Research on Small Target Detection for Drone Based on Improved YOLOv5

Due to their strong scalability, drones are widely used in military reconnaissance, agricultural monitoring, and industrial inspection, often encountering small, hard-to-detect targets. This paper proposes an improved YOLOv5 object detection algorithm that modifies the network's HEAD by replacing the original detection head with a smaller one, adding the Dyhead module, incorporating the CBAM spatial attention mechanism, and changing the post-processing method from NMS to SoftNMS. Experimental results on the VisDrone2019 dataset show an 11% increase in recognition accuracy compared to the original algorithm. Subsequently, the Lamp pruning method is applied to reduce model parameters and enhance frame rate, achieving 70% model compression and a 30% increase in inference speed. After knowledge distillation, the accuracy loss is only 1.7%.