摘要
针对酒品包装检测算法在质检任务中存在的精度低、速度慢、复杂度高导致部署性不佳等问题,提出一种基于YOLOv8n改进的轻量化酒品包装检测算法。在主干网络的SPPF模块中增加平均池化支路,对EMA注意力机制进行结构补强,扩增可变核大尺度卷积支路,并将其嵌入到SPPF模块中作为输出端;结合ADown和HWD2种下采样方法,作为新的下采样模块以减少冗余参数,并保持更丰富的特征信息,加强模型特征表达能力;采用卷积权重共享策略,对检测头进行轻量化改进,并结合深度可分离卷积和分组卷积的模块组合,进一步降低模型复杂度;采用Focaler-PIoU损失函数优化定位损失,加速算法收敛。采用自制酒品包装数据集进行训练和验证,并在阿里云天池的瓶装白酒瑕疵品公开数据集上进行泛化性测试。试验结果表明,相较于基准模型YOLOv8n,改进算法在mAP50,mAP50-95上,分别提高3.5,4.8个百分点,在参数量和计算量上,分别降低33.3%,37.0%;在瓶装白酒瑕疵品公开数据集上,改进算法在mAP50,mAP50-95上分别提高1.5,1.1个百分点,验证改进算法的泛化性。研究为酒品包装的质量检测提供理论支持。
To address the issues of low accuracy,slow speed,and high complexity leading to poor deployability of liquor packaging detection algorithms in quality inspection tasks,an improved lightweight algorithm based on YOLOv8n is proposed.An average pooling branch was added to the SPPF module in the backbone network,the structure of the EMA attention mechanism was reinforced,and a large-scale convolutional branch with variable kernels was expanded and embedded into the SPPF module as the output.Combining ADown and HWD downsampling methods,a new downsampling module was designed to reduce redundant parameters while preserving richer feature information,enhancing the model's feature representation capability.A convolutional weight-sharing strategy was adopted to lightweight the detection head,and a module combining depthwise separable convolution and grouped convolution was used to further reduce model complexity.The Focaler-PIoU loss function was employed to optimize localization loss and accelerate algorithm convergence.A self-made liquor packaging dataset was used for training and validation,and generalization testing was conducted on the publicly available Alibaba Cloud Tianchi dataset of flawed bottled liquor.Experimental results show that compared to the baseline YOLOv8n model,the improved algorithm increased mAP50 and mAP50-95 by 3.5 and 4.8 percentage points,respectively,while reducing parameter count and computational load by 33.3%and 37.0%.On the publicly available flawed bottled liquor dataset,the improved algorithm increased mAP50 and mAP50-95 by 1.5 and 1.1 percentage points,respectively,demonstrating its strong generalization capability.This research provides theoretical support for quality inspection of liquor packaging.
作者
向硕
曾水玲
贺刚健
林方聪
XIANG Shuo;ZENG Shuiling;HE Gangjian;LIN Fangcong(School of Communication and Electronic Engineering,Jishou University,Jishou 416000,China)
出处
《包装与食品机械》
北大核心
2025年第4期1-12,共12页
Packaging and Food Machinery
基金
国家自然科学基金项目(61966014)
湖南省自然科学基金项目(2024JJ7413)
湖南省研究生科研创新项目(QL20230255,CX20221107)
吉首大学科研项目(JGY2023071,JDX202409,JDX202420)。
作者简介
向硕(1999),男,硕士研究生,研究方向为计算机视觉,E-mail:764291127@qq.com;通信作者:曾水玲(1975),女,博士,教授,研究方向为模式识别,通信地址:416000湖南省吉首市人民南路120号吉首大学,E-mail:zengflsl@163.com。
