Abstract
Efficient deployment of deep learning models for aerial object detection on resource-constrained devices requires significant compression without com-promising performance. In this study, we propose a novel three-stage compression pipeline for the YOLOv8 object detection model, integrating sparsity-aware training, structured channel pruning, and Channel-Wise Knowledge Distillation (CWD). First, sparsity-aware training introduces dynamic sparsity during model optimization, effectively balancing parameter reduction and detection accuracy. Second, we apply structured channel pruning by leveraging batch normalization scaling factors to eliminate redundant channels, significantly reducing model size and computational complexity. Finally, to mitigate the accuracy drop caused by pruning, we employ CWD to transfer knowledge from the original model, using an adjustable temperature and loss weighting scheme tailored for small and medium object detection. Extensive experiments on the VisDrone dataset demonstrate the effectiveness of our approach across multiple YOLOv8 variants. For YOLOv8m, our method reduces model parameters from 25.85M to 6.85M (a 73.51% reduction), FLOPs from 49.6G to 13.3G, and MACs from 101G to 34.5G, while reducing AP50 by only 2.7%. The resulting compressed model achieves 47.9 AP50 and boosts inference speed from 26 FPS (YOLOv8m baseline) to 45 FPS, enabling real-time deployment on edge devices. We further apply TensorRT as a lightweight optimization step. While this introduces a minor drop in AP50 (from 47.9 to 47.6), it significantly improves inference speed from 45 to 68 FPS, demonstrating the practicality of our approach for high-throughput, re-source-constrained scenarios.
Abstract (translated)
在资源受限设备上高效部署深度学习模型以进行航空目标检测,需要大幅压缩模型而不会牺牲性能。本文提出了一种新颖的三阶段压缩管道,用于YOLOv8对象检测模型,并整合了稀疏感知训练、结构化通道裁剪以及通道级知识蒸馏(CWD)。首先,稀疏感知训练在模型优化过程中引入动态稀疏性,有效地平衡参数减少和检测精度。其次,我们通过利用批归一化缩放因子来消除冗余通道,从而显著减小模型大小和计算复杂度来进行结构化通道裁剪。最后,为了缓解修剪引起的准确率下降,我们采用CWD从原始模型转移知识,使用可调温度和损失加权方案,专为小中型对象检测设计。 在VisDrone数据集上进行的大量实验验证了我们的方法对多个YOLOv8变体的有效性。对于YOLOv8m版本,我们的方法将模型参数减少了73.51%,从25.85M降至6.85M,并且将FLOPs(浮点运算次数)从49.6G减少到13.3G,MACs(乘加操作次数)从101G减少到34.5G,仅使AP50精度降低2.7%。压缩后的模型达到47.9 AP50,并将推断速度从基线YOLOv8m的每秒26帧提升至45帧,实现了在边缘设备上的实时部署。 此外,我们进一步应用了TensorRT作为轻量级优化步骤。虽然这导致AP50略微下降(从47.9降至47.6),但将推断速度显著提升到每秒68帧,展示出我们的方法对高吞吐量、资源受限场景的实用性。
URL
https://arxiv.org/abs/2509.12918