基于深度残差学习的彩色图像去噪研究

目的 当噪声存在时，尤其是等级相对较大的噪声，会导致彩色图像的视觉质量下降，为了有效去除噪声的同时使去噪后的图像有更好的视觉效果，提出一种基于深度残差学习的彩色图像去噪方法。方法 首先设计由多个残差单元模块组成的残差层，然后在每个残差单元模块之间添加跳跃连接，构成由噪声图像到去噪图像的非线性映射，并优化残差单元个数，使网络能学习到更多的图像细节特征，以提升网络的去噪性能，同时将每个残差单元模块中的激活函数提到卷积层前面，以加速网络收敛。结果 与常用去噪算法相比，文中方法在Kodak24和CBSD100数据集上的主观视觉打分MOS值以及客观指标(PSNR和SSIM)上，较其他方法有更好的效果。结论 提出的基于深度残差学习的彩色图像去噪方法能有效去除图像中的噪声，尤其是较严重的噪声，并取得了良好的视觉效果，表明该方法具有良好的去噪性能。

Color Image Denoising Based on Depth Residual Learning

When the noise exists, especially the relatively serious noise level, the visual quality of color image will be reduced. The work aims to propose a color image denoising method based on depth residual learning, in order to remove noise effectively and make the denoised image have better visual effect. Firstly, a residual layer consisting of several residual unit modules was designed, and then the skip connection was added between residual unit modules to form the non-linear mapping from noise image to denoised image. The number of residual units was optimized, so that the network could learn more image details to improve the denoising performance. At the same time, the activation function of each residual unit module is moved to the front of the convolution layer to accelerate the network convergence. Compared with common denoising algorithms, the proposed method had better effects in subjective visual score MOS values and objective indicators (PSNR and SSM) on Kodak24 and CBSD100 datasets. The proposed color image denoising method based on depth residual learning can effectively remove the noise in the image, especially when the noise is serious, and obtain satisfactory visual effect, which shows that the proposed method has good denoising performance.