An Improved Swin Transformer-Based Method for Electric Spark Detection in Distribution Network Lines

Spark detection in distribution line videos remains challenging due to weak visual saliency, multi-scale variation, and complex environmental interference. To address these issues, this paper presents an enhanced Swin Transformer architecture for video-based recognition of electric sparks. The proposed model incorporates a Batch-Channel Normalization Transformer module to enhance sensitivity to small-scale targets while maintaining network stability. In addition, a Hybrid-Scale Multi-level Feature Pyramid module is introduced to strengthen multi-scale semantic representation through cross-layer feature fusion and atrous convolution. The resulting model achieves a favorable balance between detection accuracy and computational efficiency. Experimental results on a real-world transmission line dataset demonstrate that the proposed method outperforms the original Swin-T baseline by 2.2% in accuracy.