Real‐Time Efficient Detection of Substation Electrical Equipment in Infrared Images Based on Improved YOLOv11
Runtong Zhu, Hu Liu, Yanhao Gao, Yan Yue · IET Image Processing · 2025
ABSTRACT To address the challenges in automatic infrared image detection within complex substation environments, including missed detection of small targets, severe interference from thermal backgrounds, and balancing detection accuracy with efficiency–this paper proposes an intelligent detection method for electrical equipment based on an improved lightweight YOLOv11n architecture. Specifically, we propose a novel sparse attention and convolution mixing (SACM) module that integrates sparse attention with convolutional operations, aiming to enhance small object detection capability and feature selection efficiency while maintaining low computational complexity. Furthermore, the backbone network is augmented with the self‐calibrated convolutions (SCConv) module, which performs structured feature reconstruction to suppress redundant information and improve computational efficiency, thereby providing a low‐noise foundation for subsequent multi‐scale feature fusion. In addition, the original PANet is replaced with bidirectional weighted feature pyramid network (BiFPN) to construct an efficient fusion pathway that aligns with the SCConv‐enhanced features, thus strengthening cross‐scale feature interaction. A lightweight up sampling module, CARAFE, is also introduced and combined with SACM to establish a synergy between feature enhancement and fine‐grained reconstruction, effectively recovering detailed features of small objects and reducing missed detections. Finally, the shape‐intersection over union (Shape‐IoU) loss function is adopted in place of traditional bounding box regression losses to further improve localization accuracy. Experimental results demonstrate significant performance improvements in complex substation scenarios: the proposed method achieves an mAP@50 of 97.0%, representing a 3.2 percentage point increase over the baseline model, while maintaining low computational complexity with only 2.82 M parameters and 6.1G FLOPs. This study provides an innovative solution that balances high accuracy and efficiency for intelligent electrical equipment detection, offering both theoretical significance and practical engineering value for advancing intelligent substation operation and maintenance.