Development of a feature differentiation fusion semantic segmentation model (DFF-SSN) for robust substation visual perception

To address the challenges of equipment density, extreme lighting variations, and small target recognition in substations, this paper proposes a feature differentiation fusion infrared-visible semantic segmentation network DFF-SSN (Differentiated Future Fusion Semantic Segmentation Network). Based on a dual-encoder architecture with ResNet-50 as the backbone, the network incorporates a shallow fusion module SFM (Shallow Fusion Module) and a deep fusion module DFM (Deep Fusion Module) to achieve differentiated fusion of cross-modal features through spatial and channel attention, respectively. Combined with a Global Information Module (GIM) and an Information Aggregation Module (IAM), it achieves multi-scale context modeling and fine-grained boundary prediction, while mitigating class imbalance through weighted cross-entropy loss. On a self-built infrared-visible light dataset of transmission lines and substations, DFF-SSN outperforms algorithms like BiSeNet, LASNet, and RTFNet in both mIoU and mAcc metrics. It achieves up to 6.5% improvement for small object categories and maintains stable performance under complex lighting and occlusion conditions. This method provides a high-precision visual perception solution for intelligent substation inspection.