Enhanced and lightweight design of small object detector based on YOLOv5s model

Owing to the challenges of dense target distribution and complex backgrounds in small detection targets, existing small object detection algorithms suffer from poor performance and high model complexity, which is extremely difficult to deploy on embedded platforms. To address above issues, we optimized the YOLOv5s model structure to enhance detection accuracy. To avoid incurring extra computational expenses, we introduced a local pruning strategy to reduce redundancy, which enables the detection model more suitable for embedded systems. Considering pruning may cause accuracy degradation, we employ knowledge distillation techniques combining feature distillation and output distillation. Specifically, we transfer the knowledge from a high-precision teacher model to a student model, enabling exceptional real-time performance. The experimental results on the VisDrone2019 dataset show that compared to the original algorithm, our model has reduced the parameter count by 50.38%, computation by 51.81%, and model size by 52.94%, totaling just 8 M. The average precision (mAP@0.5) improved to 42.2%. Our proposed model outperforms the current state-of-the-art methods for small object detection in terms of both accuracy and computational efficiency.