WDIG: a wavelet domain image generation framework based on frequency domain optimization

Abstract In the end-to-end image generation task, spatial domain of pixel space cannot explicitly separate low-frequency general information such as texture and color from high-frequency detail structure identity. The loss function calculated in fails to effectively constrain maintenance information, generated quality is insufficient. this paper, a wavelet (WDIG) framework proposed preserve frequency images, which functions are constructed space. space, characteristic signal obtained by setting appropriate Gaussian kernel adopting fuzzy method. $$\ell_{1}$$ ℓ 1 norm for information. corresponding channel sub-band coefficients transform, separated into respectively coefficients. WDIG can model training more accurately optimize precisely, so better maintain details image. evaluated applications including style transfer, translation Generative Adversarial Nets (GAN) Inversion. Experimental results show that retain images generate realistic improve above generation.