Study On Retinex Model-based Low Light Image Enhancement Algorithms

Due to insufficient photons and low signal-to-noise ratio in low-light environment,the captured image signal has serious quality problems such as low brightness and contrast,color deviation,and noise interference.These problems can affect the visual subjective perception while making it difficult to extract information from images effectively.Therefore,low-light image enhancement has become a critical and challenging topic in computer vision and multimedia applications.To address the above problems,this paper takes low-light images as the research object,explores the effects of the illumination and reflectance components in the Retinex model on image enhancement from the Retinex decomposition,and proposes three low-light image enhancement algorithms to solve the problems faced by existing algorithms.The main contributions and innovations of this thesis are summarized as follows:1.The low brightness and contrast of low-light images leads to the inconspicuous structure and texture information of the low-light image.To solve this problem,a structure and texture revealing Retinex model is proposed.By fully considering the difference of the distribution of the illumination component and the reflectance component in the gradient field,the distribution difference is utilized to construct local relative total variation-based structure-aware constraints and texture-aware constraints.The constrains aim to improve the spatial piece-wise smoothness of the illumination component and the piece-wise continuity of the reflectance component,and improves the perceive ability to structure and texture information of the model.2.Low-light images are affected by the weak scene light and the performance limitations of electronic sensors when capturing the image,so the low-light image will inevitably be disturbed by noise.Aiming at this problem,a constraint low-rank approximation Retinex model is proposed.The nuclear norm constraint item is introduced into the variational model for reflectance component.Besides,the nuclear norm minimization method is utilized to solve the low-rank matrix approximation problem of the image to reduce the effect of the noise.The illumination component and reflectance component will affect each other in the iterative solving process,resulting in noise amplification problem.To address this problem,the alternating direction multiplier method is introduced to separate the estimation problem of the illumination component and the reflectance component into two separate problems to avoid the noise amplification.3.Deep learning-based low-light image enhancement methods deeply rely on training datasets with normal light condition,but the collection of training datasets limits the performance of deep neural network models.To solve this problem,an unsupervised deep learning network architecture based on the Retinex model is proposed.The network has two U-Net branches,which are used to estimate the reflectance component and the illumination component respectively.Since the reflectance component contains more image information than the illumination component,a deeper network structure is adopted in the reflectance component estimation branch,and several squeeze-and-excitation modules are introduced to enhance the details of the image.It aims to weaken the background noise and make the final output feature map contain more useful semantic information.