| Image restoration includes various tasks such as image denoising,image deraining,and low-light image enhancement.Traditional image restoration algorithms are based on a large number of artificially designed features,which have many problems such as fixed features and poor universality.Image algorithms based on deep learning can not only make up for the lack of universality of traditional methods,but also improve performance.Starting from the "models" and "datasets" involved in deep learning algorithms,this paper studies image restoration tasks from two perspectives.First,from the "model" perspective,this paper proposes an image restoration method based on QTP loss.Since supervised methods suffer from the domain transfer problem,unsupervised methods are more suitable for handling real-world image restoration tasks.However,unsupervised image restoration methods often suffer from three problems:pseudo color or poor image quality,insufficient restoration,and loss of semantic information due to content distortion in images before and after restoration.In response to the above problems,this paper proposes a new hybrid loss function,namely the quality,task and perception loss function QTP.The quality part solves the pseudo color and poor image quality by enhancing the image quality,make the quality score of the restored image and its not strictly paired reference clean image as close as possible;the task part proposes to apply the task probability network to solve the insufficient restoration effect,and the unsupervised image The restoration is transformed into a supervised classification problem,and the task probability network is proposed in this paper;the perception part deals with the problem of missing semantic information by ensuring the multi-scale consistency of the Fourier phase spectrum before and after image restoration.In the experimental part,this paper conducts both unsupervised and supervised experiments for the three problems of low-light image enhancement,image deraining and image dehazing.The experimental results demonstrate the effectiveness of our QTP method.Secondly,from the "dataset" perspective,this paper proposes a two-stage dualQTP unsupervised low-light image enhancement method（DQTP）.In the unsupervised domain,enhancement methods based on pseudo low-light images are very effective.However,due to the lack of specific constraints on low-light images,there is a gap between these pseudo low-light images and real low-light images.This problem greatly affects pseudo low-light images enhancement methods for low-light images.In response to the above problems,this paper proposes the DQTP method.In the first stage,this paper proposes a new prior theory for the low-light image enhancement task.The comprehensive consistency of the dual space,that is,assuming the pseudo-low light learned in the low-light image space The image should fully match a real low-light image.The daytime image enhanced by the pseudo low-light image in the enhanced image space should be consistent with the daytime image enhanced by the real lowlight image,and the above two pairs of consistency constitute a dual space symmetric structure.Then in the first step to generate pseudo-low-light images,the pair of QTP loss functions are embedded in the dual space to constrain the comprehensive consistency of the network;In the second stage,this paper adds additional Supervised constraints to optimize the performance of unsupervised augmentation networks.Finally,comprehensive experiments on both supervised and unsupervised datasets demonstrate the superiority of the DQTP algorithm.In addition to algorithm part,this paper also deploys the above restoration algorithm on the cloud platform from an engineering point of view to realize user terminal access and online invocation of image restoration services,so that the deep learning algorithm can be implemented. |
