Multiscale image denoising algorithm based on UNet3+

To fully exploit the multiscale information for image denoising, we introduce the idea of full-scale skip connections in the image segmentation network UNet3+. However, existing UNet3+ networks aggregate multiscale information by directly stitching feature maps, leading to the existence of redundant information. To address this problem, we propose a multiscale selection block for feature selection across multiple convolutional streams as well as within a single scale. Specifically, the selective feature concatenation block dynamically adjusts the receptive field through a self-attentive mechanism to selectively fuse features from multiple resolutions. The dual-stream attention unit performs feature selection from both channel and spatial dimensions within each scale. Additionally, we utilize PixelShuffle for feature reconstruction to enhance multiscale semantic information and maintain information integrity. Based on the above, a novel multiscale image denoising network based on UNet3+ is proposed in this paper. Qualitative and quantitative experimental results show that our network achieves significant results in removing additive white Gaussian noise.