Pre-training And Fine-tuning Methods With High Generalization Ability For Few-shot Recognition

Object recognition plays an important role in various fields,such as national economy and national security.However,existing deep learning based methods require a large number of training samples,which greatly limits their applications.Few-shot object recognition,which aims at using few samples to quickly learn and recognize objects from new classes,has become an active research topic.Pre-training and fine-tuning a model is one main framework for solving few-shot object recognition problems with good performance.However,existing methods have problems with inconsistent task forms in the pre-training and fine-tuning stages,as well as model uncertainty caused by data scarcity,resulting in poor generalization of few-shot object recognition models.To address the above issues,this paper conducts research on pre-training and fine-tuning methods with high generalization ability for few-shot recognition,aiming to improve the generalization ability of few-shot object recognition model by improving the generalization ability of the pre-training and fine-tuning stage.Specifically,the main research content and possible innovative points of the paper include:1.Focusing on the model pre-training stage,a Mixed Label MetaLearning(MLML)method is proposed to pre-train the object recognition model by constructing a series of meta tasks and solve the problem of inconsistent task forms between the pre-training stage and fine-tuning stage.Meanwhile,MLML generates mixed label of sample to guide the meta-learning of object recognition model.Experimental results on benchmark datasets show that MLML can effectively improve the generalization of pre-trained models;2.Focusing on the model fine-tuning stage,a Bayesian SharpnessAware Prompt Tuning(BSAPT)method is proposed to learn the parameter distribution of the visual prompt token and solve the problem of model uncertainty caused by data scarcity.BSAPT uses the Sharpness-Aware Minimization(SAM)regularization strategy to learn the model weight with low and flatten training loss to improve the generalization ability of the fine-tuned model.Experimental results show that BSAPT achieves SOTA on mainstream few-shot learning benchmark.