Research On Generative Adversarial Network-Based Data Augmentation Method For Asphalt Pavement Disease

Nowadays,China’s road network construction is improving day by day,and the large-scale road construction brings heavy road maintenance tasks at the same time.In the face of large-scale road inspection tasks,it has important engineering value and research significance to identify pavement diseases quickly and accurately,and to accurately prevent and treat road diseases.Compared with traditional manual visual inspection,deep learning-based pavement inspection technology is more efficient and accurate,but its accuracy is constrained by the quality and quantity of the training dataset,the pavement disease image screening is time-consuming and expensive to produce,and the existing open source pavement disease dataset has a single type of disease and low quality,which limits the application of deep learning technology in pavement disease inspection work.To address the above problems,this paper proposes a deep convolutional generation adversarial network-based pavement disease image data enhancement method to analyze the generation effect for different pavement diseases and to optimize and improve the generation model.The main research work of this paper is as follows:(1)Construct pavement disease image augmentation dataset.The captured pavement video is transformed into frame images,the images containing pavement diseases are screened out,and the images are cropped and resized to obtain the original disease dataset.The images are pre-processed using mean filtering,median filtering,grayscale transformation,etc.(2)Pavement damage dataset sample expansion.To address the problems of blurred generated images,unstable training and poor diversity of generated samples in the DCGAN model for the pavement potholes image generation task,the model is improved in terms of network structure and loss function.A hybrid attention module with selfattention mechanism and channel attention mechanism is introduced to enable the network to effectively extract target features and coordinate local features of potholes with global features;the network depth is improved and the residual module is introduced to generate higher resolution images;the bulldozer distance with gradient penalty is used instead of JS scatter to improve the training stability of the model.The generated images are analyzed qualitatively and quantitatively by comparison experiments and ablation experiments.The experimental results show that the images generated by the improved RM-DCGAN model are better than the original DCGAN model in terms of quality and diversity.(3)Evaluation of the amplification effect of the dataset.By amplifying the dataset using traditional methods and the RM-DCGAN model proposed in this article,two datasets were used to train the YOLOv5 model.The experimental results show that the detection accuracy of the model trained on the expanded dataset using traditional methods is 9.31% higher than the average detection accuracy of the model trained on the benchmark dataset;The average detection accuracy of the model trained using RMDGCAN to generate image augmentation datasets has increased by 15.17%;The average detection accuracy of the model trained using the RM-DCGAN model to generate image augmentation datasets is 5.86% higher than that of the datasets amplified using traditional methods.Compared with the traditional data enhancement methods,the RM-DCGAN Generative model proposed in this paper significantly improves the quality and diversity of the pavement disease data set,and can significantly improve the detection accuracy of the model.