饮料包装缺陷检测的轻量化算法研究  

Research on lightweight algorithms for beverage packaging defect detection

作　　者：付赫 王桂英[1] FU He;WANG Guiying(College of Home and Art Design,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学家居与艺术设计学院,哈尔滨150040

出　　处：《包装与食品机械》2025年第1期32-39,共8页Packaging and Food Machinery

基　　金：国家自然科学基金项目(32071692)。

摘　　要：针对饮料包装生产线上缺陷检测精度低和速度慢等问题,提出一种基于改进YOLOv8的轻量化饮料包装缺陷检测算法(Light YOLOv8-DP),以提升小目标检测精度,降低计算资源消耗。在YOLOv8的Backbone骨干网络中引入RepStreamGhost(RSG)模块以替代C2f模块,实现梯度流通优化;将YOLOv8的Neck特征融合网络替换为Small Target Boost Pyramid(STBP)结构,通过AdaptiveOK(AOK)模块增强特征提取的多尺度表现;采用重参数轻量化共享卷积的Rep Shared Convolutional Detection(RSCD)检测头,减少参数量和计算复杂度;通过构建饮料包装缺陷数据集进行试验验证。结果表明,Light YOLOv8-DP算法平均精度达到85.5%,召回率达到82.1%,精确率达到83.8%,较原始YOLOv8分别提高2.3%,3.3%,3.9%;检测速度达到259.26帧/s,较原始YOLOv8提高43.65帧/s。对改进后的算法进行实时检测验证,F1分数提高1.7,单张图片的处理速度提高0.7 ms。研究为食品包装的自动化检测提供新思路。To address the challenges of low accuracy and slow inference speed in defect detection on beverage packaging production lines,this study proposes a lightweight defect detection algorithm named Light YOLOv8-DP,an improved YOLOv8-based framework.The algorithm enhances small object detection accuracy while reducing computational resource consumption.Specifically,in the YOLOv8 backbone network,the original C2f module is replaced with the RepStreamGhost(RSG)module to optimize gradient propagation.The neck network is redesigned as a Small Target Boost Pyramid(STBP)structure integrated with an AdaptiveOK(AOK)module to strengthen multi-scale feature representation.Additionally,a Reparameterized Shared Convolutional Detection(RSCD)head is introduced to minimize parameter size and computational complexity.Experimental validation was performed on a custom-built beverage packaging defect dataset.Results show that Light YOLOv8-DP achieves a mean average precision(mAP)of 85.5%,recall of 82.1%,and precision of 83.8%,representing improvements of 2.3%,3.3%,and 3.9%,respectively,over the baseline YOLOv8.The inference speed reaches 259.26 FPS(frames per second),43.65 FPS faster than the original model.Real-time validation further demonstrates a 1.7-point increase in F1-score and a 0.7 ms reduction in per-image processing latency.This research provides a novel approach for automated quality inspection in food packaging systems.

关 键 词：YOLOv8 缺陷检测 RSG模块 AOK模块 RSCD检测头 

分 类 号：TS203[轻工技术与工程—食品科学] TH165.4[轻工技术与工程—食品科学与工程] TH487[机械工程—机械制造及自动化]