EPDNet: Light-weight small target detection algorithm based on pruning and logical distillation

Abstract

Drone detection technology plays a crucial role in various fields. However, due to the limited computational resources of edge devices onboard drones, achieving efficient detection using large-parameter algorithms remains challenging. Small target detection in drone-based applications faces several difficulties, including the small size of targets, limited feature information, and vulnerability to environmental interference. Moreover, existing lightweight small target detection methods often compromise detection accuracy while reducing model parameters, failing to meet the dual requirements of accuracy and efficiency in drone scenarios. To address these challenges, this paper proposes EPDNet, a lightweight small target detection algorithm designed for drone applications. First, ConvNextV2 replaces the original backbone network, incorporating a fully convolutional masked autoencoder framework combined with a self-supervised learning strategy to enhance the extraction of essential low-level features. Additionally, the EC2f feature extraction module is introduced to enable interactive modeling of contextual detail features across different target scales, orientations, and shapes. Furthermore, an adaptive channel pruning scheme is designed to reduce redundant parameters and computational complexity, thereby enhancing algorithm efficiency. Finally, the detection performance of the pruned model is further optimized using knowledge distillation. Experimental results on the VisDrone2019 aerial photography dataset demonstrate that EPDNet improves detection precision (P) by 2.6%, increases mean average precision (mAP) by 3.0%, reduces the number of parameters by 29.6%, and decreases computational cost by 17.8%. These results indicate that EPDNet effectively meets the lightweight deployment requirements of drone-based applications.