Low-light image enhancement is a challenging task that requires consideration of vari-ous complex issues,including brightness restoration,color distortion,and noise reduction.Simple ad-justments to the brightness of low-light images inevitably magnify these artifacts.To address these challenges,an attention guided low light enhancement network(AGNet)was proposed.AGNet consists of two parts:an attention guided network and a low-light enhancement network.The attention guided network is used to learn the illuminance-attention mapping in low-light images and applied in low-light enhancement network to guide the brightness enhancement and denoising tasks.The low-light enhancement network is composed of multi-scale convolution and residual blocks,which extract detail and texture features from low-light images through feature pyramid structure.Furthermore,a multi-scale color recalibration module(MCRM)is introduced in the network to further enhance the color and contrast of the output image.Experimental results demonstrate that AGNet not only outperforms other methods on major low-light datasets(LOL-v1 and LOL-v2-synthetic)in terms of objective met-rics(the PSNR of both datasets improved by 2.13/2.52),but it also has subjective advantages in com-parison.
维普
万方数据