Research And Application Of Intrinsic Image Decomposition Algorithms Based On Retinex

Intrinsic image decomposition is a challenging task in the area of computer vision,which aims at recovering intrinsic components from the observation.Since each component represents a different physical element that benefits many image processing tasks,intrinsic image decomposition has been studied extensively.Traditional methods usually construct the regularization terms based on prior knowledge in energy models to solve this problem.Though they provide theoretical guarantee,their calculation procedures often consume much resource.Recently,network-based approaches have achieved good results in runtime and quality.However,the network structures of these methods are usually designed heuristically.This kind of model is like a black box,and its generalization ability is limited by data.To overcome these limitations,this paper proposes two kinds of methods from different perspectives.The first method considers combining the advantages of deep learning and traditional optimization,which first designs the energy model according to Retinex theory,and then replaces the solution of prior terms in the unrolled optimization process with the deep prior from network.So the performance and speed of model can be improved while the theoretical guarantee is provided.Although the first method shows good performance,some of its parameters require manual adjustment.In the second method,this paper trains an end-to-end network called ERDNet to solve the task of intrinsic image decomposition.The overall structure of ERDNet is designed based on Retinex theory.The architecture of its basic module ERDB refers to and absorbs the advantages of different network structures in the field of deep learning,so that it can make full use of the hierarchical features of image.At the same time,this paper adds reconstruction loss in the design of loss function,which makes the whole model subject to the physical imaging law,so as to ensure the consistency of the observed images and reconstruction results.In general,based on the Retinex theory and deep learning technology,this paper proposes two kinds of algorithms for intrinsic image decomposition from different perspectives,including the traditional optimization and deep learning.In addition,this paper successfully apply our algorithm to low-light image enhancement according to the relationship between tasks.Experimental results show that the proposed models have better numerical results and visual effects compared with other state-of-the-art methods.