Objective Pipeline safety monitoring is evolving from post-event inspection to real-time early warning, demanding faster and more accurate anomaly localization. Image Salient Object Detection (ISOD) effectively identifies key image regions and holds significant application potential but faces challenges like high annotation costs, complex object shapes, and sample imbalance. In oil and gas pipeline inspection, pixel-level ground-truth annotations are costly and scarce, with limited public datasets. Weakly Supervised Image Salient Object Detection (WSISOD) addresses these issues by learning salient regions from weak annotations, reducing annotation costs and enhancing cross-scene generalization.

Methods A weakly supervised image salient object detection method for oil and gas pipeline inspection was proposed. First, considering the confidentiality of oil and gas pipelines and the lack of public datasets, the Remote Sensing Oil and Gas Pipeline Salient Object Detection Dataset (RSOGP-SODD), containing 1,000 images, was constructed. Second, based on the Segment Anything Model 2 (SAM2), a multi-box joint inference mechanism was adopted to convert low-cost bounding box annotations into high-quality pseudo-labels. Finally, the Heterogeneous Feature Collaboration Network (HFCNet), optimized for remote sensing images, was used as the backbone network, and iterative optimization of pseudo-labels was employed to improve model accuracy.

Results Experimental results on the RSOGP-SODD dataset demonstrated that, without pixel-level ground-truth annotations, the proposed method achieved S-measure and E-measure scores of 0.749 1 and 0.830 9, respectively. Compared to fully supervised methods, the weakly supervised model showed only about a 14％ performance decrease while significantly reducing manual annotation costs. Additionally, the model exhibited robust structure preservation and boundary detection capabilities in complex remote sensing scenes.

Conclusion The proposed weakly supervised image salient object detection method offers a feasible, cost-effective, and efficient solution for intelligent oil and gas pipeline inspection. Future work can focus on enhancing the boundary detail quality of pseudo-labels, integrating multimodal data and active learning to improve model generalization across complex geographic and multi-pipeline scenes, and extending the framework to practical engineering applications such as pipeline defect detection and leakage monitoring.