基于改进的卷积神经网络的图像超分辨率重建

Image super-resolution reconstruction based on improved convolutional neural networks

利用卷积神经网络技术对单幅低分辨率图像进行超分辨率重建。利用五层卷积的网络模型用以特征提取和端到端的低分辨率与高分辨率之间的映射;使用自适应矩估计优化算法加快网络的收敛速度;将ReLU激活函数修改为Leaky ReLU激活函数,解决遇到导数为0时的导致神经元不能进行参数更新的问题,同时调整卷积核大小以及数目。提出算法在Set5和Set14数据集上进行实验验证,并与Bicubic、ScSR、SR_NE_ANR、SRCNN等主流方法进行对比,实验结果表明,该方法在重建精度和收敛速度等方面都有很好的效果。

Convolution neural network was used to reconstruct the single low resolution image.A five-layer convolution network model was proposed for feature extraction and mapping between end-to-end low resolution and high resolution.The adaptive moment estimation optimization algorithm was used to accelerate the convergence speed of the network.The ReLU activation function was modified to Leaky ReLU activation function to solve the problem that neurons can not update parameters when the derivative is zero,and the size and number of convolution nuclei were adjusted.Results of experiments on Set5 and Set14 datasets show that the proposed method has good performance in reconstruction accuracy and convergence speed,compared with Bicubic,ScSR,SR_NE_ANR,SRCNN and other mainstream methods.