基于U-Net融合的保留纹理的图像去噪方法

图像的噪声阻碍了高级视觉任务对图像的理解,且去除图像的噪声是一个具有挑战性的任务.现有的基于卷积神经网络的图像去噪方法在去除噪声的同时,对图像纹理会引入一定程度的破坏,导致去噪后图像无法保留图像的纹理.为了解决这个问题,本文提出一种用二分支U-Net网络来融合特征和保留纹理的图像去噪方法.首先选取一种去噪方法的两个不同去噪参数的预训练模型分别得到同一张噪声图像的不同去噪结果,其中一个结果中去噪效果比纹理保留效果好,另一个结果中纹理保留比去噪效果好.然后将这两个去噪图像作为卷积神经网络的输入,利用两个编码器分别提取图像的特征,并同时放入融合模块融合图像的特征,最后利用解码器重建出无噪声图像.实验结果表明,与现有的方法相比本文的方法更有效,在去除噪声的同时能保留更多的图像纹理信息.

Texture-preserving Image Denoising Method Based on U-Net Fusion

Image noise hinders understanding of images in high-level vision tasks,and image denoising is a challenging task.Existing image denoising methods based on convolutional neural networks remove image noise while destroying the image texture to a certain extent,resulting in the image being unable to retain the image texture after denoising.In order to solve this problem,this paper proposes an image denoising method using two-branch U-Net network to fuses features and preserve texture.First,two pre-trained models with different denoising parameters of a denoising method are selected to produce two different denoising results of the same noisy image,one of which has a better denoising effect than the texture preserving effect,while the other result has a better texture preserving effect than the denoising effect.Second,the two denoising results are used as the input of the proposed convolutional neural network.Specifically,two encoders extract image features from the two inputs,and a fusion module fuses the extracted features.Finally,the decoder reconstructs a noise-free image from the fused features.Experimental results show that the proposed method is more effective than the existing method,and can retain more texture information while removing noise.