Lightweight YOLOv8 Real-Time Object Detection via Progressive Pruning and Feature-Aware Knowledge Distillation

This study addresses the challenges of high computational complexity and large model size of YOLOv8 in edge device deployment by proposing a lightweight compression method combining progressive channel pruning and feature-aware knowledge distillation. The research first introduces the significance of real-time object detection in applications like intelligent transportation and security surveillance, highlighting YOLOv8's performance but its limited adaptability to edge devices due to high parameters (11.2M) and FLOPs (28.6G). To solve this, a progressive pruning strategy based on Batch Normalization (BN) layer scaling factors is developed, iteratively pruning and fine-tuning to reduce parameters by 20.5% and FLOPs by 21.7% while preserving critical shallow-layer features. Complementing this, feature-aware knowledge distillation is designed to recover accuracy loss from pruning by aligning student model features with teacher model patterns, restoring [email protected] from 0.738 (pruned model) to 0.748 (only 0.5% lower than the original). Experimental results validate the method's effectiveness in balancing efficiency and accuracy, with the distilled model achieving 165 FPS and 32.1 MB size. Feature map visualization further reveals pruning-induced feature sparsity and distillation-driven recovery, providing insights for optimization. Limitations, such as quantization's parameter ineffectiveness and complex scenario robustness, are also discussed, suggesting future improvements with dynamic pruning and multi-scale distillation.