Deep Learning-based Multi-effect Image Restoration Algorithm

RAW images are digital images that have not undergone digital signal processing and contain richer structure and texture information.Joint demosaicing,denoising,and super-resolution tasks can perform multi-effect restoration on images,fully utilizing the information in RAW images.By improving image quality and increasing image detail and texture information while reducing image quality loss,this supports subsequent image processing and analysis tasks better.In recent years,scholars have begun to pay attention to the tasks.However,these scholars only discussed the feasibility of the tasks,lacking research on the characteristics of the three subtasks.Based on the above research,thesis focuses on the deep learning-based multi-effect image restoration algorithm,with the main work and contributions as follows:A multi-effect image restoration algorithm based on deep learning has been proposed.Thesis adopts the Unet network structure to capture global information and constructs a basic module through the attention mechanism to shield the influence of local noise.For demosaicing and super-resolution tasks,more efficient feature extraction modules and stronger supervised loss functions are introduced to maximize the extraction of image details.A fast multi-effect image restoration algorithm based on frequency domain decomposition has been proposed.Thesis decomposes the high-frequency information and the midlow frequency information of the image through the Laplacian pyramid,and constructs a dual-branch network to separately learn high-frequency and mid-low frequency information.In denoising and demosaicing tasks,the output is upsampled without parameters as the mid-low frequency branch information? in the super-resolution task,a lightweight structure is used to learn high-frequency information to supplement the texture details of the image.The contribution of thesis lies in proposing two multi-effect image restoration networks to fully utilize the information of RAW images,improve image quality and resolution,and provide better image input for high-level visual task applications.