Lightweight YOLOv5 model based small target detection in power engineering

Abstract

Deep learning architectures have yielded a significant leap in target detection performance. However, the high cost of deep learning impedes real-world applications, especially for UAV and UGV platforms. Moreover, detecting small targets is still of lower accuracy in contrast to the large ones. Aiming to comprehensively handle these two issues, a novel SP-CBAM-YOLOv5 architecture is proposed. The main novelty of our hybrid model lies in the cooperation of the attention mechanism and the typical YOLOv5 architecture, which can largely improve the performance of the small target detection. Moreover, the depth convolution and knowledge distillation are jointly introduced for lightening the model architecture. To evaluate the performance of our proposed SP-CBAM-YOLOv5, we built a novel dataset containing challenging scenes of power engineering. Experimental results on this benchmark demonstrate that our proposed SP-CBAM-YOLOv5 achieves a competitive performance in contrast to the other YOLO architectures. Besides, our lightweight YOLOv5 has more than 70% decrease of parameters. Moreover, the ablation study is conducted to demonstrate the compact architecture of SP-CBAM-YOLOv5.