Research On Low-light Image Enhancement Based On Deep Learning

Poor lighting causes a variety of problems in photographs,including general darkness,features that are concealed by low light,color distortion,low contrast,low signal-to-noise ratio,and restricted dynamic range.Due to their poor visual quality,these photos make it difficult to see objects and conduct in-depth analysis.Although low-light image enhancement has seen several advances to date,there are still certain difficulties and inadequacies in practical applications.To research low-light picture improvement,this study integrates deep learning techniques to address the issues of low brightness,missing dark features,significant noise,and color distortion in low-light photographs as a result of insufficient or uneven light.The details of the research project are as follows:We propose an image enhancement method based on the Attention U-Net Dual Discriminator-Generative Adversarial Networks(AUDD-GAN)with a generator incorporating convolutional attention modules and dual discriminators.First,the U-Net network is improved using the Convolutional Block Attention Module(CBAM)and serves as a generator to extract both shallow and deep features to obtain richer detail information,effectively avoiding image distortion caused by overexposure or underexposure.Next,a dual discriminator is employed,with Patch GAN as the local discriminator,demonstrating excellent enhancement performance for images with insufficient local lighting.Finally,the loss function is improved,and multiple loss functions are combined as a joint loss function.Experimental results show that,compared to existing algorithms,the AUDD-GAN algorithm effectively improves the brightness and contrast of low-light images,providing a more natural visual effect.Based on the AUDD-GAN algorithm,we propose a low-light image enhancement method combining Retinex-Net and AUDD-GAN based on Retinex theory is proposed.First,the low-light image is decomposed into illumination and reflection components.Then,the adaptive learning of image noise distribution characteristics is achieved through a convolutional neural network,removing the noise from the reflection map and further improving the image quality.Simultaneously,a U-Net model with the CBAM is added to the light map to better catch the image’s details and texture information.Finally,the enhanced image is obtained by synthesizing the adjusted illumination and reflection components,which are then trained using discriminative network adversarial learning.Experimental results show that this algorithm not only improves the denoising performance but also effectively enhances the brightness and contrast,making the processed image largely resemble an image taken under normal lighting conditions.