Algorithm Research Of Few-shot Domain Adaptation

With the popularization and development of artificial intelligence(AI),the trend of AI customization is rising.How to transfer the trained model to different scenarios quickly and well has become one of the technical difficulties.Domain adaptation is an important assumption to solve this problem.Traditional domain adaptation algorithms are mostly based on unsupervised learning.Unsupervised domain adaptation(UDA)algorithms require a large number of target training samples,which is expensive and not always available.To overcome this problem,this thesis proposes a novel few-shot domain adaptation algorithm,called Few-shot Adversarial Discriminative Domain Adaptation(FSADDA),which is inspired by contrastive learning,metric learning,generative adversarial networks and other related researches.This work uses only a few samples to achieve better performance then UDA on the key datasets.To further verify the generalization ability of FSADDA,we then introduce it into the object detection task.Combined with the latest researches,a one-step few-shot domain adaptive object detection algorithm is devised over FSADDA.Extensive experimental results demonstrate that our works significantly outperform state-of-the-art methods across multiple domain adaptation benchmark datasets.The main research contents and contributions of this thesis are as follows:(1)Firstly,from the perspective of data augmentation,this thesis proposes two novel algorithms,called Dual-domain Samples Mixup and Dual-domain Fine-grained Mosaic,which arm to bring down the effect of over-fitting in the process of few-shot domain adaptation.Secondly,in the ablation experiments that used for finding the most suitable neural network design patterns during domain adapting step,we explore four convolutional neural network design patterns and propose a new domain classification discriminator.At the same time,we improve the Adversarial Discriminative Domain Adaptation algorithm by Dual-domain Samples Pairing algorithm and using metric loss as a strong constraint.Finally,focus on reducing training instability in adversarial domain adaptation algorithm,we elucidate a bag of tricks that apply to adversarial domain adaptation during neural network training step,which contains network weights initialization,hyper-parameter search and reinforcement learning,further improve the performance and availability of FSADDA.(2)Most of the previous domain adaptation algorithms focus on the classification task,and the researches about the object detection task are scarcity and inefficiency.To this end,we extend the data augmentation algorithms,augment the data both in the levels of image and instance.In addition,we propose a novel Multi-scale One-step Few-shot Domain Adaptive Object Detection Network,which is based on FSADDA and adversarial criterion.We extensively compare this approach to the state-of-the-art in multiple domain adaptation experiments.Results show that our algorithm is useful both theoretically and experimentally.