基于双重注意力残差模块的低照度图像增强

Low-light Image Enhancement Based on Dual Attention Residual Blocks

低照度图像增强(Low Light Image Enhancement,LLIE)是将光照不足条件下获取的图像恢复成正常曝光的图像,基于深度学习的LLIE算法常用堆叠卷积或上/下采样的方式设计,这样缺少相关语义信息的指导,导致增强后的图像存在噪声增多、色彩失真、细节丢失等问题。为此,本文提出一种基于双重注意力残差模块的LLIE算法。该算法提出融合双重注意力单元的残差模块(Dual Attention Residual Block,DA-ResBlock),在通道域和空间域提供的语义信息引导下,通过多级串联的DA-ResBlock对有效特征进行稳定提取,并且使用跳跃链接与卷积神经网络来恢复图像细节信息。此外,使用复合损失函数对增强任务进行约束。最后,在2个真实图像的公共数据集上与近几年主流算法进行对比。实验结果表明,本文算法在主观视觉上在有效提高图像亮度的同时,更好地抑制了噪声、恢复了图像色彩与细节纹理,客观评价上在PSNR、SSIM、LPIPS这3个指标上均优于所对比的主流算法。

Low Light Image Enhancement(LLIE),which is to restore images captured under insufficient lighting conditions to normal exposure images.The existing LLIE algorithms based on deep learning often use stacked convolution or up/down sampling methods,which lacks the guidance of relevant semantic information,resulting in problems such as increased noise,color distor⁃tion and detail loss in the enhanced image.To address this issue,a novel LLIE algorithm based on dual attention residual mod⁃ules is proposed.This algorithm proposes a residual block that integrates dual attention units(Dual Attention Residual Block,DA-ResBlock),which provides semantic information guidance in both channel and spatial domains.Through multi-level cas⁃caded DA-ResBlocks,effective features are stably extracted,and skip connections and convolutional neural networks are used to restore image detail information.In addition,a composite loss function is used to constrain the enhancement task.Finally,we compare our algorithm with mainstream algorithms in recent years on two public datasets that provide real images.The experimen⁃tal results show that the proposed algorithm effectively improves image brightness while better suppressing noise,restoring image color and detail texture in subjective vision.In the objective evaluation,the three indexes of PSNR,SSIM and LPIPS are supe⁃rior to the compared mainstream algorithms.