EFENet: An Efficient Edge-Feature Enhanced Network for power insulator detection

Accurate detection of high-voltage transmission line insulators is crucial for ensuring the safe and stable operation of power systems. However, traditional machine vision detection techniques face limitations in detection accuracy under complex backgrounds and strong interference from visually similar objects. To address this challenge, this paper proposes an Edge-Feature-Enhanced Transformer Network (EFENet), which achieves precise insulator detection through an Edge-Frequency-Multi-Scale Feature Fusion Framework. Initially, a Dual path Edge detection Branching Convolution network (DEBC) is constructed to enhance the edge feature extraction capability of ResNet50. This module synergistically leverages ResNet50 and the canny edge detector to effectively capture subtle structural information of target edges, enabling precise localization of insulator contours. Subsequently, an FFT Stage Fusion Module (FFT) is designed. This module deeply integrates the edge features extracted by the dual-path backbone network with the features output by the ResNet50 network within the frequency domain, effectively suppressing background noise interference on edge features and significantly enhancing the model’s robustness in complex backgrounds. Furthermore, to fully exploit the multi-scale features generated by frequency-domain fusion, an Efficient Encoder (EE) is proposed. This module optimizes feature representation in complex scenes by facilitating efficient multi-scale information interaction and feature fusion across channels, thereby improving the model’s generalization performance across diverse detection scenarios. To ensure comprehensive evaluation, we create a specialized dataset featuring diverse insulator samples under varied background complexities and occlusion scenarios. Extensive experiments demonstrate EFENet’s superior precision and robustness, achieving an Average Precision at 50% IoU ( A P 50 ) of 98.49% and an F1-score of 0.98. These results surpass state-of-the-art plan, confirming its efficacy for intelligent power equipment inspection. • Proposes a Dual-path Edge Branching ConvNet (DEBC) with an Edge Enhanced Module for robust insulator recognition. • An FFT Stage Fusion Module merges edge/semantic features in the frequency domain to suppress background noise. • An Efficient Encoder enables multi-scale interaction and cross-channel fusion for better generalization. • A novel high-quality insulator dataset from diverse scenes supports training for complex real-world tasks.