曹海文1,3　陈德平3,4　杨丹妮3  王楠3  钟震宇3  段二强2

（1.广东工业大学，广东 广州 510006

2.佛山市云米电器科技有限公司，广东 佛山 528308

3.广东省科学院智能制造研究所，广东 广州 510070

4.汕头大学，广东 汕头 515063）

摘要：针对瓶装产品褶皱的旋转角度随机、多尺度，以及与瓶体背景对比度低等特性，提出基于改进YOLOv8的瓶装产品褶皱检测模型。该模型利用Sobel算子提取边缘特征，通过融合边缘特征与空间特征，增强低对比度区域的边缘特征提取能力；通过融合卷积与加性注意力机制，使YOLOv8模型能够同时提取图像的全局特征与局部特征，改进模型对不同尺度褶皱的特征融合能力，从而增强对旋转目标的鲁棒性；利用基于注意力的尺度内特征交互模块替换快速空间金字塔池化模块，增强骨干网络在同一尺度内的褶皱特征提取能力，解耦特征提取与多尺度特征融合任务；利用双向特征金字塔网络改进路径聚合网络，通过加权特征融合的方式动态调整特征权重，提升模型的多尺度特征融合能力；基于该模型构建了一套瓶装产品褶皱检测系统。实验结果表明，改进的YOLOv8模型的AP50为82.3%，相较于YOLOv8-OBB模型提升了6.6%，能更好地完成瓶装产品褶皱的检测任务。经实际生产线验证，该瓶装产品褶皱检测系统的检测精度、推理速度和稳定性均符合工业应用要求。

关键词：瓶装产品褶皱检测；改进YOLOv8；Sobel算子；加性注意力机制；基于注意力的尺度内特征交互模块；双向特征金字塔网络

中图分类号：TP391           文献标志码：A           文章编号：1674-2605(2025)04-0002-11

DOI：10.12475/aie.20250402                                 开放获取

Wrinkle Detection Model for Bottled Products Based on Improved YOLOv8

CAO Haiwen1,3　CHEN Deping3,4　YANG Danni3　                                WANG Nan3  ZHONG Zhenyu3　DUAN Erqiang2

（1.Guangdong University of Technology, Guangzhou 510006, China

2.Foshan VIOMI Electrical Technology Co., Ltd., Foshan 528308, China

3.Institute of Intelligent Manufacturing, Guangdong Academy of Sciences, Guangzhou 510070, China

4.Shantou University, Shantou 515063, China）

Abstract: Aiming at the random rotation angles, multi-scale characteristics, and low contrast with the bottle background of wrinkles in bottled products, this paper proposes an improved YOLOv8-based wrinkle detection model for bottled products. The model utilizes the Sobel operator to extract edge features and enhances edge extraction capability in low-contrast regions by fusing edge features with spatial features. By integrating convolution and additive attention mechanisms, the YOLOv8 model can simultaneously extract global and local features of the image, improving its ability to feature fusion wrinkles of different scales and enhancing robustness against rotated targets. The fast spatial pyramid pooling module is replaced with an intra-scale feature interaction module based on attention, strengthening the backbone network's ability to extract wrinkle features within the same scale and decoupling feature extraction from multi-scale feature fusion tasks. The path aggregation network is improved using a bidirectional feature pyramid network, which dynamically adjusts feature weights through weighted feature fusion to enhance the model's feature fusion capability. Based on this model, a wrinkle detection system for bottled products is constructed. Experimental results show that the improved YOLOv8 model achieves an AP50 of 82.3%, a 6.6% improvement over the YOLOv8-OBB model, demonstrating superior performance in wrinkle detection for bottled products. Verified on actual production lines, the system meets industrial requirements in terms of detection accuracy, inference speed, and stability. 

Keywords: bottled products wrinkle detection; improved YOLOv8; Sobel operator; additive attention mechanism; attention-based intra-scale feature interaction module; bidirectional feature pyramid network