A Small-Sized Defect Detection Method for Overhead Transmission Lines Based on Convolutional Neural Networks

Insulator defect detection is essential to the reliable operation of overhead transmission lines. However, current automatic algorithms struggle to extract critical features due to the small size of insulator defects in inspection images, which may lead to potential failures. To address this issue, this article proposes a novel and high-accuracy defect detection method based on deep learning technology, named insulator defect detection network (I2D-Net), which incorporates several innovative modules. First, we design a three-path feature fusion network (TFFN) to improve the network's ability to extract features from shallow layers. This hierarchical feature fusion mechanism across different network layers preserves spatial and semantic information, thereby maintaining the quality of features at different levels of the pyramid. Second, an enhanced receptive field attention (RFA+) block is incorporated to enable the network to adapt to different-scale defects and effectively distinguish them from the background. Finally, the context perception module (CPM) is introduced to better understand the surrounding features and their relationship with the defects. This improves defect localization capacity in the presence of interfering factors. Experimental results on the transmission line dataset demonstrate that the proposed method can accurately detect insulator defects and electrical components, even in challenging scenarios.