Abstract: Objective With the rapid development of long-distance oil and gas pipeline construction and the continuous expansion of pipeline network scale, pipeline safety monitoring is gradually shifting from post-event inspection to real-time early warning, posing higher requirements for rapid and accurate anomaly localization. Image Salient Object Detection (ISOD) is capable of efficiently identifying key regions in images and naturally meets the demand for fast detection of critical objects in pipeline monitoring scenarios, such as leakage, damage, and foreign object intrusion. However, practical applications of ISOD are often challenged by high annotation costs, complex object morphologies, and sample imbalance. Methods In oil and gas pipeline inspection tasks, pixel-level fine-grained annotations are costly and difficult to obtain, while publicly available datasets remain scarce. Weakly Supervised Image Salient Object Detection (WSISOD) enables the learning of salient regions under weak annotation conditions, offering the potential to reduce annotation costs and improve cross-scene generalization. To this end, this paper proposes a weakly supervised salient object detection method tailored for oil and gas pipeline inspection. Considering the confidentiality constraints and the lack of public datasets in this domain, we first construct a remote sensing image oil and gas pipeline salient object detection dataset containing 1, 000 images, named Remote Sensing Oil and Gas Pipeline Salient Object Detection Dataset (RSOGP-SODD). Subsequently, leveraging the segmentation foundation model Segment Anything Model V2 (SAM2), a multi-box joint inference strategy is adopted to transform low-cost bounding box annotation into high-quality pseudo labels. Finally, the Heterogeneous Feature Collaboration Network (HFCNet), optimized for remote sensing imagery, is employed as the backbone network, and an iterative pseudo label refinement strategy is introduced to further enhance detection performance. Results Experimental results on the RSOGP-SODD demonstrate that, without using pixel-level annotations, the S-measure of proposed method is 0.7491 and the E-measure is 0.8309. Compared with fully supervised methods, the weakly supervised model incurs only approximately 14% performance degradation while significantly reducing manual annotation costs. Moreover, it exhibits strong structural preservation and boundary perception capabilities in complex remote sensing scenes. Conclusion A low-cost and high-efficiency solution for intelligent oil and gas pipeline inspection is proposed in this paper. Future work will focus on improving the quality of pseudo labels, incorporating multimodal information or active learning mechanisms to enhance generalization across complex geographical environments and diverse pipeline scenarios, and extending the proposed framework to engineering applications with higher practical value, such as pipeline defect identification and leakage monitoring.