Research On Domain Generalization Based On Image Frequency Domain Features

Domain generalization(DG)is a very important fundamental problem in computer vision and pattern recognition.Adaptation to out-ofdistribution data is a tough challenge for all deep learning algorithms that strongly rely on the independent and identically distributed assumption.It leads to unavoidable labor costs and confidence crises in realistic applications.For that,domain generalization aims at mining domainirrelevant knowledge from the source domain that can generalize to unseen target domains,and the key is how to guide the model to learn domain invariant knowledge.In this thesis,we focus on domain generalization for image classification tasks.The previous arts mainly focus on how to learn with inter-domain interpolation or mine domain-irrelevant knowledge,but less effort is paid to studying the very question of which components of images carry the semantic information shared across domains.In this thesis,we tackle domain generalization problems with empirical prior knowledge.First,we propose a disentanglement and interaction method based on high-and low-frequency features of images(H&L)in this thesis.By visualizing and experimental analyzing the high-and low-frequency components of the image,we find that the high-frequency information of images depicts object edge structure and contains high-level semantic information,which is naturally consistent across different domains,and the low-frequency component is much more domain-specific but contains object global structure of images.Motivated by the above analysis,we design an encoder-decoder structure to disentangle high-and lowfrequency features of an image and propose a spatial-mask-based information interaction mechanism to ensure the helpful knowledge from both two parts can cooperate effectively,which can obtain more representational features.Then,we propose a frequency-domain-based data augmentation method(FDAG)in this thesis,which introduces additive noise and multiplicative noise in the image frequency domain.On the one hand,the method enriches the number of source domain data and the diversity of distribution,and reduces the risk of overfitting the source domain data;on the other hand,it can alleviate the sensitivity of the model to frequency fluctuations,thus improving the robustness of the model in extracting highand low-frequency features of the image.Therefore,FDAG can be effectively combined with H&L to further improve the generalization ability of the model.The proposed method obtains state-of-the-art performance on three widely used domain generalization benchmarks.Moreover,extensive ablation studies are implemented to conduct an in-depth study and analysis of the method.