摘要
目前,图像去噪任务的主流方法多关注常规图像的噪声去除,而在低照度条件下拍摄的图像含有更严重的噪声,这种富有挑战性的去噪问题亟待解决。实验发现,将现有图像去噪方法直接应用于低照度图像任务中,会存在难以抑制真实图像噪声、丢失图像暗部和亮部细节信息问题。针对这些问题,提出一种基于Transformer的编解码U型网络。首先,提出对每张真实的噪声图像采用暗阴影校正方法进行预处理,改变其空间上的不均匀性,从而降低真实噪声学习的复杂性;然后,每张预处理后的图像通过提出的U型网络进行图像去噪,提出的U型网络是在UNet架构基础上引入Transformer Block模块构建分层编解码器,使网络可以从全局特征变化角度出发,多尺度分析图像亮暗细节信息;此外,采用窗口增强自注意力模块代替全局自注意力,更好地捕捉上下文特征依赖关系;最后,在解码器阶段引入可变形卷积残差模块,进一步增强局部特征的提取能力,提高网络去噪性能。在低照度真实噪声图像数据集SID和ELD上,方法的PSNR和SSIM较其他主流方法的最优值提升了0.53 dB和0.004。主观结果表明,该方法能在有效抑制低照度图像噪声的同时,更好地恢复细节信息。
Image denoising focuses on removing noise from conventional images,but an image taken under low-light conditions contains more serious noise.It is urgent that solve this challenging denoising problem.According to the experimental results,the existing methods of image denoising lack the ability to suppress the real image noise,lose details from the dark and light parts of the image,and distort the color of the image when applied directly to the low-light image denoising task.Aiming at these problems,this paper proposes a U-shaped network based on the Transformer.First,this paper uses dark shadow correction to reduce the complexity of real noise learning by removing spatial inhomogeneity from the noise images.Then,the image is denoised by the proposed U-shaped network.Transformer Block is used to build a hierarchical codec based on the UNet architecture in our network.The network can analyze the light and dark details of the image at multiple scales from the perspective of global feature changes.Moreover,use window-enhanced self-attention modules rather than global self-attention modules to better capture contextual feature dependencies.To further enhance the ability of local feature extraction and to improve the denoising performance of the network,a deformable convolution residual module is introduced in the decoder stage.On the low-light real-noise image datasets SID and ELD,the PSNR and SSIM of this method improve by 0.53 dB and 0.004 compared with the optimal values of other mainstream methods.
出处
《工业控制计算机》
2023年第5期69-72,共4页
Industrial Control Computer
基金
上海市自然科学基金资助项目(19ZR1419100)。
