Multi-Modal Pipeline Defect Localization

NeurIPS 2024 Workshop MusIML Submission15 Authors

Keywords: Object detection, Segmentation, Domain shift, Label noise, Pipeline

Abstract:

This study investigates the use of Laser and Magnetic Flux Leakage (MFL) pipeline
data to develop a deep learning model for accurate detection and segmentation of
pipeline defects. Laser images are used to precisely identify defect regions and
provide labels for training a Mask R-CNN model for localizing and segmenting
defects in MFL signals. Unlike conventional datasets where ground-truth labels
are pixel-wise accurate, our labels are derived from a different sensor modality,
resulting in misalignment and feature discrepancies between the laser and MFL
data. These discrepancies lead to label noise and domain shift. Our experiments
show that training advanced object detection and segmentation models using only
laser-derived labels does not achieve accurate defect localization in MFL signals.
This underscores the need for models capable of handling label discrepancies and
adapting across domains to ensure robust and scalable performance in real-world
pipeline defect detection.

Submission Number: 15