阶梯式图像去噪方法

为了提高对原始红绿蓝（RAW）和标准红绿蓝（sRGB）2种格式的真实噪声图像的去噪性能，提出了一种基于卷积神经网络的阶梯式图像去噪方法。第一阶段，利用单个通道内空间结构信息对图像各个通道独立去噪，获得初步去噪结果。第二阶段，利用噪声图像在不同通道的相关性进一步去噪，获得增强的去噪结果。在所提方法中引入误差反馈机制来减少采样带来的信息损失；使用密集残差连接模块使得提取到的噪声图像特征能更有效地复用和传播；利用通道注意力使得网络有选择性地增强信息量大的特征，抑制无用特征。将所提方法与常用的其他去噪方法比较，实验结果表明，在达姆施塔特噪声数据集的RAW/sRGB数据集上，所提方法分别达到了49.55 dB和39.55 dB的峰值信噪比（PSNR）；在跨通道数据集达到了39.52 dB的PSNR，较目前绝大多数方法具有更好的性能。

A Ladder-Type Denoising Method

A ladder-type denoising method is proposed to improve the denoising performance for raw red green blue(RAW) and standard red green blue(sRGB) real-world images. In the first stage, each channel of the noisy image is denoised separately utilizing intra-channel structure information. In the second stage, the inter-channel correlation information of the noisy image is utilized to further denoise the whole image, and the final boosted denoising result is obtained. Error feedback mechanism is introduced to reduce the information loss caused by sampling. Additionally residual dense connection makes features more effective for reuse and propagation; channel attention selectively enhances features with large amount of information and suppress useless features. The proposed method is compared with other denoising algorithms, and the results show that the proposed method achieves 49.55 dB peak signal to noise ratio(PSNR) in RAW images and 39.55 dB PSNR in sRGB images on Darmstadt noise dataset, and 39.52 dB PSNR on cross-channel dataset, which realizes competitive performance in comparison with other denoising algorithms.