Adaptive Focusing Multi‐Scale Feature Network for Pinning Defect Detection in Transmission Lines

ABSTRACT With the development of smart grids, transmission line UAV intelligent inspection technology has been widely used. Pin defect detection is a common task in the intelligent inspection process, but due to the small size of the pin bolts in the inspection image, it is difficult for the existing detection algorithms to accurately recognise the pin defects in the complex background. In this paper, an Adaptive Focusing Multi‐Scale Feature Network (AFMFNet) is proposed. First, the Path‐Interleaved Deformation Convolution (PIDC) is proposed to further enhance the feature extraction ability for the irregular pose of the pin bolt. Second, the Small Target Enhanced Pyramid (STEP) is constructed. It realises the effective fusion of multi‐scale features of small targets through the differentiated processing between different layers and the global perception capability granted by CSP_OmniKernel. Finally, the improved Wise‐MPDIoU loss function is utilised to improve the convergence speed and regression accuracy of the model. AFMFNet enhances detection accuracy for normal and defective pins by 7.5% and 13.4%, respectively compared to the baseline model, achieving a reasoning speed of 141.2 f/s on PC, meeting real‐time detection needs. Its robustness is verified in complex scenarios, offering a new intelligent approach for transmission line inspection.