Self-Supervised Image Denoising with Blind-Spot Network and Residual Learning

The utilization of deep learning methods in image denoising has witnessed remarkable growth, primarily due to the training on extensive datasets. However, acquiring vast amounts of clean, uncontaminated images in practical settings is both costly and challenging, thus spurring the rise and evolution of self-supervised deep learning approaches. Within this framework, we have identified two primary challenges: enhancing the efficacy of self-supervised learning and ensuring its adaptability across diverse network architectures. To tackle these issues, we introduce a self-supervised image denoising methodology rooted in the Blind2-Unblind framework. This method mitigates information loss inherent in blind-spot-driven denoising techniques and crafts an efficient denoising network leveraging the U-Net and ResNet architectures. By designing a deeper convolutional network structure and carefully tuning training parameters, we enhance the performance of Blind2Unblind while preventing gradient explosions. Experimental results demonstrate that our proposed method achieves an average improvement of 0.16 dB in Peak Signal-to-Noise Ratio (PSNR) and 0.0022 in Structural Similarity Index (SSIM) compared to the original network, thus validating the robustness of the Blind2Unblind training strategy across various network structures.