P-DETR: A transformer-based algorithm for pipeline structure detection

Pipelines are essential transportation infrastructure for oil and gas, but they are vulnerable to defects such as cracks, joint failure, and corrosion due to extreme weather conditions. These defects can result in oil and gas leakage, which prompts environmental and economic damages. Hence, regular inspection of pipelines is necessary. The industry has increasingly relied on using drones for pipeline inspections, though the inspection is still done manually by the drone operator or offline via recorded video footage from the drone. This paper proposes using the Pipe Detection Transformer (P-DETR), a novel transformer-based model designed for pipeline detection and potential integration with aerial robots or drones to enable autonomous pipeline inspection. P-DETR introduces significant improvements to the original Detection Transformer (DETR) framework to enhance its detection performance, particularly for small-sized pipes - a key limitation of the baseline DETR. The major contribution is a Feature Normalization and Transformation (FNT) module, which fuses multiple layers of the convolutional backbone to provide a focused representation of small-sized features before processing by the transformer module. Experimental results validate the superiority of P-DETR, achieving an overall mAP of 55 %, a 3 AP improvement over DETR, and significantly increasing precision for small-sized pipe detection by 8.6 AP (from 1.9 to 10.5). Additionally, precision improvements for medium- and large-sized pipes were 10.8 AP (from 10.8 to 21.6) and 2.2AP (from 64.4 to 66.6), respectively, with an overall recall of 73.9 %, a 4 AP improved performance over DETR. The results from extensive experiments highlight the superior performance of the proposed P-DETR model over the original DETR, UP-DETR, R-DETR, Skip-DETR, and other standard object detection models, including YOLOv3 and SSD.