Enhanced real-time detection transformer (RT-DETR) for robotic inspection of underwater bridge pier cracks

Abstract

The inadequate visual environment reduces the accuracy of underwater bridge pier fracture detection. Consequently, this paper suggests enhancing the backbone of the Real-Time Detection Transformer(RT-DETR) model to serve as the backbone of the YOLOv8 model. This will be achieved by substituting the Faster Implementation of CSP Bottleneck with 2 convolutions(C2f) module with the Poly Kernel Inception(PKI) Block, which is composed of the PKI Module and Context Anchor Attention(CAA) Block. Its strong capability to distinguish cracks and background features enables accurate recognition of underwater bridge pier cracks. To provide data for detecting these cracks, the enhanced Unpaired Image to Image Translation(CycleGAN) network converts land-style bridge crack images to underwater-style fracture images. The proposed model achieved an F1 score of 0.85 and a mAP50 of 0.84. The real-time detection of underwater bridge fractures by the underwater robot was facilitated by the FPS index of 87.47, which optimizes the detection efficiency.