A local generation-mix cascade network for image translation with limited data

Abstract

Image translation based on deep generative models often overfits with limited data. Current methods overcome this problem through mix-based data augmentation. However, if latent features are mixed without considering semantic correspondences, augmented samples may exhibit visible artifacts and mislead model training. In this paper, we propose a Local Generation-Mix Cascade Network (LogMix), a data augmentation strategy for image translation tasks with limited data. Through cascading a local feature generation module and mixing module, LogMix enables the generation of a reference feature bank, which is mixed with the most similar local representation to form a new intermediate sample. Furthermore, we design a semantic relationship loss based on the mixed distance of latent features ensures consistency in the distribution of features between the generated and source domains. LogMix effectively mitigates the overfitting problem by learning to translate intermediate samples instead of memorizing the training data Experimental results across various tasks demonstrate that, even with limited data, LogMix data augmentation reduces image ambiguity and offers significant advantages in establishing realistic cross-domain mappings.