Unsupervised Contrastive Learning for Signal-Dependent Noise Synthesis
We present a simple yet robust noise synthesis framework based on unsupervised contrastive learning. With access to clean images only, the proposed contrastive noise synthesis framework trains a Glow-based generative model to synthesize image noise in a self-supervised fashion. We utilize the signal-dependency of the synthetic noise as a discriminative feature for the instance-wise discrimination pretext task and introduce a noise contrastive loss based on maximum mean discrepancy. The empirical results show that, with only 4312 parameters, the noise synthesized by the proposed framework shows advantages over the noise synthesized by traditional statistical models both qualitatively and quantitatively. The proposed framework fills a methodological gap in learning-based noise synthesis and can be used as an alternative to traditional statistical models.
PDF Abstract