基于改进YOLOv8的纸箱外形异常检测算法  

A Carton Shape Anomaly Detection Algorithm Based on Improved YOLOv8

作　　者：张杰韬 张可义 徐汉均 宋江旭 张万青 ZHANG Jie-tao;ZHANG Ke-yi;XU Han-jun;SONG Jiang-xu;ZHANG Wan-qing(Beijing Research Institute of Automation for Machinery Industry Co.,Ltd.,Beijing 100120;BZS(Beijing)Technology Development Co.,Ltd.,Beijing 100120)

机构地区：[1]北京机械工业自动化研究所有限公司,北京100120 [2]北自所(北京)科技发展股份有限公司,北京100120

出　　处：《制造业自动化》2025年第3期168-174,共7页Manufacturing Automation

摘　　要：为了实现对现有物流行业中输送线上纸箱外形异常检测问题的优化,在YOLOv8框架的基础上,改进了骨干网络,推出一种融合ShiftConv模块的轻量化骨干网络。提出了一种新的检测头DSDET,使网络在轻量化的同时保持较高精度,为解决在目标遮挡情况下检测效果不佳的问题,使用了Repulsion Loss损失函数,通过损失函数使同类间的预测框尽可能远离,从而使检测网络模型在存在较多遮挡情况的纸箱外观异常的检测任务中有更高的精度。通过在自制纸箱数据集上对目标检测网络模型进行测试,平均检测精度可达93.8%,与原有YOLOv8相比检测精度提升2.4%,这种方法可应用于自动化物流实际场景下的纸箱外形异常目标检测。To optimize the detection of abnormal shapes of cardboard boxes on conveyor belts in the existing logistics industry,an improved backbone network has been developed based on the YOLOv8 framework,introducing a lightweight backbone network that integrates a shiftconv module.A new detection head,DSDET,has been proposed to maintain relatively high accuracy while keeping the network lightweight.To solve the problem of poor detection performance in case of target occlusion,a Repulsion Loss function has been used.This loss function encourages predictions of the same class to be as distant as possible from each other,thereby increasing the accuracy of the detection network model in tasks involving carton shape anomaly detection with a lot of occlusions.Testing of the object detection network model on a self-made carton dataset has shown an average detection accuracy of 93.8%,which is a 2.4%improvement in detection accuracy compared to the original YOLOv8.This method can be applied to the detection of carton shape anomalies in automated logistics scenarios.

关 键 词：深度学习 目标检测 YOLO 轻量化 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]