NoiseTransfer: Image Noise Generation with Contrastive Embeddings

Deep image denoising networks have achieved impressive success with the help of a considerably large number synthetic train datasets. However, real-world is still challenging problem due to dissimilarity between distributions real and noisy Although several datasets been presented, (i.e., pairs clean images) limited, acquiring more noise laborious expensive. To mitigate this problem, numerous attempts simulate models using generative studied. Nevertheless, previous works had multiple handle different distributions. By contrast, we propose new model that can synthesize images Specifically, adopt recent contrastive learning learn distinguishable latent features noise. Moreover, our generate by transferring characteristics solely from single reference image. We demonstrate accuracy effectiveness for both known unknown removal.