Research On Deep Learning-Based Pavement Disease Detection

In recent years,the significant increase in road usage has led to a rise in pavement diseases,resulting in a rapid reduction in the lifespan of roads.Simultaneously,the failure to repair pavement diseases promptly has resulted in frequent traffic accidents,making the detection of pavement diseases a prominent topic in current research.Particularly,there is a focus on the accuracy and real-time performance of detection algorithms.This thesis proposes an efficient object detection method based on deep learning to realize the automatic detection of pavement diseases.This method can effectively reduce manpower and resources,thereby contributing to the safe development of transportation.Traditional deep learning-based object detection methods often require the construction of large,high-precision convolutional neural networks.However,due to limited computational resources on small embedded and mobile devices,it becomes challenging to deploy such large network models in such application scenarios.To address this issue,this thesis first enhances the task-oriented nature of data using generative adversarial networks.Furthermore,an improved lightweight detection model called Ghost Net-YOLOX is proposed,enabling accurate and real-time automated detection of road defects.Lastly,the detection system is implemented on an embedded platform,and software development is completed using PyQt.The innovation and main contributions of this thesis are as follows:(1)This thesis proposes a pavement diseases detection method based on Generative Adversarial Networks(GANs)to address the issue of low learning efficiency caused by the loss of image details in pavement diseases images.The proposed method is an enhanced algorithm called UGAN-2G,based on the Wasserstein Generative Adversarial Networks,for reshaping key target images of transverse cracks,longitudinal cracks,and potholes.It also enhances the multitasking capability of target detection.The method introduces a self-mapping gradient combined with a U-shaped network as the generator,incorporating the Res Net concept.It achieves unsupervised reshaping of defect target images,thereby improving network efficiency.To enhance the effectiveness of reshaping,this thesis combines image pixel loss with adversarial loss as the training loss for the generator,thereby improving the robustness of the model.Finally,the generated data is applied to YOLOv5 and YOLOX models through image synthesis techniques for experimentation,demonstrating that this method improves the accuracy of automated pavement disease detection.(2)To address the issue of low utilization of the model due to limited computational resources in hardware environments,this thesis proposes a lightweight pavement diseases recognition model called Ghost Net-YOLOX,based on YOLOX.The model first replaces the YOLOX backbone network with the optimized Ghost Net to reduce the computational parameter volume of the network.It constructs the DAM(Dimensional Attention Model)to replace the SE module in the Ghost Bottleneck module,thereby effectively utilizing the limited network capacity for enhanced feature learning.Then,the DFM(Deep Fusion Model)module is proposed to improve PANet and perform a deep fusion of high and low-level feature layers to obtain richer feature information.Finally,the loss function is optimized to accurately fit the detection box positions,and the Image-Multitasking data augmentation method is introduced to strengthen the task-oriented nature of the target images,thereby improving the network’s generalization ability and robustness.(3)A pavement diseases detection system was designed based on the Jetson TX embedded platform.The system decodes the data stream using GStreamer technology,accelerates the model using Tensor RT,and solves video lagging issues through multithreading.Experimental results demonstrate that this detection system can meet the real-time and accuracy requirements of road surface detection.Lastly,the human-computer interaction software design for the detection model was implemented using PyQt.