基于深度卷积神经网络的图像去噪研究

摘要

In order to remove the noise in the image more effectively,a method based on deep learning is proposed.The method uses the integrated image as the input and output of the network,and uses hidden layer to compose a nonlinear mapping from the noisy image to denoised image.The network has a symmetric network structure consisting of convolution subnet and deconvolution subnet.Convolution subnet learns about image features,and the deconvolution subnet recovers the original image on the basis of characteristic graph and obtains more texture details combining rectified linear unit.This method uses the VOC2012 data set as the training set and Tensorflow to train the network model in the GPU environment.Experimental result demonstrates that this method can remove the noise in the image more effectively and obtain a higher Peak Signal to Noise Ratio(PSNR).In addition,it takes shorter time,and has better visual effect and perfect practicability.%为更有效地去除图像中的噪声,提出一种深度学习的图像去噪方法.以完整图像作为网络的输入及输出图像,通过隐含层构成由含噪声图像到去噪图像的非线性映射,由卷积子网和反卷积子网构成一种对称式的网络结构,卷积子网学习图像特征,反卷积子网根据特征图恢复原始图像,并结合修正线性单元获取更多的纹理细节.以VOC2012数据集作为训练集,使用Tensorflow在GPU环境下训练网络模型.实验结果表明,与GSM,KSVD,CN2,MLP方法相比,该方法能更有效地去除图像中的噪声,获得更高的峰值信噪比,耗时更短,视觉效果更佳.