【摘要】 关系网络是通过度量分析样本之间相似性的小样本分类方法，其固有的局部连通性限制了对样本全局特征的利用，并且在数据量较少时，模型的泛化能力不足.提出一种混合自监督学习和自适应感知关系网络的小样本分类方法.首先，通过结合自监督的实例级和场景级辅助任务、有监督的小样本分类辅助任务和自适应双相关注意任务提升模型特征表示和泛化能力.其次，引入动态权重平均策略，用于自适应优化辅助任务之间的权重.实例级辅助任务用于学习旋转样本未知类别的转移知识，场景级辅助任务确保不同旋转数据集的分类器预测结果一致性，小样本分类辅助任务则对扩展数据集进行有监督的分类预测平均，优化分类效能.自适应感知关系网络任务通过自适应层对图像特征变化进行自动调节，通过双关联注意力机制增强特征间相互作用，促进关键特征辨识.在数据集miniImageNet、tieredImageNet和CUB-200-2011上进行了验证，提出的方法在不同的骨干网络上都能较好地提升关系网络的分类性能，表明该方法是可行有效的.更多还原

【Abstract】 Relation networks, as a method for few-shot classification through metric analysis of sample similarities, are limited by their inherent local connectivity which restricts the utilization of global features of samples. Furthermore, these networks demonstrate insufficient generalization ability when data is scarce. This paper proposes a hybrid method of few-shot classification combining self-supervised learning with adaptive perception relation networks. Firstly, it enhances model feature representation and generalization ability by integrating self-supervised instance-level and scene-level auxiliary tasks, supervised few-shot classification auxiliary tasks, and adaptive dual-relation attention tasks. Additionally, a dynamic weight averaging strategy is introduced to adaptively optimize weights between auxiliary tasks. Instance-level auxiliary tasks focus on learning transfer knowledge of unknown categories in rotated samples, scene-level tasks ensure consistency in classifier predictions across different rotated datasets, while few-shot classification auxiliary tasks average supervised predictions on expanded datasets, optimizing classification efficacy. The adaptive perception relation network tasks automatically adjust image feature variations through an adaptive layer, and enhance inter-feature interactions via a dual-relation attention mechanism, thereby promoting key feature recognition. The proposed method has been validated on the miniImageNet, tieredImageNet and CUB-200-2011 datasets, demonstrating its capability to significantly enhance the classification performance of relation networks across various backbone networks, proving the feasibility and effectiveness of the proposed approach.更多还原