High resolution imagery is of crucial importance for the performance on visual recognition tasks. Super-resolution (SR) reconstruction algorithms aim to enhance the image resolution beyond the capability of the image sensor being used. Traditional SR algorithms approach this inverse problem using physical models for the image formation combined with a regularization function to prevent instabilities in the solution. Recently deep neural networks have been put forward as an alternative approach to the SR reconstruction problem. They learn a mapping from low resolution images to their high resolution counterparts from pairs of training images, which allows them to capture more specific information about the space of possible solutions than traditional regularization functions. These networks have achieved state-of-the-art performance on single image SR for sets of generic test images. Here we investigate whether the same performance can be realized when these neural networks for single image SR are applied specifically in the maritime domain. In particular we investigate their ability to reconstruct undersampled images of ships at sea, and demonstrate that the performance is similar to what is achieved on generic test images. In addition we quantify the gain in performance that is achieved when the networks are trained specifically on images of ships, which allows the networks to capture more prior knowledge about the space of possible solutions. Finally we show that the performance deteriorates when the resolution of test images is limited by image blur, for example due to diffraction, rather than undersampling. This highlights the importance of using representative training data that account for the part of the image formation process that limits the resolution in the sensor data.
|