UAV Imagery Based Fine-Grained Defect Recognition System of Noise Barrier Fasteners in High-Speed Railroad

Noise barriers along high-speed railroads have the function of blocking the high-pressure airflow generated by the passage of high-speed trains and preventing the intrusion of foreign objects. However, under the influence of complex operation environment and natural factors, along with the increase of service time of noise barriers, various defects generated by noise barriers pose a threat to the safe operation of railroads, especially the defects of fasteners that are used for fixing noise barrier bases. In addition, classical learning-based deep networks are difficult to make decisions on fine-grained fastener defect categories. This paper proposes a fine-grained defect recognition system of noise barrier fasteners using Region Shuffling and Position Regression Network (RPNet). First, RPNet leverages tailored shuffling rule to achieve destructive fastener images, which can assist the RPNet to focus on the local vital details. Second, a position regression structure integrated in RPNet are employed to learn and regress the original location distribution after feature extraction of RepVGG-based backbone. Third, based on dual-branch fully connection structure, RPNet enables the decision-making of subtle defect categories. Finally, the comprehensive experiment results represent that proposed system can acquire a faster inference speed yet maintains a higher detection rate in the complex UAV -based noise barrier dataset.