DSBR-CSNet: a cross-stage small targets fault detection in transmission lines based on dynamic sparse bi-level routing mechanism

Abstract To address the challenge of accurately detecting small target faults in drone-based transmission line inspections, this paper proposes dynamic sparse bi-level routing cross-stage network (DSBR-CSNet), a novel deep learning-based detection framework. First, to improve feature extraction capability under complex backgrounds, we introduce the Biformer visual transformer into the backbone network. This module utilizes a dynamic sparse attention mechanism with bilayer routing, which achieves content-aware computation and effectively captures long-range dependencies. Second, to alleviate the computational burden associated with standard and depthwise convolutions in cross-stage partial darknet, we design a group partial convolution module to serve as a more efficient alternative. Furthermore, the effective complete intersection over union loss function is adopted to enhance localization precision of small targets and accelerate model convergence. The experimental results show that DSBR-CSNet performs well in detecting small fault targets in complex environments. The detection accuracy reaches 94.7%, the recall rate is 93.1%, the mAP@.5 is 92.6%, and the mAP.5:.95 is 69.2%. In addition, the model has a running speed of 54.1 FPS for 640 × 640 input images, indicating its potential for real-time detection applications.