自适应深度残差椒盐噪声滤除算法

为了在去除图像中椒盐噪声的同时最大程度地避免产生色彩失真与边缘模糊等瑕疵, 提出基于深度残差网络的椒盐噪声自适应滤除算法. 将图像去噪分解为 2 步. 首先, 为了让网络模型能够处理不同尺度密度的椒盐噪声, 提高网络模型的鲁棒性, 先对图像进行自适应预处理以去除高频信息; 其次, 构建深度残差网络模型, 训练出能将预处理后的图像映射到干净图像的函数. 大量实验结果表明, 文中算法不仅在保留图像边缘细节和去除高密度椒盐噪声方面均优于传统和基于机器学习的椒盐噪声去除技术, 可有效地避免出现色彩失真和条纹等瑕疵. 同时, 其在BSD300 数据集上去噪效果优于其他算法.

To remove salt-and-pepper noise with minimal degradation (e.g., edge blurring, color deviation, and stripe) of image intrinsic properties, we present an adaptive salt-and-pepper denoising method based on a deep residual network. The main idea of this paper is to simplify image denoising into two steps. Firstly, in order to enable the network model to handle different-densities salt-and-pepper noises and improve the robustness of the network model, we remove the high frequency information by adaptive windows as a first step. Secondly, we construct an effective deep residual network model to train a function which can map the pre-processed images to their corresponding ground truths. Qualitative and quantitative experiments show that not only can our method avoid problems such as color distortions and streaks, but also our method outperforms the state-of-the-art learning-based and traditional approaches, in terms of both handing inputs with different levels of noises and revealing high-fidelity image edges. Meanwhile, the performance on BSD300 evaluated in PSNR shows superiority over the competitors.