In the intricate and dynamic industrial landscape of oil and gas pipeline transportation, crack detection faces significant challenges. Crack morphology varies unpredictably, while image quality is constantly compromised by various interferences—ranging from corrosive residues and background noise to misleading textural similarities—making it difficult for traditional methods to achieve a satisfactory balance between accuracy and efficiency. To address this pressing challenge, advanced object detection algorithms characterized by structural flexibility, robust performance, and exceptional adaptability to crack detection tasks have become a research focus. Such algorithms are particularly well‐suited to the rigorous demands of pipeline inspection, and their refined, customizable architectures can be seamlessly integrated into practical engineering scenarios, laying a solid foundation for both innovative model improvements and enhanced research applicability. Leveraging these advantages, this study proposes a novel detection framework—DSM, which strikes a balance between computational efficiency and improved accuracy. DSM is an integrated abbreviation for two key modules: the multikernel denoising pooling block (MKDP) and the deformable spatial strip attention block (DSSA). The framework ingeniously combines these specialized components—where MKDP is designed to meticulously filter out irrelevant background noise, and DSSA aims to extract both global and local contextual information—significantly enhancing detection precision in complex environments. Experimental results highlight the effectiveness of the proposed framework, achieving 64.0% mAP@0.5, which represents a 6.8% improvement over the YOLOv8‐n baseline on a real‐world pipeline crack dataset. This improvement significantly enhances the reliability of defect identification in complex industrial environments. Moreover, DSM‐Det maintains a lightweight structure with only 5.8M parameters, achieving an excellent balance between accuracy and computational efficiency. These advantages make the framework particularly suitable for large‐scale, real‐time pipeline inspection applications, demonstrating strong practicaland academic value.
Yang et al. (Sun,) studied this question.