摘要
针对铝型材瑕疵尺寸差异显著、瑕疵识别率低以及边缘设备实施检测难等问题,提出一种融合迁移学习与测试时增强的轻量化铝型材瑕疵分类方法。首先,使用预训练MobileViT-XXS模型进行迁移微调学习,对顶层MobileViT模块、最终卷积层、分类层权重实施微调。其次,设计了微调阶段和测试阶段对称的弱数据增强策略:水平翻转和图像缩小,同时在微调阶段的数据增强过程中引入随机性,按50%概率进行水平翻转。最后,利用模型对每个增强样本独立预测,以概率均值融合集成预测结果。实验结果证明,在相同微调策略下,相比于无特定增强的MobileViT-XXS迁移微调模型,这种轻量化方法准确率相对提高2.74%,F1值相对提高2.81%。
To address the challenges of significant variation in aluminum profile defect sizes,low recognition rates,and difficulties in deploying detection systems on edge devices,a lightweight defect classification method integrating transfer learning and test-time augmentation is proposed.First,the pre-trained MobileViT-XXS model is fine-tuned via transfer learning.This involves adjusting the weights of the top MobileViT module,the final convolutional layer,and the classification layer.Second,a symmetric weak data augmentation strategy is designed for both the fine-tuning and testing phases,incorporating horizontal flipping and image scaling.During fine-tuning,randomness is introduced with a 50%probability of applying horizontal flipping.Finally,the model generates independent predictions for each augmented test sample,and the results are fused by averaging the probability outputs.Experimental results demonstrate that under the same fine-tuning strategy,this lightweight approach improves accuracy by 2.74%and the F1-score by 2.81%compared to the baseline MobileViT-XXS transfer learning model without specific augmentation.
作者
黄清兰
游贵荣
乐宁莉
郑佳芳
HUANG Qinglan;YOU Guirong;LE Ningli;ZHENG Jiafang(Information Technology Center,Fujian Business University,Fuzhou 350012,China)
出处
《邵阳学院学报(自然科学版)》
2025年第6期39-45,共7页
Journal of Shaoyang University:Natural Sciences Edition
基金
福建省中青年教师教育科研项目(JAT242007)。
