Global-Local Generation Adversarial Learning Based Low-Light Image Enhancement

In the field of low-light image enhancement,the existing unsupervised methods still have some problems,such as the lack of authenticity and the unclear effect of image enhancement in extremely dark conditions.To address this issue,we propose a highly-effective unsupervised approach to directly recover normal light image from low-light image by designing an effective cyclic generation adversarial network,Firstly,in order to solve the problem of insufficient memory caused by large-size input,a global and local generator has been designed.Secondly,we develop an adaptive two-stage discriminator.Among it,by discriminating the whole image,the local region that needs to be discriminated again is obtained adaptively.Finally,the loss in frequency domain is employed to avoid image distortion.We employ the focal frequency loss which allows the model to adaptively focus on frequency components that are hard to synthesize.The PIindex of the method in this paper reaches on NPE,LIME,MEF,DICM and VV datasets respectively 2.81,2.78,2.40,3.15,3.69.On LOL datasets,PSNR and SSIM index reached 19.89 dB,0.7823,with good robustness.