Multi Image Super-resolution (MISR) refers to the task of enhancing the spatial resolution of a stack of low-resolution (LR) images representing the same scene. Although many deep learning-based single image super-resolution (SISR) technologies have recently been developed, deep learning has not been widely exploited for MISR, even though it can achieve higher reconstruction accuracy because more information can be extracted from the stack of LR images. One of the primary obstacles encountered by deep networks when addressing the MISR problem is the variability in the number of LR images that act as input to the network. This impedes the feasibility of adopting an end-to-end learning approach, because the varying number of input images makes it difficult to construct a training dataset for the network. Another challenge arises from the requirement to align the LR input images to generate high-resolution (HR) image of high quality, which requires complex and sophisticated methods.

In this paper, we propose a self-learning based method that can simultaneously perform super-resolution and sub-pixel registration of multiple LR images. The proposed method trains a neural network with only the LR images as input and without any true target HR images; i.e., the proposed method requires no extra training dataset. Therefore, it is easy to use the proposed method to deal with different numbers of input images. To our knowledge this is the first time that a neural network is trained using only LR images to perform a joint MISR and sub-pixel registration. Experimental results confirmed that the HR images generated by the proposed method achieved better results in both quantitative and qualitative evaluations than those generated by other deep learning-based methods.