MSR-Net: A Novel Noise Elimination Method for Real CMOS Image Sensor

In the imaging process of Complementary Metal Oxide Semiconductor (CMOS) image sensor, noise is inevitably introduced at various stages. This effect is particularly serious when detecting weak signals, such as ultraviolet light. Through theoretical analysis of CMOS noise, we deduce that the main noise under illumination is non-uniformity noise by averaging the images to eliminate granular noise. Existing image denoising methods rely on simulated noise, which perform poorly when applied to real detectors. To address this issue, we establish a novel CMOS image dataset. Initially, we obtain the non-uniformity noise blocks from the CMOS imaging system designed in this paper. Then, we utilize a GAN network to augment the noise data. Next, we randomly combine these noise blocks with high-quality images from the DF2K dataset to form paired image datasets. In recent years, the development of various deep learning algorithms has significantly improved the effectiveness of image noise reduction compared to traditional algorithms. This paper combines convolutional neural network and proposes the MSR-Net denoising algorithm, which is based on the U-Net network and incorporates the Res2Net module as its main network structure. It provides features with different scale receptive fields and enhances image details. Additionally, to more accurately reflect the visual perceptual effects of the images, we propose a novel image evaluation metric, Uniform pixel outliers (UPO), making the image evaluation more adequate. Experiments were conducted on our proposed image dataset, results indicate that compared with similar noise reduction algorithms, this method performs better in both qualitative and quantitative aspects, effectively suppressing noise dominated by non-uniform noise.