针对现有的基于样本学习的图像超分辨率方法参数较多、运算速度较慢等问题, 结合基于卷积神经网络的超分辨率方法, 提出一种快速图像超分辨率方法。设计一学习网络, 以低分辨率图像作为网络的输入, 从根本上减少网络的运算负担, 加速网络运算; 减小卷积核尺寸使得网络训练参数减少, 提高运算速度; 最后以亚像素卷积层同时实现网络的映射和图像融合过程。将所提方法在通用测试集上进行测试, 并与其他方法的测试结果进行了对比, 所提方法生成的图像具有更高的峰值信噪比, 且具有更好的主观视觉效果。实验结果表明所提方法不仅运算速度能够得到大幅提升, 而且能够生成更高质量的超分辨率图像, 具有更佳的超分辨率性能。

Existing image super-resolution methods that are based on sample-learning involve a large number of associated parameters and protracted calculation time. Hence, a method for rapid image super-resolution has been proposed that incorporates convolutional neural networks. The proposed method directly processes low-resolution images as network inputs so that the networking computational cost can be reduced and the associated calculation speed can thereby be accelerated. The size of the convolutional kernel was decreased for reducing the number of network training parameters and for improving the computation speed. Furthermore, nonlinear mapping and image fusion were simultaneously achieved through the sub-pixel convolution layer. The proposed method was applied to images of universal test datasets and then compared with other methods for evaluation. The evaluation results obtained from the proposed method indicate a higher peak signal-to-noise ratio and better subjective visual effects. The experimental results confirm that the proposed method can improve the calculation speed by a large margin and can also produce super-resolution images of superior quality.