Automatic detection of internal corrosion defect in a natural gas gathering pipeline using improved YOLOv5 model
Bingjie Chen, Li Ma, Shan Liang
Abstract
The detection of internal corrosion of the very long gas pipeline is a fundamental task for the prevention of possible failures, also one of the major challenges facing gas companies. Pipeline endoscopy based on the corrosion inspection method provides direct visual observation of defects, but the post-video analysis is time-consuming and not practical. In this work, we propose an improved YOLOv5 model-based software approach to automatically detect inner corrosion defects in nature gas-gathering pipelines. Video streaming of pipe interiors is provided by an endoscope robot. Sample augmentation strategies such as affine transformation and defect segmentation with background fusion are used to generate defect images and expand the data set. The region-based recursive localization method can be effectively optimized to improve the localization of corrosion regions.