Research On Data Architecture And Optimization Method For Deep Learning Pavement Disease Detection

With the vigorous development of China’s transportation industry and the continuous expansion of the highway network,the highway system has gradually transitioned from the construction period to the maintenance and management period.Disease detection is one of the core tasks of road maintenance management,but traditional manual detection methods have shortcomings such as low efficiency,poor accuracy,and high labor intensity,which cannot meet the needs of modern maintenance management.In recent years,with the rapid development of artificial intelligence technology,automated detection technology for road surface diseases has been widely applied.Due to the complexity of the road environment and the diversity of road surface diseases,the detection accuracy of road surface disease detection models based on deep learning is not high.In response to this issue,most current research has chosen to improve deep learning model algorithms,and data is the foundation of deep learning.Even the best model is limited by the lack of training data,so optimizing the dataset is very necessary.This article takes the road disease training dataset as the research object.In response to the problems of unreasonable data distribution,imbalanced category distribution,and complex background interference during model training,a method for constructing the road disease training dataset is proposed.The initial dataset is constructed using the Kmeans++clustering precision segmentation algorithm,image augmentation balance algorithm,and the creation of image anti-interference labels to improve the accuracy of the road disease detection model.By providing feedback on experimental results,targeted data expansion is carried out on the dataset to optimize the pavement disease dataset and further improve the generalization of the model.The main research content of this article is as follows: Firstly,different initial pavement disease datasets are constructed by dividing the training set and validation set through random partitioning and stratified sampling.By comparing the evaluation indicators of pavement disease detection models under different YOLO models and different dataset construction methods,the feasibility of improving the composition of the dataset to improve model accuracy was demonstrated,and the problems in the pavement disease detection model were analyzed from the data level.Secondly,in response to the problems with the model,a method for constructing a training dataset for road disease detection is proposed.This method includes the Kmeans++ clustering refinement algorithm,image augmentation and equalization algorithm,and the creation of image anti-interference labels.By constructing the original pavement disease dataset II,we obtained an anti-interference fine classification balanced dataset.Through experimental comparison,the pavement disease detection model trained on the anti-interference fine classification balanced dataset has a 22.1% improvement in m AP0.5:0.95 compared to the original pavement disease dataset II,and the model’s detection performance has been significantly improved.Next,in response to the missed and false detections of diseases in the actual test section,the anti-interference fine division balance dataset was supplemented and optimized to obtain a replacement and augmented fine division balance dataset.In the secondary test section,the comprehensive index F1 value of the pavement disease detection model increased by 2.2%,and the balance index of the real pavement disease dataset increased by 6%.Finally,build a road surface disease dataset management system to achieve automated architecture and optimization of the road surface disease dataset.The dataset architecture and optimization methods studied in this article improve the distribution rationality and balanced diversity of the pavement disease detection dataset,enhance the training effect of the model,have high accuracy and universality,and can provide a new idea and method for automated pavement disease detection.It can obtain more scientific and effective pavement disease detection results with less investment cost,and has certain practical engineering application value.