摘要
针对面点预制菜速冻水饺外观质量检测中细微露馅人眼难以辨别、露馅颜色特征不明显、正常褶皱与露馅存在相似性等因素导致的人工质检容易误判和漏检等问题,提出一种基于改进YOLOv8的用于面点预制菜流水线质检场景的外观质量检测算法CL-YOLO,该算法在backbone阶段引入2个新的相互关联的卷积注意力模块,在保证模型轻量化的同时加强对饺子特征中的各个维度进行权重分配与学习,使其更好地提取卷积中针对饺子露馅特征的信息。试验结果表明,相较于YOLOv8n,CL-YOLO在流水线饺子识别任务中mAP精度提升0.9%,召回率提升1.8%,算法的检测速度达到159FPS,模型大小及参数量均能够满足边缘端部署需求;在EUM-DET数据集中,mAP精度提升1.8%,召回率提升0.5%。该算法为食品流水线质检场景提供了新的实现思路。
In response to the challenges encountered in the visual quality inspection of frozen dumpling preprocessed foods,such as the difficulty for the human eye to discern minor stuffing leaks,the indistinct color characteristics of the leaks,and the similarity between normal folds and leaks leading to frequent misjudgments and omissions in manual inspections,a novel appearance quality detection algorithm for conveyor line quality inspection in the pre-processed food industry,named"CL-YOLO",based on an improved yolov8,was proposed.In the backbone stage of this algorithm,two new interconnected convolutional attention modules were introduced.These modules,while ensuring model lightness,enhanced the weight distribution and learning across various dimensions of dumpling features,thereby improving the extraction of convolutional information specific to dumpling stuffing leaks.Prior to entering the neck network,a triplet attention module further enabled the feature map to stack more effectively with the neck network,intensifying the learning of dumpling characteristics.Experimental results demonstrated that compared to YOLOv8n,CL-YOLO achieved a 0.9%increase in mAP accuracy and a 1.8%increase in recall rate in the conveyor line dumpling recognition task.Moreover,with only 2,830,839 parameters,the algorithm reached a detection speed of 159FPS,meeting the deployment requirements for edge computing.In the EUM-DET dataset,an improvement of 1.8%in mAP accuracy and 0.5%in recall rate was observed,validating the effectiveness of the proposed structure.The structure presented in this research offers a new perspective for quality inspection in food conveyor line scenarios.
作者
杨嘉诚
黄清华
李少勇
钟卓伦
谢秋波
YANG Jiacheng;HUANG Qinghua;LI Shaoyong;ZHONG Zhuolun;XIE Qiubo(Guangdong Institute of Modern Agricultural Equipment,Guangzhou 510630,China;Shaoguan Xinghe Biotechnology Co.,Ltd.,Shaoguan 512136,China;Shaoguan Benniu Agricultural Development Co.,Ltd.,Shaoguan 512136,China;Guangdong University of Technology,Guangzhou 510006,China)
出处
《现代农业装备》
2024年第4期25-34,共10页
Modern Agricultural Equipment
基金
2022年广东省级现代农业产业园——韶关市曲江区预制菜产业园(GDSCYY2022-013)。
关键词
流水线
面点预制菜
YOLOv8
质量监测
深度学习
注意力机制
conveyor line
noodle prefabricated dishes
YOLOv8
quality inspection
deep learning
attention mechanisms
