Research On RAW Image Denoising Algorithm Based On Image Signal Processing

Raw image is the original data stored after the acquisition equipment obtains photon information.Compared with RGB image,the noise of raw image is more in line with statistical characteristics,and better image processing results can be obtained based on the analysis of imaging principle and the inverse deduction of ISP process.All devices introduce noise when capturing images due to the physics of the sensor.Especially in some extreme scenes,such as night scenes,dark light,etc.,the image quality will be seriously polluted.The phenomenon of noise in the initial stage seriously affects the subsequent use.Through the analysis of image signal processing,we can get better image processing performent.Focusing on the field of RAW image denoising,this thesis studies the denoising algorithm framework based on neural network,and optimizes the existing algorithms for two important indicators:denoising performance and efficiency.The main work and contributions are as follows:Firstly,a raw image denoising method based on SUnet++ is proposed.Combining the performance advantages of Unet++ and the efficiency advantages of deep separable convolution,a novel deep separable Unet++ neural network model is designed.By using SUnet++ neural network model to learn the mapping from very low light noise raw image to long-time exposure RGB image,the very low light noise raw image is joint denoising and demosaicing.Secondly,a raw image denoising method based on Vi T-Unet is proposed.According to the good learning ability of convolutional neural network for basic knowledge under the condition of small-scale samples and the efficient performance of self attention mechanism in dealing with complex computer vision tasks,a novel Vi T-Unet neural network model is designed.By improving the Vi T method,the network develops a dynamic Vi T module with multi-stage customized blocking process,and embeds it into the deep position of the classical Unet neural network model to ensure that the raw image has better denoising effect in small-scale samples.This research is based on the SID method.The SUnet++ neural network and Vi T-Unet neural network proposed in this topic are applied to the joint denoising and demosaicing problem of raw images,and the comparative experiments are carried out on the two data sets of SID and ELD.Experiments show that this study has achieved better results in the generalization ability,detail clarity and color accuracy of denoising,which proves the effectiveness of the improvement of Unet neural network by SUnet++ neural network and Vi T-Unet neural network.