摘要
小样本图像分类解决了传统图像分类在数据量不足时表现不佳的问题,其难点在于如何充分利用稀缺的样本标签数据预测真实的特征分布。一些最新方法采用随机遮挡或混合插值等数据增强方法来提高数据标签样本的多样性和泛化性,但仍然存在以下问题:1)随机遮挡具有不确定性,会出现完全遮挡或暴露前景的情况,导致样本关键信息丢失;2)由于混合插值后的数据分布过于平均,模型难以准确区分不同类别之间的差异和边界。针对上述问题,提出一种基于显著性掩模混合的数据增强方法。首先,通过视觉特征隐蔽融合和置信度裁剪选择策略,对图像关键特征信息进行自适应的筛选与保留;其次,采用视觉特征显著性融合方法,计算出图片中各个区域的重要性,引导图片融合,增加所得图片的多样性和丰富性,使类别边界更加清晰。所提方法在多个标准小样本图像分类数据集(miniImageNet,tieredImageNet,Few-shot CIFAR100和Caltech-UCSD Birds-200)上表现出色,优于最先进方法约0.2%~1%,在小样本图像分类中具有显著的潜力和优势。
Few-shot image classification addresses the problem of poor performance in traditional image classification when data is scarce.The challenge lies in effectively utilizing sparse sample label data to predict the true feature distribution.To tackle this,some recent methods adopt data augmentation techniques such as random mas-king or mixed interpolation to enhance the diversity and generalization of data label samples.However,there are still the following issues:1)Due to the uncertainty of random masking,situations where the foreground is either completely masked or exposed may occur,leading to the loss of crucial information in samples;2)Because the data distribution after mixed interpolation tends to be overly uniform,models find it difficult to accurately distinguish differences between different classes,thus failing to effectively delineate boundaries between different categories.To address these problems,this paper proposes a data augmentation method based on Saliency Mask Mixup.Firstly,through Mask Mix(M-Mix)and Confident Clip Selector(CCS),adaptive selection and retention of key feature information in images are performed.Secondly,using Saliency Fuse(SF),the importance of various regions in the image is calculated to guide image fusion,making the resulting images more diverse and rich,thereby making category boundaries clearer.The proposed method demonstrates outstanding performance on multiple standard few-shot image classification datasets(such as miniImage-Net,tiered-ImageNet,Few-shot CIFAR100,and Caltech-UCSD Birds-200),outperforming state-of-the-art methods by approximately 0.2~1%.These results indicate significant potential and advantages of the proposed method in few-shot image classification.
作者
陈亚当
高宇轩
卢楚翰
车洵
CHEN Yadang;GAO Yuxuan;LU Chuhan;CHE Xun(School of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)
出处
《计算机科学》
北大核心
2025年第6期256-263,共8页
Computer Science
基金
国家自然科学基金(62473201,62477026)
江苏省重点研发计划产业前瞻与关键核心技术项目(BE2022161)
无锡市产业创新研究院先导技术预研项目。
关键词
小样本学习
图像分类
对比学习
数据混合
数据增强
显著图
Few-shot learning
Image classification
Contrastive learning
Date mixing
Data augmentation
Saliency map
