Frequency Domain Feature Learning with Wavelet Transform for Image Translation

Huan Zhao, Yujiang Wang, Tingting Li, Song Wang, Lixuan Li, Xupeng Zha, Zixing Zhang

Abstract: Image-to-image translation serves as an essential field of research in computer vision. Existing models frequently cause accidental distortion over non-target attributes, leading to overfitting of the generated image to the reference domain and poor visual quality. To address this problem, we propose Frequency Domain Feature Learning with Wavelet Transform, namely FDFL-WT, which with better non-target attributes retention and more precise image capture of style. This method utilizes the wavelet transform to capture the image’s approximation coefficients and diagonal coefficients, then we suggest wavelet reconstruction loss and wavelet translation loss. The former comprehensively records the context information of the source image to make the generated image realistic, whereas the latter improves the generator’s capacity to decouple attributes by assisting the model in efficiently retaining image content attributes. Experimental results on CelebA-HQ dataset indicate that FDFL-WT achieves about a 7.03\(\%\) performance improvement comparing with methods in the FID score of realism and disentanglement tests.