基于边缘融合与监督-密集块特征提取去噪网络设计

Design of a denoising network based on edge fusion and supervised-dense block feature extraction

去噪卷积神经网络(denoising convolutional neural network,DnCNN)在去噪方面展现出了优异的性能,但仍存在着过度平滑和细节丢失的问题。针对此问题,提出一种图像边缘融合与监督-密集块充分特征提取的方法,该方法在主干网络中使用叠加的密集块对图像特征进行充分地提取;在辅助网络中通过提取含噪图像的边缘信息将其融合在主干去噪网络中并分阶段监督主干网络促使主干网络在去噪的同时更好的保留图像细节,提高去噪图像的成像质量。实验证明在噪声水平为25的条件下,模型在Set12数据集上的平均峰值信噪比(PSNR)和结构相似性(SSIM)比传统的DnCNN模型分别高出0.14 dB和0.011。同时该模型还可以用来训练去除高斯盲噪声和脉冲盲噪声,使用高斯盲噪声去噪网络去除噪声水平为25的高斯噪声,在Set12数据集上PSNR和SSIM值较DnCNN-B网络分别提升0.16 dB和0.005;使用脉冲盲噪声去噪网络去除10%的脉冲噪声在Set12数据集上PSNR值和SSIM值分别可以达到37.16 dB和0.960。模型在去除噪声的同时还能尽可能多的保留图像的细节。

The denoising convolutional neural network(DnCNN)has shown remarkable performance in denoising,but there are still issues of excessive smoothing and loss of details.To address this problem,this paper proposes a method for image edge fusion and stage-wise supervision-dense block sufficient feature extraction.This method uses stacked dense blocks to fully extract the features of the image in the main network;in the auxiliary network,it fuses the edge information of the noisy image by extracting it and supervises the main denoising network in stages to better preserve the image details while denoising,and improve the imaging quality of the denoised image.Experimental results show that under the condition of a noise level of 25,the model's average peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)on the Set12 dataset are 0.14 dB and 0.011 higher than those of the traditional DnCNN model,respectively.At the same time,this model can also be used to train to remove Gaussian blind noise and pulse blind noise.When using the Gaussian blind noise denoising network to remove Gaussian noise with a noise level of 25,the PSNR and SSIM values on the Set12 dataset are improved by 0.16 dB and 0.005,respectively,compared to the DnCNN-B network.When using the pulse blind noise denoising network to remove 10%of the pulse noise,the PSNR and SSIM values on the Set12 dataset can reach 37.16 dB and 0.960,respectively.This model can remove noise while preserving as much detail in the image as possible.