结合非对称卷积与复合感受野结构的图像降噪方法

Image Noise Reduction Method Based on Asymmetric Convolution and Complex Receptive Field Structure

为提高图像降噪精度,在去除图像噪声的同时尽可能保留图像边缘细节,在对多种传统降噪算法以及现代卷积神经网络架构研究的基础上,结合非对称卷积与复合感受野结构,提出一种新的降噪卷积神经网络模型。该模型在多尺度上获得了不同感受野下的图像特征,能更好地学习含噪图像到降噪图像的端到端映射。非对称卷积减少了模型参数量,使其更易于训练与验证。同时,该模型使用的残差学习、批量规范化、ReLU激活函数可加快卷积神经网络的收敛速度并提高其降噪性能。实验结果表明,在标准测试Set12上对图像加入均值为0、标准差为25的高斯噪声进行测试,降噪图像的峰值信噪比均值高达30.64dB。与目前优秀的降噪模型相比,该模型降噪性能良好,适用于多种强度含噪图像的降噪工作。

In order to improve the precision of image noise reduction,and retain image edge details as much as possible while image noise is removed,Based on the research of various traditional noise reduction algorithms and modern convolutional neural network architecture,a new noise reduction convolutional neural network modelcombining asymmetric convolution and complex field structure is propose。The image features under different sensing fields are obtained on multiple scales,which can help us better learn the end to-end mapping from noise-containing image to noise-reducing image.Asymmetric convolution reduces the number of parameters of the model and makes it easier to train and verify the model.At the same time,residual learning,batch normalization and ReLU activation function are used to accelerate the convergence rate and improve the noise reduction performance of the convolutional neural network.The experimental results show that,on the open standard test Set12,Gaussian noise with mean value of 0 and standard deviation of 25 is added to the image,and the mean peak signal to noise ratio of the denoised image is as high as 30.64dB.Compared with today's excellent noise reduction models,this model showed good performance and is suitable for noise reduction of various intensity images.