Road Defect Recognition Based On Deep Neural Networks

In recent years,China’s highways have achieved rapid development.However,with the passage of vehicles,many difficult maintenance problems have arisen.Cracks are one of the most common,easily occurring,and earliest forms of damage on road surfaces,which accompany the entire service life of the road and become more serious as the road ages.Therefore,accurate detection of road cracks is of great significance for road maintenance.Traditional image processing algorithms have certain defects in road crack detection,such as being sensitive to image interference such as lighting and shadows,requiring a large amount of human and material resources,and having poor accuracy and robustness.Road crack detection methods based on computer vision have higher accuracy,stronger generalization ability,higher automation,and faster detection speed,and can better meet the needs of road crack detection in practical scenarios.This article focuses on the research of road crack detection and segmentation algorithms based on computer vision.(1)This thesis proposes an improved YOLOv5 crack detection model based on attention mechanism.In order to address the problem that existing object detection networks tend to ignore small-width road cracks in images,this thesis introduces a hybrid attention mechanism to enhance the feature extraction ability of the model.Through experiments,the performance differences of introducing attention modules at different network stages are demonstrated,and a temperature parameter is introduced to improve the loss function,which can solve the problem of imbalance between positive and negative samples of cracks and backgrounds in the image.After improvement,the average precision of the model on the crack dataset is increased by 2.5%.(2)This thesis proposes a detection-segmentation multitask crack extraction network model.In order to meet the practical maintenance requirements of quickly and accurately detecting cracks,determining their contour size and location,this thesis proposes a multitask network crack extraction model.By using the multitask network,instance-level annotations can be used to train the classification task,and pixel-level annotations can be used to train the segmentation task,achieving sufficient utilization of difficult-to-obtain pixel-level annotations and easily obtained instance-level annotations.By sharing the backbone network,the same feature maps can be avoided from being computed repeatedly,and the already extracted features can be fully utilized,reducing the computation and memory consumption of the model and improving its efficiency,generalization ability,and accuracy.The experimental results show that the multitask shared backbone network can mutually promote each other through joint training,achieving better performance.(3)This thesis proposes a weakly supervised semantic segmentation model for crack extraction based on the idea of adversarial networks.To address the problems of requiring a large number of samples for segmentation model training and difficulty in creating and obtaining segmentation datasets,this thesis designs two different segmentation networks as generating networks based on the idea of adversarial networks to increase label diversity and strengthen the association between segmentation categories.The discriminator network judges whether the two outputs generated by the two generating networks are correct through two different inputs,thereby promoting competition and cooperation between the two generating networks and enhancing the model’s ability to learn crack features.In the training process of the model,a weakly supervised learning method is adopted,and a small amount of labeled crack images are used for training.Through the competition and cooperation mechanism of the adversarial network,semantic segmentation of unlabeled images is achieved.Experimental results show that the proposed model performs well in crack segmentation tasks,has strong practical value and application prospects,and exhibits good performance and generalization ability.Overall,the thesis has made progress in using deep learning algorithms for crack detection,addressing issues such as small crack width and a lack of pixel-level annotated samples in segmentation datasets.