Paper Reading - 1、Registration based Few-Shot Anomaly Detection

Registration based Few-Shot Anomaly Detection: 无需微调即可推广，上交大、上海人工智能实验室等提出基于配准的少样本异常检测框架

RegAD的模型架构：

The model architecture of the proposed RegAD. Given paired images from the same category, features are extracted by three convolutional residual blocks each followed by a spatial transformer network. A Siamese network acts as the feature encoder, supervised by a registration loss for feature similarity maximization （给定来自同一类别的图像对，通过三个卷积残差块提取特征，每个残差块后跟一个空间变换网络。孪生网络充当特征编码器， 由配准损失来监督特征相似度最大化）.