【作者】 余悦；

【导师】 成科扬；

【作者基本信息】 江苏大学 ， 计算机科学与技术， 2024， 硕士

【摘要】 基于生物学的证据,越来越多的学者使用少样本学习去解决带标签数据有限的问题。然而,大多数少样本学习模型要求训练数据域和测试数据域处于相同分布中,当这两个阶段的数据域存在领域偏差时则会导致少样本模型难以在实际应用中推广。此外,在少样本设置下,神经网络可能存在简单性偏见现象,即容易关注最简单的非因果特征(颜色和背景纹理等),而倾向于忽略复杂的语义特征,这些简单特征足以使模型在预定义的训练数据域中区分少样本类别,但会导致在领域分布存在显著差异的情况下难以学习到可用于泛化的纯粹域不变特征。本文以该领域的研究现状为基础,针对现存的问题展开基于不确定性元学习和域感知度量的跨域少样本图像分类方法研究,具体的研究工作包括以下内容:(1)针对元训练与元测试存在的领域偏移问题,提出基于不确定性生成元学习的跨域少样本图像分类方法,摆脱加性仿射变换的局限性,以在合理范围内生成更有挑战性的特征分布,最大化提高模型的跨域适应能力。首先,在梯度不确定性生成方法中,图像特征被定义为服从高斯分布的概率表示,通过引入分类结果的预测梯度对不确定性的概率表示进行建模,从中随机抽取新的充分统计来生成全新的特征。其次,提出了一个因果不变信息方法来均匀化任务梯度的同质性,为不确定性分布的边界构建提供进一步的保障。实验结果表明所提出的算法在弱泛化基准和强泛化基准上平均提高2.23%和2.01%的跨域泛化精度。(2)针对跨域设置存在的简单性偏差问题,提出基于域感知的跨域度量方法,辅助模型消除特征信道中的虚假相关性,从而缓解简单性偏差造成的负面影响,以提取更纯粹的域不变特征。首先,提出领域显著化感知方法,增强源域特征信道中语义特征的判别性,并在后续利用域鉴别器帮助模型筛选和聚焦于其中的可泛化鲁棒特征。其次,提出一个语义感知度量器来学习经过鉴别和提取的类特征之间的局部信息,从而进一步提取可用于跨域新类的语义知识。实验的可视化结果表明所提出的方法可以提高语义特征的可判别性,并帮助模型学习其中的纯粹知识。(3)基于本文所提出的研究方法,设计并实现一个少样本跨域图像分类原型系统。该系统主要包括元学习器构建、基学习器构建与评估和图像预测三个模块。经过其开发测试结果表明,所搭建的系统具有良好的实用价值和应用前景,能广泛应用于真实世界的不同领域之中。更多还原

【Abstract】 Based on biological evidence,more and more scholars are using few shot learning to solve the problem of limited labelled data.However,most few-shot learning models require that the training data domain and the test data domain be in the same distribution,which makes it difficult to generalise the few-shot model to practical applications when there is a domain bias in the data domains of these two phases.Furthermore,in the few-shot setting,neural networks may suffer from the phenomenon of simplicity bias,i.e.a tendency to focus on the simplest non-causal features(colour and background texture,etc.)and a tendency to ignore complex semantic features that are simple enough to allow the model to discriminate between a few classes in the training in a predefined homogeneous domain,but that can lead to difficulties in learning purely domain-invariant features that can be used for generalisation in the presence of a significantly differentiated domain distribution.In this thesis,based on the existing research status in the studied area,we carry out a research on cross-domain few-shot image classification methods based on uncertainty meta-learning and domain-aware metrics to address the existing problems,and the specific research work includes the following:(1)An uncertainty-based generative meta-Learning approach for cross-domain few-shot image classification is proposed to get rid of the limitations of additive affine transformations in order to maximise the cross-domain adaptability of the model by generating more challenging feature distributions within a reasonable range.Firstly,in the gradient uncertainty generation method,image features are defined as probabilistic representations obeying a Gaussian distribution,and the probabilistic representations of uncertainty are modelled by introducing a predictive gradient of the classification result,from which new sufficient statistics are randomly extracted to generate completely new features.Secondly,a causal invariant information method is proposed to homogenise the homogeneity of the task gradients,providing further guarantees for the boundary construction of uncertainty distributions.Experimental results show that the proposed algorithm improves the cross-domain generalisation accuracy by an average of 2.23%and 2.01% on both weak and strong generalisation benchmarks.(2)A domain-aware cross-domain classification metric is proposed to assist the few shot model in eliminating spurious correlations in the feature channel thereby mitigating the negative impact caused by simplicity bias in order to extract purer domain-invariant features.Firstly,the domain saliency perception method is proposed to enhance the discriminative nature of semantic features in the source domain feature channel by enhancing the discriminative nature of semantic features in the source domain feature channel and subsequently using a domain discriminator to help the model filter and focus on the generalisable robust features therein.Secondly,a semantic-aware metric is proposed to learn the local information between the discriminated and extracted class features to further extract the semantic knowledge that can be used to cross-domain new classes.The visualisation results of the experiments show that the proposed approach improves the discriminability of semantic features and helps the model to learn pure knowledge from them.(3)Based on the methodological study presented in this thesis,a prototype system for cross-domain image classification with few samples is designed and implemented.The system mainly includes three modules: meta-learner construction,base-learner construction and evaluation,and image prediction.The results of its development and use show that the constructed system has good practical value and application prospects,and can be widely used in different fields of real scenarios.更多还原