Research On Intelligent Detection And Recognition Of Road Cracks Based On Transfer Learning

With the rapid development of the transportation industry,the tasks and load of highway transportation are continuously increasing,leading to various types of road diseases,including cracks.Cracks not only affect the road surface’s lifespan but also pose serious safety hazards for vehicle travel.Traditional manual detection methods have limitations,while deep learning neural networks have the potential to improve detection accuracy and efficiency.This thesis use a one-stage object detection network which has a fast detection speed,improved feature extraction and fusion performance,and efficient model training through transfer learning technology to achieve accurate detection and recognition of various types of cracks.The main contributions of this thesis are as follows:(1)Dataset preprocessing.The public road crack dataset is collected,supplemented by partial road crack images,and annotated to complete the dataset.Finally,the dataset is augmented and grayscale processed to form a complete road crack dataset.(2)Proposed crack detection and recognition algorithm with fusion attention mechanism.Based on the single-stage detection and recognition network,the convolution layers are added to the backbone network to extract deeper features,and an attention mechanism and residual structure are used to enhance feature extraction performance.The feature fusion part utilizes multi-scale feature stitching and a self-attention mechanism to improve the fusion effect and improve the accuracy of crack detection and recognition.(3)Experimental verification and result analysis.The performance of the model is experimentally verified and evaluated from multiple aspects,including detection accuracy,robustness,and detection speed.The results demonstrate that the crack detection algorithm with fusion attention mechanism improves the detection accuracy and recognition accuracy of four types of cracks,namely transverse,longitudinal,alligator,and mesh,while also improving the detection speed.Additionally,the model transfer method is used to pre train the model,accelerating the training speed of the model.(4)The road crack detection and recognition system is designed and implemented using the B/S architecture.The system includes system management,model management,and crack detection and recognition modules,which enable functions such as model deployment and update,crack detection and recognition,and image management.The thesis proposes a crack detection and recognition algorithm with a fusion attention mechanism to detect and recognize four types of road cracks: transverse,longitudinal,alligator,and mesh.The algorithm improves detection accuracy and efficiency.At the same time,a road crack detection and recognition system has been designed and implemented,promoting the application of road crack detection and recognition algorithms in practical life,which has certain theoretical research significance and practical application value.