Edge-Guided Network for Salient Object Detection in Substation Equipment Defects

With the rapid development of computer vision, defect detection leveraging this technology has become an increasingly prominent trend. However, due to the interference from complex substation environments, traditional networks exhibit certain limitations in edge detection and defect information extraction. To address these challenges, this paper proposes the Edge-Guided Attention Network (EGANet), which consists of the Feature Extraction Module (FEM), the Edge Detection Module (EDM), and the Edge-Guided Attention Module (EGM). The FEM employs dilated convolutions with varying receptive fields to capture multi-scale contextual information. The EDM utilizes the Sobel operator, guided by high-level semantic features, to extract edge information from defect features. Meanwhile, the EGM integrates edge features and multi-scale features to further enhance critical edge information while suppressing irrelevant small structures. Experimental results demonstrate that EGANet performs exceptionally well in substation equipment defect detection tasks, surpassing state-of-the-art methods in terms of accuracy, generalization capability, and robustness.