represents the entry of convolutional neural networks into the field of target detection,and the subsequent advent of YOLO algorithm represents the entry of a mature stage in the field of target detection,with YOLOv5 algorithm being the most mature.This topic is based on the YOLOv5 s algorithm for multi-stage improvement.The specific content includes:(1)Aiming at the large number of samples in the asphalt pavement crack data set,a sliding window based system sampling algorithm is designed to sample the excessive number of samples,and then the training set and test set are divided.(2)According to the characteristics of pavement cracks,a Mosaic data enhancement algorithm based on multi enhancement fusion is proposed in the data enhancement stage of YOLOv5 s algorithm,so as to obtain a model with better generalization ability.(3)In the Backone stage of the YOLOv5 s algorithm,the focus on important information is not strong,and there is no attention mechanism for feature selection at the end of the Backone stage.The attention mechanism CBAM module is added in the final stage of Backone.(4)At the same time,a Sim SPPFCSPC module was designed to replace the SPPF module that did not work well after the end of the Backbone class.(5)In the Neck stage of the YOLOv5 s algorithm,GSConv is introduced,replacing the original CBS and C3 modules with corresponding GSCBS and Vo VGSC3 modules,fully and effectively utilizing the advantages of deep separable convolution and reducing its disadvantages.(6)In the Head stage of the YOLOv5 s algorithm,the original detection head is decoupled into two,FC-Head focuses on detection and classification tasks,and Conv-Head focuses on positioning and regression tasks.Through comparative experiments,it has been shown that the accuracy rate has been improved by about 12%,the recall rate has been improved by about 15%,and the m AP has been improved by about 18%,achieving very good results. |
