Zero-Shot Image Denoising with W-MSA and Residual connection

In this paper, we propose an image denoising approach incorporating zero-shot learning with Window-based Multi-head Self-Attention (W-MSA). Our scheme extends a conventional small-scale image-to-image neural network by adding a convolutional layer, W-MSA, and a fully connected layer. Additionally, we introduce residual connections at the output of the image-to-image neural network to shift the learning objective. Instead of modeling noise components, the network predicts an image that retains the noise of the input image. These methods, integrated into the Zero-Shot Noise2Noise (ZS-N2N) framework, enable high-quality image restoration. Experimental results on artificial and real-world noise demonstrated that our method achieved superior noise removal compared to other approaches.