Single Image Denoising via a New Lightweight Learning-Based Model

Abstract

Restoring a high-quality image from a noisy version poses a significant challenge in computer vision, particularly in today’s context where high-resolution and large-sized images are prevalent. As such, fast and efficient techniques are required to address noise reduction in such images effectively. Deep CNN-based image-denoising algorithms have gained popularity due to the rapid growth of deep learning and convolutional neural networks (CNNs). However, many existing deep learning models require paired clean/noisy images for training, limiting their utility in real-world denoising scenarios. In this paper, we propose a fast residual denoising framework (FRDF) designed based on zero-shot learning to address this issue. The FRDF first employs a novel downsampling technique to generate six different images from the noisy input, which are then fed into a lightweight residual network with 23K parameters. The network effectively utilizes a hybrid loss function, including residual, regularization, and guidance losses, to produce high-quality denoised images. Our innovative downsampling approach incorporates zero-shot learning principles, enabling our framework to generalize to unseen noise types and adapt to diverse noise conditions without needing labelled data. Extensive experiments conducted on synthetic and real images confirm the superiority of our proposed approach over existing dataset-free methods. Extensive experiments conducted on synthetic and real images show that our method achieves up to 2 dB improvements in PSNR on the McMaster and Kodak24 datasets. This renders our approach applicable in scenarios with limited data availability and computational resources.

Keywords

• Deep learning
• image denoising
• self-supervision
• downsampling
• zero-shot learning