| Low-light images suffer from low brightness,low contrast,and unclear details.Limited by shooting conditions,there are plenty of lowlight images in daily production and life.Low-light image enhancement is significant.On the one hand,it can improve the subjective perception of low-light images;On the other hand,it can be treated as a preprocessing step for downstream computer vision tasks,such as face recognition and object detection.In recent years,deep learning-based low-light image enhancement has made significant progress,especially the self-regularized learning method has become a trending research topic because only lowlight images are required in the training process,which reduces the cost of acquiring training datasets.However,existing methods suffer color deviation and fail to adapt to various lighting conditions.To solve the above problems,this thesis proposes a self-regularized low-light image enhancement algorithm based on Retinex.The work of this thesis includes three points:(1)Aiming at the problem of color deviation,this thesis designs a novel low-light image enhancement network based on the HSV color space.As a result,the brightness information(Value)is enhanced separately while all chromaticity information(Hue,Saturation)is preserved,reducing color deviation.(2)Aiming at the problem of poor generalization performance to brightness,this thesis proposes to treat the reflectance component,which is irrelevant to various illumination based on Retinex theory,as the enhanced value,using the insensitivity of the reflectance component to illumination to improve the generalization performance to brightness.Therefore,the brightness generalization problem can be transformed into the reflectance component estimation problem.(3)The above research work has derived a new problem:it is impossible to distinguish the relationship between the estimation result of a single value and the reflection component.This thesis designs a novel self-regularized reflectance component estimation model.The model first uses a random disturbance algorithm to generate another value of the same scene and then estimates the reflectance component shared in the original form and the disturbed form of value by the reflectance estimation network.Extensive experiments demonstrate that the proposed algorithm proposed has achieved good results quantitatively and qualitatively.Specifically,the proposed method surpasses the mainstream selfregularized low-light image enhancement method by 0.72dB in the PSNR index. |
