Ghost-enhanced pruning for lower training cost and faster inference in YOLOv8-based nameplate detection

This study investigates Ghost module enhanced pruning on YOLOv8n (nano) for nameplate detection, aiming to reduce model redundancy and the computational cost of pruning. Standard convolutional and bottleneck layers were replaced with GhostConv and GhostBottleneck modules, effectively lowering parameters and FLOPs prior to pruning. Experimental results show that Ghost-enhanced models preserve competitive detection accuracy, precision, and recall across various pruning ratios, while the primary tradeoff occurs in bounding box quality (AP metrics). Integration of Ghost modules allows up to a 50% reduction in required pruning to achieve comparable model efficiency. These findings demonstrate that combining Ghost module integration with iterative pruning provides an efficient pipeline for compressing YOLOv8 models while maintaining strong detection performance, offering a practical approach for lightweight object detection in resource-constrained settings.