YOLOv8-Seg with Dynamic Multi-Kernel Learning for Infrared Gas Leak Segmentation: A Weakly Supervised Approach

Area of Science:

• Industrial Safety
• Computer Vision
• Infrared Imaging

Background:

• Gas leak detection is crucial for safety in oil and gas facilities.
• Infrared imaging offers real-time, non-contact monitoring but faces segmentation challenges.
• Existing pixel-level networks struggle with accuracy, rough edges, and jagged boundaries.

Purpose of the Study:

• To develop a novel pixel-level segmentation network training framework for enhanced gas detection.
• To improve the segmentation performance of the YOLOv8-seg network for infrared gas leak applications.
• To introduce a low-cost, weakly supervised learning approach for training segmentation networks.

Main Methods:

• Introduced a dynamic threshold using Visual Background Extractor (ViBe) with YOLOv8-det to generate training masks.
• Designed a new segmentation head with dynamic kernels, multi-branch collaboration, deformable convolution, and attention mechanisms.
• Implemented a joint Dice-BCE loss weighted by ViBe-CRF confidence to refine gas edges.

Main Results:

• Achieved a 6.4% increase in F1 score and a 7.6% improvement in mean Intersection over Union (mIoU).
• Significantly reduced roughness and jaggedness at gas edges, enhancing segmentation accuracy.
• Demonstrated a new, efficient, real-time detection algorithm for infrared gas leak imaging.

Conclusions:

• The proposed framework effectively enhances YOLOv8-seg for gas leak detection in oil and gas facilities.
• The method provides a more accurate and robust solution for real-time infrared gas leak monitoring.
• Introduced a valuable weakly supervised learning strategy for training pixel-level segmentation networks cost-effectively.