【作者】 王亮；

【导师】 赵鹏；

【作者基本信息】 安徽大学 ， 计算机技术， 2024， 硕士

【摘要】 随着深度学习的快速发展和大规模标记数据的收集,图像分类技术引起了人们的广泛关注,并在现实生活中充当着重要地位。然而,传统的图像分类模型在很大程度上依赖于大规模且准确标记的数据。但是,在许多实际应用中,比如濒危物种识别、缺陷检测、医学影像分割等领域,很难收集足够的标记数据。缺乏训练数据可能会导致模型的过度拟合并降低模型的性能。在有限训练数据的情况下,学习出具备更好泛化能力的模型仍然面临着一系列的挑战。基于此,小样本图像分类应运而生。近年来,虽然小样本图像分类取得了巨大的进展,但是在小样本设置下,现有的小样本图像分类方法仍然面临着一些挑战。首先,由于新类的标记样本很少,提取新类标记样本的全局特征往往难以表达类的分布。其次,由于基类和新类的类别是不相交的,具有较大的分布差异,因此利用基类学习到的模型在推广到新类时会出现域偏移问题。然后,由于背景信息的干扰,从图像的全局区域进行特征提取容易引入背景信息,造成学习到的类原型不具有较强的判别性。最后,由于新类的标记样本(支持集)和未标记样本(查询集)的数量具有差异,导致它们之间的分布不一致,因此,根据标记样本获得的类原型具有偏差。为了解决上述挑战,本文开展了以下的工作:(1)针对小样本图像分类方法中存在的少量标记的样本不能充分代表类别的分布,以及将学习到的模型从基类推广到新类时出现的域偏移问题,本文提出了一种基于多尺度任务相关结构图建模的小样本图像分类方法。具体来说,该方法由多尺度任务相关的结构组件学习模块和多尺度图注意力网络模块组成。多尺度任务相关的结构组件学习模块能够根据生成的不同尺度的任务信息,来获得不同尺度的结构组件特征,不仅缓解了域偏移问题,而且丰富了样本的表示,从而更好地表达类的分布。多尺度图注意力网络模块能够探索样本在所有尺度上的不同结构组件之间的相关性。借助图注意力网络的注意力机制、信息聚合和传播机制,多尺度图注意力网络可以充分融合局部和全局信息,并学习到更具代表性和判别性的结构组件特征。本文进行了大量的实验,实验结果验证了所提方法的有效性。(2)针对小样本图像分类方法中存在的背景信息干扰导致的类原型的判别性不强,以及支持集和查询集之间的分布不一致导致的类原型偏差问题,本文提出了一种基于结构原型修正的小样本图像分类方法。具体来说,该方法由显著性结构组件挖掘模块和结构组件原型修正模块组成。显著性结构组件挖掘模块通过生成图像的前景掩码来获取样本丰富的背景无关的结构组件特征,从而使结构组件原型不仅缓解了背景信息的影响,而且能够更好的表达类的分布。结构组件原型修正模块利用高相似度的查询集样本的结构组件特征对支持集的类结构组件原型进行修正,从而减小了学习到的类原型和真实类原型之间的偏差,提升了原型的表达性和判别性,并缓解了支持集和查询集之间的分布差异带来的负面影响。实验结果验证了所提方法的有效性。更多还原

【Abstract】 With the rapid growth of deep learning and the collection of large-scale labeled data,image classification technology has attracted wide attention and played an important role in real life.However,traditional image classification models largely rely on large-scale and accurately labeled data.However,in many practical applications,such as endangered species identification,defect detection,medical image segmentation,etc.,it is difficult to collect enough labeled data.The lack of training data may lead to overfitting of the model and reduce the performance of the model.In the case of limited training data,learning models with better generalization ability still faces a series of challenges.Based on this,few-shot image classification came into being.In recent years,although the few-shot image classification has made great progress,but in the few-shot setting,the existing few-shot image classification methods still face some challenges.First,since there are very few labeled samples of the new class,it is often difficult to extract the global features of the labeled samples of the new class to express the distribution of the class.Secondly,since the categories of the base class and the new class are disjoint and have large distribution differences,the model learned using the base class will have a domain shift problem when it is generalized to the new class.Then,due to the interference of background information,feature extraction from the global area of the image easily introduces background information,causing the learned class prototypes to be less discriminative.Finally,since the number of labeled samples(support set)and unlabeled samples(query set)of the new class is different,resulting in inconsistent distributions between them,the class prototype obtained based on the labeled samples is biased.In order to solve the above challenges,this thesis has carried out the following works:(1)In order to solve the problem that a few labeled samples cannot adequately represent the distribution of the class and domain shift occurs when the learned model is generalized from the base class to the new class,a few-shot image classification method based on multi-scale task-aware structure graph modeling is proposed in this thesis.Specifically,the method consists of a multi-scale task-aware structure component learning module and a multi-scale graph attention network module.The multi-scale task-aware structure component learning module can obtain structure component features of different scales based on the generated task information of different scales,which not only alleviates the domain shift problem,but also enriches the representation of samples,thereby better expressing the distribution of the class.The multi-scale graph attention network module is able to explore correlations between different structure components of a sample at all scales.With the attention mechanism,information aggregation and propagation mechanism of the graph attention network,the multi-scale graph attention network can fully fuse local and global information,and more representative and discriminative structure component features can be learned.This thesis performs extensive experiments and the experimental results validate the effectiveness of the proposed method.(2)In order to solve the problem of weak discriminability of the class prototype caused by background information interference in few-shot image classification methods,and the class prototype bias caused by inconsistent distribution between the support set and the query set,this thesis proposes a few-shot image classification method based on structure prototype rectification.Specifically,the method consists of a significant structure component mining module and a structure component prototype rectification module.The significant structure component mining module generates the foreground mask of the image to obtain the rich background independent structure component features of the sample,so that the structure component prototype not only alleviates the influence of background information,but also can better express the class distribution.The structure component prototype rectification module uses the structure component features of the query set samples with high similarity to rectify the class structure component prototype of the support set,thereby reducing the bias between the learned class prototype and the real class prototype,improving the expressiveness and discrimination of the prototype,and alleviating the negative impact of the distribution difference between the support set and the query set.Experimental results verify the effectiveness of the proposed method.更多还原