A Novel Insulator Defects Segmentation Network Integrating Dual-Branch Dynamic Snake Convolution Module and Decoupled-Selective Feature Pyramid Network

Insulator defect detection is critical for ensuring the safe and stable operation of power grids. However, existing methods still have certain limitations in terms of mask edge accuracy and overall detection performance. To address the challenges posed by complex shapes and significant scale variations in insulator defects, this paper proposes a network that integrates Dynamic Snake Convolution (DSConv) with an Decoupled-Selective Feature Pyramid Network (D-SFPN). First, a Dual-Branch Dynamic Snake Convolution Module (DB-DSCM) is designed, which combines DSConv with standard convolution in a synergistic manner to enhance the feature representation of crack and flashover defects. On this basis, a dual-branch feature extraction network is built to enhance defect feature representation. Second, a D-SFPN is proposed, which incorporates a Decoupled Information Enhancement Module (DIEM) and an Adaptive Feature Selection Module (AFSM). The DIEM enhances critical information and suppresses redundant background noise, while the AFSM adaptively optimizes semantic information. Finally, the bounding box loss function (Reinforced IoU, RIoU) is utilized to further improve the detection accuracy of insulator defect masks in terms of area deviation, center deviation, and shape deviation. The proposed method achieves a better balance between comprehensive detection accuracy and speed in insulator defect detection, demonstrating robust overall performance. Experiments show that the proposed method achieves an 11.16% improvement in overall bounding box average precision and a 17.81% improvement in mask average precision compared to the baseline.