带噪声水平评估的快速灵活盲深度降噪模型

Fast and Flexible Blind Deep Denoising Model with Noise Level Estimation

为提高快速灵活降噪网络(FFDNet)模型的降噪性能,建立一种噪声水平估计(NLE)模型,将其预测的噪声水平估计值输入FFDNet模型中,并以NLE模型作为FFDNet深度降噪模型的前置模块,使FFDNet转换为盲降噪模型。采用浅层卷积神经网络模型将噪声信号从噪声图像中分离出以得到噪声映射图,将噪声映射图的标准差作为噪声水平的初估值。鉴于噪声水平初估值与真值之间具有强相关性的特性,应用BP神经网络模型对噪声水平初估值进行修正。实验结果表明,该NLE模型与FFDNet降噪模型相结合后,降噪效果总体上与使用真实噪声水平值的FFDNet降噪模型接近,在多数噪声水平值下,两者的PSNR值相差小于0.1 dB,NLE模型的估计值可以达到与真实噪声水平值近似的效果,能够充分发挥FFDNet降噪模型的快速和灵活特性。

To improve the denoising performance of the Fast and Flexible Denoising Convolutional Neural Network(FFDNet),this paper proposes a Noise Level Estimation(NLE)model that estimates the level of noise.The estimation result is input into the FFDNet model,and the NLE model is taken as the preceding module of the FFDNet deep denoising model to transform it into a blind denoising model.Then the shallow convolutional neural network model is used to separate noise signals from noisy images to obtain the noise map,the standard deviation of which is taken as the initial estimated value of the noise level.Considering the fact that there exists strong correlation between the initial estimated value and ground-truths of the noise level,a Back-Propagation(BP)neural network model is used to correct the initial estimated value of noise level.Experimental results show that when the proposed NLE model works with the FFDNet model,its denoising performance is close to that of the FFDNet denoising model which uses the ground-truths of noise level.For most of the noise level values,the difference of Peak Signal to Noise Ratio(PSNR)values between the two models is within 0.1 dB,which means the estimation results of the proposed NLE model are similar to the ground-truths of noise level,bringing the fast and flexible characteristics of the FFDNet model into full play.