图像超分辨率(Super Resolution，SR)技术能够从低分辨率图像中恢复出高分辨率图像，已被广泛应用于遥感、医学影像、目标跟踪与识别等多个领域。随着深度学习研究的深入，该技术也被成功应用于 SR 相关研究中，但现有工作往往只关注输出图像的质量，而忽略了训练和重构效率。该文基于对图像特征和训练效率的观察，提出了一种基于多模型的 SR 框架——MMSR，能够根据不同的图像特征选择合适的网络模型，从而在不影响输出图像质量的情况下有效缩短训练时间。面向 DIV_2K 图像集的测试结果表明，该框架能够实现平均 66.7% 的性能提升，同时具有良好的可扩展性。
Super resolution (SR) technique is an important means for image resolution improvement, which has been widely used in remote sensing, medicine image processing, target recognition and tracking etc. In recent years, the deep learning techniques also have been applied in the SR domain successfully. However, researchers pay most of their attentions on the quality of the output images, but ignore the training or reconstruction efficiency. In this paper, we found that for images with different texture features, the most appropriate models are usually different. Based on this observation, a multi-model super resolution framework (MMSR) is proposed, which can choose a suitable network model for each image for training. Experimental results with the DIV_2K image set indicate that, efficiency can be improved 66.7% without the loss of image quality. Moreover, MMSR exhibits good scalability.
WU Xinzhou, YUAN Ninghui, SHEN Li. An Efficient Multi-Model Super Resolution Framework[J]. Journal of Integration Technology,2019,8(5):49-57