关键词: 关键词: 小样本学习, 关系网络, 跨域学习, 元学习, 图像分类

Abstract: Aiming at the problem of poor classification accuracy of few-shot learning models in the domain shift, a cross-domain few-shot model based on Relation network and Vision Transformer ReViT (Relation Vision Transformer) was proposed. First, vision Transformer was introduced as a feature extractor, and the pre-trained deep neural network solves the problem of insufficient feature expression ability. Then, a shallow convolutional network was used as a task adapter to enhance the knowledge transfer ability of the model, and a nonlinear classifier was constructed based on the Relation Network and the channel attention mechanism. Finally, a four-stage learning strategy of "pre-training meta-training fine-tuning meta-testing" was adopted to train the model, which further improved the cross-domain classification performance of ReViT. The experimental results show that ReViT has good performance in dealing with the cross-domain few-shot classification problem, and the classification accuracies under in-domain scenarios and out-of-domain scenarios in the Meta-Dataset dataset are improved by 5.82 and 1.17 percentage points, respectively. ReViT is improved by 1.00, 1.54 and 2.43 percentage points on the 5-way 5-shot for the three sub-problems EuroSat(European Satellite data), CropDiease and ISIC(International Skin Imaging Collaboration) of the BCDFSL(Broader study of Cross-Domian Few-Shot Learning) dataset, respectively. ReViT is improved by 0.13, 0.97, and 3.40 percentage points on the 5-way 20-shot for EuroSat, CropDiease, and ISIC, respectively. ReViT is improved by 0.36 percentage point improvement on the 5-way 50-shot for CropDiease. The good classification results show that ReViT can have applications in image classification tasks with sparse sample size, such as satellite image recognition, human skin disease recognition and crop disease recognition.

Key words: Keywords: few-shot learning, relation network, cross-domain, meta learning, image classification