【摘要】 现有的跨领域少样本分类模型受限于域特定因素的干扰，限制了其有效性。为此，提出了一种基于高斯仿射的通道鉴别网络。具体来讲，所提出的学习框架包含随机高斯仿射模块和域通道鉴别模块，在随机高斯仿射模块中，通过对特征的充分统计量进行高斯扰动以生成区别于源域数据分布的全新特征分布，从而显著化训练数据特征中域不变信息；在域通道鉴别模块中，将经过增强前后的特征图输入到域鉴别器中引导模型区分和提取其中的域不变特征，以达到提高模型泛化能力的目的。最后，在两个目标数据集进行实验，结果验证了所提出方法的可行性和有效性。更多还原

【Abstract】 Existing cross-domain few-shot classification models are limited by the interference of domain-specific factors, resulting in poor cross-domain performance. To overcome this problem, a channel knowledge discriminative network for cross-domain few-shot image classification is proposed. Specifically, the proposed learning framework contains a stochastic Gaussian affine module and a channel knowledge discrimination module. In the stochastic gaussian affine module, we salientise the domain-invariant information in the feature map by Gaussian perturbing sufficient statistics of the features to generate a new feature distribution that is distinct from the source domain data distribution. In the channel knowledge discrimination module, the feature maps before and after enhancement are fed into the domain discriminator to guide the model to distinguish and extract the domain-invariant features therein, thus improving the model generalisation capability. Finally, we conduct experiments on two target datasets and the results validate the effectiveness of the proposed method.更多还原