基于DBAFFNet的低照度图像增强

Low-Light Image Enhancement Based on DBAFFNet

针对当前低照度图像增强后存在色偏、细节损失和噪声放大的问题,提出了基于双分支自适应特征融合网络的低照度图像增强方法。首先,设计自适应特征融合模块,在深层特征中融合更多细节和颜色信息;其次,构建通道及空间注意力模块,使网络着重于图像细节和颜色的恢复;最后,根据Retinex理论设计Poisson-Retinex损失函数,抑制图像的噪声,从而提高图像的增强效果。在多个数据集上的主观和客观对比结果表明,所提方法不仅能恢复增强图像的颜色和细节,而且能更好地抑制噪声,从而获得良好的增强效果。

To solve the problems of color cast,detail loss and noise amplification after lowlight image enhancement,an improved low-light image enhancement method is proposed based on dual-branch adaptive feature fusion network(DBAFFNet).Firstly,the adaptive feature fusion(AFF)module is designed to fuse more details and color information into deep features.Secondly,the channel and spatial attention(CASA)module is established to focus on the restoration of image details and color.Finally,a Poisson-Retinex loss function based on Retinex theory is designed to suppress noise of the image,thereby improving the enhancement effect of images.The results of subjective and objective comparisons on multiple datasets demonstrate that the proposed method not only restores the color and details of the enhanced image,but also suppress the noise better,and achievea good enhancement effect.