结合感受野优化与残差激励的图像去噪

Image Denoising Combined with Receptive Field Optimization and Residual Excitation

前馈去噪卷积神经网络(DnCNN)拥有较为优异的图像去噪表现。然而,仅采用3×3依次相连的卷积核使得感受野在浅层受限,无法获得图片更多的特征信息。针对这一不足,提出一种Concatenate’s Convolutional Neural Network(CCNN)图像去噪模型。新模型使用3×3、5×5、7×7和9×9并联卷积核组成一个Concatenate block,并且自适应确定最佳网络层数。为了使得网络的特征信息不容易丢失,以及保留图像的浅层信息,加入了浅层到深层的残差激励;加入了全局残差学习,可以让网络直接提取含噪图像中的噪声信息,使得网络去噪效果更加优秀;同时为了缓解梯度消失,加速网络收敛,使用批量规范化(Batch Normalization)和非线性激活函数(ReLU)。在公开数据集中对比其他优秀方法表明,论文提出的模型去噪效果更加明显。

Feedforward denoising convolutional neural network(DnCNN)has excellent image denoising performance.Howev-er,only the convolution kernel connected in sequence by 3×3 is used,so the receptive field is limited and more feature information of the picture cannot be obtained.Therefore,this paper proposes a Concatenate's convolutional neural network(CCNN)image de-noising model.The new model uses 3×3,5×5,7×7 and 9×9 parallel convolution kernel to form a Concatenate block.At the same time,and the optimal network layer number is determined adaptively.In order to make the feature information of the network not easy to lose,and to retain the shallow information of the image,a shallow to deep residual excitation is added.The addition of global residual learning allows the network to directly extract noise information in noisy images,making the network denoising effect more excellent.Meanwhile,in order to alleviate the disappearance of gradients and accelerate network convergence,batch normalization(BN)and nonlinear activation function(ReLU)are added.Compared with other excellent methods in a public data set,the pro-posed image denoising model in this paper is more effective.