Self-Supervised Contrastive Denoising with ViT and U-Net: A Hybrid Approach

Image denoising is still challenging, particularly when no matched clean data is available. To handle this issue, a novel hybrid architecture that combines the detailed spatial reconstruction capabilities of a U-Net decoder with the global receptive field of a Vision Transformer (ViT) encoder. This approach makes use of U-Net's ability in maintaining local details and the ViT's capacity to record long-range dependencies. Clean target images are not required for training because the architecture is learned using a self-supervised contrastive learning technique. The suggested model obtains a Structural Similarity Index Measure (SSIM) of 0.94 and a Peak Signal-to-Noise Ratio (PSNR) of 35 dB on the difficult SIDD dataset. These outcomes show competitive performance, matching or outperforming a number of cutting-edge denoising techniques. Through the integration of CNN and Transformer paradigms inside a contrastive learning framework, the study demonstrates a viable path toward effective denoising in low-supervision applications.