An HWE-Net Network for Insulator Detection in Overhead Transmission Lines

The inspection of insulators in high-voltage transmission lines is essential for ensuring safety; however, detection accuracy may be compromised by complex backgrounds, surrounding objects, and partial occlusions. To address these challenges, we propose a novel network, the Haar Wave Feature Efficient Transformer Network (HWE-Net). This network incorporates the Fusion Wave Backbone, which integrates high-frequency features derived from the Haar Wavelet Transform with the original features, thereby enhancing edge and detail information to improve target detection accuracy. Furthermore, we introduce an efficient encoder model (EE) characterized by a sandwich structure that optimizes feature channels and facilitates the transfer of multi-scale feature information, thereby enhancing the detection of small objects. Additionally, we incorporate region-specific Self-Directed Attention to improve multi-scale information fusion and filter redundant information, allowing for dynamic adaptation to background interference. We have also developed a new dataset comprising insulator samples under various backgrounds and occlusions for evaluation purposes. Experimental results indicate that the proposed method achieves an Average Precision at 50% Intersection over Union (AP50) of 98.32% and an F1 score of 0.98, surpassing the performance of more complex models in detection efficacy.