Unsupervised Image Denoising with Cycle-Consistent Adversarial Networks and Residual Blocks

Abstract

Recently, denoising methods based on supervised learning have exhibited promising performance. However, their reliance on external datasets containing noisy-clean image pairs restricts their applicability. To address this limitation, researchers have focused on training denoising networks using unpaired noisy-clean images. In this study, to improve the feasibility of denoising procedures, we propose a simple yet effective unsupervised learning method for unpaired image domain translation using a cycle-consistent adversarial network (CycleGAN) framework. Low-level features are retained by adopting identity loss. In addition, perceptual and style losses are employed as constraints to maintain high-level features. Total variation (TV) loss is used to reduce the artifacts generated during the denoising process. In the CycleGAN framework, to stabilize training, residual blocks are utilized in both the encoder and decoder of the generators, and transposed convolutional layers are utilized in the decoder of the generators. For the discriminators, PatchGAN was applied to extract and distinguish the status of the local features, and spectral normalization was utilized to satisfy the 1-Lipschitz constraint and ensure a steady training process. The experimental results indicated that the proposed method can achieve state-of-the-art (SOTA) denoising performance on both synthetic and real-world datasets. Hence, our method can be applied as a preprocessing method for real-world application, i.e., medical and satellite image analysis, where observed raw images are exposed to unknown noise and collecting target images is merely impossible. Consequently, this highlights the effectiveness and practicality of the proposed method as a potential solution for various noise removal tasks.