Super-resolution based on deep learning technique for constructing digital elevation model

Abstract

In this paper, the additional learning method on the pre-trained convolutional neural network (CNN) for image super-resolution (SR) and its usage for lunar image postprocessing is proposed. Transfer learning is a popular method in convolutional network (ConvNet) research because training a ConvNet to learn basic features for classification and detection is prohibitively time consuming. Transfer learning enables the re-training of the latter layer of a ConvNet to perform a different task. In SR, the overall ConvNet structure is much different from a ConvNet structure used for classification and detection, as the size of the input and output data must be identical. Inspired by the transfer learning method, an additional CNN structure is added to the base CNN for SR, and the additional ConvNet structure is newly trained. Results show a small improvement in performance over the base ConvNet structure in some example images. The CNN for SR algorithm outperforms the Bicubic interpolation method in restoring a sample lunar image to its original resolution. Possible applications include enhancing the resolution of lunar images to perform shape from shading, de-noising, and template matching the lunar surface image to a given DEM.