Research On Low-light Image Enhancement Algorithm Using Attention Layer And GAN

With the popularization of image devices and vision systems,the use of intelligent vision algorithms such as autonomous driving and target detection has become increasingly widespread.However,in dark scenes such as night,the low-light images collected by the device have uneven surface illumination and low signal-to-noise ratio,which seriously affects the accuracy of the visual algorithm.To improve this situation,the enhancement of low-light images has attracted more and more attention.Traditional low illumination enhancement algorithms improve the surface illumination of objects by calculating pixel illumination values and building enhancement models.Current research shows that the Retinex imaging model have better ability to capture the lighting distribution of low-light images,but the local enhancement details.In recent years,researcher combine the Retinex model and deep learning method.The deep Retinex algorithm improve the efficiency of the low-light enhanced algorithm.After summarizing and analyzing the related low-light enhancement algorithms,the author have a breakthrough that due to lack the spatial structure feature extraction,the existing enhancement algorithms tend to produce Pictures with poor enhancement.In this article,integrating the attention mechanism and the generation confrontation,the AM-Retinex Net and AM-Retinex GAN algorithms are proposed.As the subsequent experiments show,a good enhancement effect is obtained.The AM-Retinex Net algorithm introduces a lighting map,which is an enhanced network that can perform "two-stage" adjustments to the enhancement process.In the first stage of the AM-Retinex Net network,the low-light image is processed through the decomposition network to obtain a constant reflection map of the object and a slowly changing and smooth illumination map.In the next stage,the lighting map extracts spatial and texture information through the attention layer.Similarly,in the AM-Retinex Net training process,a low-light image synthesis method is used to expand the amount of data used in the training set.A new loss function is constructed,and a color loss function that guides the convergence of contrast is added.The final experimental data shows that AM-Retinex Net can improve the distortion,color spots,and other phenomena in the LOL and SID experiment sets;in terms of enhancing the evaluation index,the PSNR index increased by 15% and the SSIM index increased by9.5%.However,the training process of the AM-Retinex Net algorithm have signs to heavy depends on the paired low-light data.The collection of paired low-light data requires a lot of manpower.In some special cases,there are not easy to collect paired low-light data.Toovercome these shortcomings,AM-Retinex GAN that take advantage of effective and unsupervised adversarial networks is proposed in Cha.4 of this paper.AM-Retinex GAN uses adversarial technology effectively train the deep network without paired low-light data.Among a GAN network,the generator part of AM-Retinex GAN uses the attention mechanism to generate normal-light enhanced images;the discriminator part uses a global-local discriminator structure to distinguish the generated image from the real image as much as possible;the generator and discriminator continue to play the game Finally,the generator continuously optimizes the network weights to meet the enhancement results close to normal illuminance.In the AM-Retinex GAN loss function section,this paper adds self-normalized perceptual loss fusion.The final experiments on the LOL and SID experiment datasets show that AM-Retinex GAN is superior to the benchmark algorithm in both subjective and objective visual assessment.Also,experiments conducted on the BDD100 K dataset show that AM-Retinex GAN can be better used in conjunction with autonomous driving tasks to improve the accuracy of autonomous driving in low-light environments.