独立性视角下的相频融合领域泛化方法

Domain generalization method of phase-frequency fusion from independent perspective

针对现有的领域泛化(DG)方法对领域特征处理粗糙和泛化能力弱的问题,提出一种基于频域特征独立性这一独特视角解决领域泛化问题的方法。首先,设计频域分解算法,将图像的深度特征快速傅里叶变换(FFT)后,再从相位信息中获得领域无关特征,以提高模型对领域无关特征的识别能力;其次,基于独立性视角,通过对样本的特征赋权,进一步消除频域特征中各属性的相关性,提取最有效领域无关特征,解决样本特征之间相关性带来的泛化能力差的问题;最后,提出幅度融合策略,拉近源域和目标域的距离,进一步提升模型对未知领域的泛化能力。在流行的图像领域泛化的数据集PACS和VLCS上的实验结果表明,所提方法的准确率均值比StableNet分别高0.44、0.59个百分点,且在各个数据集上均取得了优秀的性能。

The existing Domain Generalization(DG)methods process the domain features poorly and have weak generalization ability,thus a method based on the feature independence of the frequency domain was proposed to solve the domain generalization problem.Firstly,a frequency domain decomposition algorithm was designed to obtain domainindependent features from phase information by the Fast Fourier Transform(FFT)of depth features of the image,improving the recognition ability of domain-independent features.Secondly,from the independence perspective,the correlation of attributes in frequency domain features was further eliminated by weighting the features of samples,and the most effective domain-independent features were extracted to solve the poor generalization problem caused by correlation between sample features.Finally,the amplitude fusion strategy was proposed to narrow the distance between the source domain and the target domain,so as to further improve the generalization ability of the model to the unknown domain.Experimental results on popular image domain generalization datasets PACS and VLCS show that the average accuracy of the proposed method is 0.44,0.59 percentage points higher than that of StableNet,and the proposed method achieves excellent performance on all datasets.