PGI-Net : Using Programmable Gradient Information and Group Lasso Network Slimming for object detection of UAV images

This paper addresses object detection in unmanned aerial vehicle (UAV) imagery by proposing the PGI-Net model and the Group Lasso channel pruning method to improve performance and efficiency. Existing methods often struggle with detecting small objects and adapting to resource-constrained environments. PGI-Net enhances detection by integrating Programmable Gradient Information (PGI) into YOLOv8, employing a main branch and an auxiliary invertible branch for collaborative learning. Meanwhile, Group Lasso pruning applies grouped L1 and L2 regularization to effectively reduce model size and computational cost. Experiments on the SeaDronesSee dataset show that PGI-Net achieves improvements of 2.74% in mean average precision (mAP) at 0.5 and 2.31% at 0.5:0.95. After pruning, the model maintains detection performance while significantly reducing computational overhead, demonstrating its suitability for resource-limited UAV applications.