Research On Low Illumination Image Enhancement Algorithm Based On Deep Learning And Retinex Theory

As an effective information carrier,images are widely used in computer vision,civil photography,outdoor surveillance and other fields.Affected by inadequate lighting conditions at night and other factors,images captured in low-light environments often have problems such as low brightness,poor contrast,and loss of detailed information,which seriously affect people's information acquisition and subsequent computer vision system's performance.In the night environment,many security problems are prone to occur,such as traffic accidents,night theft,etc.The low-illuminance images taken by the equipment are not suitable for accident prevention and analysis;At the same time,in the field of national defense security,the images collected by night surveillance and surveys will also suffer from the influence of low light.In order to effectively improve the quality of low-il luminance images,this paper conducts research on low-illuminance image enhancement algorithms under the theoretical framework of deep learning and Retinex,in order to recover high-quality images with complete details,well lit,and full colors from low-illuminance images.The main research contents of this paper are as follows:1.This paper proposes a fully-supervised low-illuminance image enhancement algorithm based on illumination decomposition.Inspired by the Retinex theory,this paper builds an illumination decomposition network by fitting the existing illumination estimation algorithm to estimate the illumination and reflectance of low-illuminance images.On this basis,an estimation method of illumination mapping curve based on deep convolutional neural network is designed.By learning the mapping relantionship between paired images,a illumination map with increased brightness is obtained to enhance low-illuminance images.At the same time,in view of the problems of color distortion and noise amplification in the enhanced image,a color restoration and image denoising network is designed to achieve effective image post-processing and significantly improve image's quality and visual effects.Compared with the methods in the same period,the PSNR index on the classic dataset LOL has been improved up to 3.01dB,and the SSIM index has been improved up to 2.77%.2.On the basis of the above work,this paper further proposes a semi-supervised low-light image algorithm based on Retinex decomposition.The existing fully supervised learning method only learns the mapping function by reducing the error between image pixels,ignoring the subjective visual preference of the human eye for the image.Therefore,this algorithm is based on generative adversarial learning and introduces subjective visual information to obtain high vision pleasant image.First,a discriminator is designed to calculate the probability that the input image belongs to a high-quality image,and the unpaired high/low-quality image dataset is used to train the discriminator.Based on the trained discriminator,a.composite loss function is designed for the above-mentioned image enhancement network.The purpose is to enhance the image quality while satisfying the visual preference of the human eye as much as possible.Through continuous training of the discriminator and generator,the final discriminator cannot distinguish whether the input is a generated image or a high-quality image.The image generated by the generator is similar to the high-quality image,thereby improving the image quality.The semi-supervised low-illuminance image algorithm based on Retinex decomposition proposed in this paper can significantly improve the subjective visual effect of the image while ensuring the objective index of the enhanced image.Compared with the method of the same period,the improvement of the SSIM index on the classic dataset of LOL is 3.48%.