Image processing algorithms for synthetic image resolution improvement

The aim of this thesis is to develop a Machine Learning algorithm for Multi-image Super-resolution (MISR). Super-resolution is a well known image processing problem, whose aim is to process low resolution (LR) images in order to obtain their high resolution (HR) version. The super-resolution process tries to infer the possible values of missing pixels in order to generate high frequency information as coherently as possible with the original images. In general, we can distinguish between two different super-resolution approaches: Single-image Super-resolution (SISR) and Multi-image Super Resolution (MISR). The former tries to build the best LR-HR mapping analysing the features of a single LR image, while the latter takes as input multiple LR images exploiting the information derived from the small differences between the images such as changes in the point of view position and orientation, in the lightening condition and in the image exposition and contrast. The thesis had as first objective to take part to the competition called “PROBA-V Super Resolution”, organized by the Advanced Concept Team of the European Space Agency. The goal of the challenge was to obtain High-resolution images from low resolution ones from a dataset of pictures took from the PROBA-V satellite of the ESA. The work has been developed under the direction of the PIC4SeR (PoliTO Interdepartmental Centre for Service Robotics), which aims to integrate it into its agricultural related projects, for the satellite monitoring of the fields status. The analysis of the state of the art for what concerns super-resolution reveals that Machine Learning approaches outperform the classical algorithms proposed for this problem. In particular, Neural Networks have been widely used in literature for Single-image Super-resolution, while this approach for Multi-image Super-resolution is relatively new. An original model to deal with competition problem has been studied, trained and tested. The obtained results show that the multi-image approach can help in the improvement of existing algorithms for super-resolution. However, several issues can be further addressed to increase the model efficiency and performance, making this particular topic interesting for future work development.