基于残差连接的并行网络去噪

Parallel Network Denoising Method Based on Residual Connection

针对大部分的深度卷积神经网络存在着层次难以达到很深以及神经网络的深层次性能退化问题,论文基于前馈去噪卷积神经网络的模型,提出一种结合批量重整化去噪网络中的并行网络。新方法是将原有的网络层一分为二,通过增加网络的宽度而不是深度深层结构达到获取更多特征的目的。再使用残差学习和批量归一化的方法,以改善去噪质量,并且加速训练。结果表明,相比于当前比较成熟的去噪算法,论文提出的残差连接的并行网络去噪新算法的去噪效果更为突出,去噪耗时也大大降低。

Aiming at remedy the defeat that most deep convolutional neural networks have the problem of deep level and deep degradation of performance,a parallel network method of feedforward denoising convolutional neural network combined with batch renormalization denoising network is proposed.The new method is to seperate the primitive network layer into two parts,and obtain more features by increasing the width of the network rather than the depth deep structure to achieve the purpose of image denoising.Then the residual learning and batch normalization methods are used to improve the denoising quality and accelerate the training.The results show that,compared with the current mature denoising algorithms,the new method has more prominent denoising effect and greatly reduces the denoising time.