Image denoising under low signal-to-noise ratio (SNR) and non-Gaussian noise is still a challenging problem in image processing. In this study, the authors prose a kind of improved convolution neural network auto-encoders for image denoising. Different from other priors based methods, the denoising auto-encoders (DAEs) can learn end-to-end mappings from noisy images to the target ones. This study research statistical features of image residual between the restored images and target images. According to the maximum entropy principle, the training loss function of the ordinary DAEs was modified with residual statistics as the constraint condition, and an improved training algorithm was proposed based on augmented Lagrange function method. Thus, the quality of restored image can be improved through removing image information from residual more efficiently. Experiments show not only the denoising effects of improved DAEs is superior to the original mean-square-error loss function DAEs in both peak SNR and Riesz feature similarity metric indexes, but also has the ability to suppress the different types of noises with different levels through a single model.

中文翻译：

---

基于残留熵最大值的图像DAE

低信噪比（SNR）和非高斯噪声下的图像去噪仍然是图像处理中的一个难题。在这项研究中，作者提出了一种用于图像去噪的改进的卷积神经网络自动编码器。与其他基于先验的方法不同，去噪自动编码器（DAE）可以学习从嘈杂图像到目标图像的端到端映射。本研究研究了恢复图像和目标图像之间图像残留的统计特征。根据最大熵原理，以残差统计为约束条件，对普通DAE的训练损失函数进行了修正，提出了一种基于增强拉格朗日函数法的改进训练算法。从而，通过更有效地从残差中去除图像信息，可以提高恢复图像的质量。实验表明，改进后的DAE的去噪效果不仅在峰值SNR和Riesz特征相似性指标上均优于原始均方误差损失函数DAE，而且还具有通过抑制不同级别的不同类型噪声的能力一个模型。