HFQE-Det: frequency-enhanced multi-scale defect detection with confidence calibration for transmission line inspection

Abstract During the inspection of transmission lines using unmanned aerial vehicles, complex background interference and multi-scale defect characteristics often lead to missed or false detections, which may compromise the safe operation and maintenance of power grids. A multi-scale defect detection network based on frequency domain enhancement and confidence calibration, abbreviated as HFQE-Det, is proposed in this study to address the above challenge. The network leverages frequency domain decomposition to separate low-frequency structural features from high-frequency detail features, and incorporates a multi-scale edge enhancement mechanism to strengthen target boundaries and semantic context while suppressing background noise. Furthermore, by integrating dual-dimensional attention with a multi-scale contextual awareness structure, the network enables hierarchical fusion of global structural information and local detailed features, thereby improving robustness in detecting targets across varying scales. Additionally, by combining shared convolutional weights with position quality estimation, classification confidence is dynamically optimized through statistical characteristics of bounding box distributions, significantly enhancing localization accuracy for rare defects. To evaluate the detection performance of HFQE-Det, extensive experiments were conducted on a self-built power line defect dataset (PLDD-2024). This dataset contains 2483 images with a uniform resolution of 640 × 640, covering 12 common defect categories. The model was compared with 10 state-of-the-art detection methods, achieving a detection accuracy of 95.1%.