Development of a Petrochemical Plant Pipe Counting Method Based on Improved Mask R-CNN and Transformer

Abstract

In petrochemical plant construction, pipe installation critically the affects overall progress and costs, making accurate on-site pipe quantification essential. However, the wide range of pipe diameters, high-density layouts, and complex site conditions poses significant challenges for automated counting methods. To address these issues, this study constructs the “PetroPipe” dataset, comprising 2,380 images (340 originals and 2,040 augmented) with pipes ranging from 0.75 to 60 inches in diameter under various resolutions. Multiple data augmentation strategies and a multi-stage training process enhance robustness under dynamic construction conditions. Building on this dataset, a pipe counting method is proposed that integrates instance segmentation with a counting head and a transformer encoder. Based on the Mask R-CNN framework, the model applies a discrete-bin classification approach to achieve accurate global quantity predictions, while maintaining high-quality instance-level detection and segmentation. The experimental results demonstrate that the developed model outperforms the standard Mask R-CNN model on the test set, increasing the counting accuracy from 65.55% without a counting head to 85.50%. When evaluated on an unlabelled 18-second video captured in a real-world scenario, it attained an accuracy of approximately 98.1%, demonstrating its capacity to adapt to practical industrial conditions. These results indicate that the proposed approach effectively accommodates varying pipe diameters, dense layouts, and complex backgrounds. Consequently, it provides a scalable, efficient pipe counting, progress monitoring, and resource optimization in petrochemical construction environments.