Deep Learning Defect Detection Research for Road Maintenance

With the growth of traffic flow and climate impacts exacerbating the problem of road defects, the limitations of traditional detection methods are highlighted. This study focuses on this dilemma and constructs a road defect detection method using deep learning techniques to provide a reference for intelligent urban maintenance. Due to the diversity of road defects, the dataset used in this paper covers a variety of defects such as cracks, potholes, etc., and road images under different lighting and weather conditions are collected, and image enhancement is carried out through contrast adjustment, rotation, random cropping and other methods. In terms of deep learning model construction, this paper compares a variety of deep learning models that are more applicable in the field of road defect recognition, and the dataset is divided into 7:3 for training. Experiments show that EfficientNet has the lowest loss value and its F1 index reaches 0.92, which has a significant advantage in recognition accuracy, and the confusion matrix shows that it is effective in classifying most of the defects, which provides effective support for intelligent road maintenance management.