Towards a Lightweight Object Detection through Model Pruning Approaches

Object detection tasks represent one of the most prevalent areas of study in computer vision, leading to the introduction of numerous techniques. Among these, the You Only Look Once (YOLO) series of object detection models continued to evolve and progress. The latest iterations within the YOLO family exhibit enhanced performance and quicker inference times. However, the increased capacities and memory demands of these models present real-world challenges in terms of practical deployment. This underscores the importance of developing lightweight versions of the updated YOLO models to ensure their applicability in real-life scenarios. In this context, this study introduces YOLOv7 lightweight, building upon a prior channel pruning technique employed for YOLOv5. By adopting the foundational method to align with the YOLOv7 architecture, we effectively managed to reduce the model's complexity. Furthermore, this research delves into identifying the appropriate pruning levels and model configurations tailored specifically for human detection tasks. In the course of our investigation, we evaluated the trade-off between performance degradation and reductions in parameters and computational complexity. This analysis led us to select a pruning protection ratio of 50% as the most optimal value. Moreover, this article presents the optimization of the lightweight YOLOv7 model for efficient human detection. In essence, our research not only suggests enhancements to existing methodologies for updated models but also emphasizes the practical application of such methods through a comprehensive grasp of the unique characteristics of updated models.