科技创新2030 “新一代人工智能”重大项目(2018AAA0102001); 国家自然科学基金(61922043, 61872421)
现存的图像去噪算法在处理加性高斯白噪声上已经取得令人满意的效果, 然而其在未知噪声强度的真实噪声图像上泛化性能较差. 鉴于深度卷积神经网络极大地促进了图像盲去噪技术的发展, 针对真实噪声图像提出一种基于自监督约束的双尺度真实图像盲去噪算法. 首先, 所提算法借助小尺度网络分支得到的初步去噪结果为大尺度分支的图像去噪提供额外的有用信息, 以帮助后者实现良好的去噪效果. 其次, 用于去噪的网络模型由噪声估计子网络和图像非盲去噪子网络构成, 其中噪声估计子网络用于预测输入图像的噪声强度, 非盲去噪子网络则在所预测的噪声强度指导下进行图像去噪. 鉴于真实噪声图像通常缺少对应的清晰图像作为标签, 提出了一种基于全变分先验的边缘保持自监督约束和一个基于图像背景一致性的背景自监督约束, 前者可通过调节平滑参数将网络泛化到不同的真实噪声数据集上并取得良好的无监督去噪效果, 后者则可借助多尺度高斯模糊图像之间的差异信息辅助双尺度网络完成去噪. 此外, 还提出一种新颖的结构相似性注意力机制, 用于引导网络关注图像中微小的结构细节, 以便复原出纹理细节更加清晰的真实去噪图像. 相关实验结果表明在SIDD, DND和Nam这3个真实基准数据集上, 所提的基于自监督的双尺度盲去噪算法无论在视觉效果上还是在量化指标上均优于多种有监督图像去噪方法, 且泛化性能也得到了较为明显的提升.
Existing image denoising algorithms have achieved decent performance on the images with the additive white Gaussian noise (AWGN), while their generalization ability is not good on real-world images with unknown noise. Motivated by the significant progress of deep convolution neural networks (CNNs) in image denoising, a novel two-scale blind image denoising algorithm is proposed based on self-supervised constraints. Firstly, the proposed algorithm leverages the denoised results from the small-scale network branch to provide additional useful information for the large-scale image denoising, so as to achieve favorable denoised results. Secondly, the used network is composed of a noise estimation subnetwork and an image non-blind denoising subnetwork. The noise estimation subnetwork is firstly used to predict noise map, and then image denoising is carried out through the non-blind denoising subnetwork under the guidance of the corresponding noise map. In view of the fact that the real noise image lacks the corresponding clean image as the label, an edge-preserving self-supervised constraint is proposed based on the total variation (TV) priori, which generalizes the network to different real noisy datasets by adjusting smoothing parameters. To keep the consistency of the image background, a background guidance module (BGM) is proposed, which builds a self-supervised constraint based on the information difference between multi-scale Gaussian blurred images and thus assists the network to complete image denoising. In addition, the structural similarity attention mechanism (SAM) is proposed to guide the network to pay attention to trivial structural details in images, so as to recover real denoised images with cleaner texture details. The relevant experimental results on the SIDD, DND, and Nam benchmarks indicate that the proposed self-supervised blind denoising algorithm is superior to some deep supervised denoising methods, and the generalization performance of the network is improved significantly.