针对超分辨率卷积神经网络(SRCNN)卷积层较少、训练时间长、不易收敛且表达和泛化能力受限等问题,提出了一种残差反卷积SRCNN(RD-SRCNN)算法。首先利用不同大小的卷积核进行卷积操作,以更好地提取低分辨率图像中的细节特征;然后将获取的图像特征输入由不同大小卷积核构成的卷积层和指数线性单元激活层组成的残差网络,并通过短路径连接各个特征提取单元,以解决梯度消失、实现特征重用、减少网络冗余;最后,通过加入反卷积层增大感受野,得到清晰的高分辨率图像。实验结果表明,RD-SRCNN算法在视觉和客观评价标准上均取得了较好的效果。

Aiming at the problems of super-resolution convolutional neural network (SRCNN) with fewer convolutional layers, long training time, difficulty in convergence, and limited expression and generalization capabilities, a residual deconvolution SRCNN (RD-SRCNN) algorithm is proposed in this work. First, different size convolution kernels are used for convolution operation to better extract the detailed features in low resolution images. Then, the acquired image features are input into the residual network composed of convolution layer composed of convolution kernels of different sizes and activation layer of exponential linear unit, and each feature extraction unit is connected by short path to solve the problem of gradient disappearance and realize the feature reuse, and reduce the network redundancy. Finally, a clear high-resolution image is obtained by adding a deconvolution layer to increase the receptive field. Experimental results show that the RD-SRCNN algorithm achieves good results in both visual and objective evaluation criteria.