Efficient Deep Multi-Image Super-Resolution for low-power devices: performance optimization through fusion and registration techniques

Image super-resolution is a well-known problem in computer vision that aims to enhance a low-resolution image into a high-resolution one. In recent years, deep neural networks have achieved remarkable results in tackling image super-resolution tasks, often employing large models and subsequently huge amounts of computational resources. However, the wide diffusion of mobile devices and portable photography necessitates the design of computationally efficient solutions. This thesis aims to address these challenges by adapting an existing lightweight neural network designed for single-image super-resolution to perform burst super-resolution while preserving its mobile-friendly characteristics. The main goal is to improve the quality of super-resolved images, assessed using the Peak Signal-to-Noise Ratio (PSNR) metric, by inputting more than one image into the network, while trying to keep the model’s inference time as low as possible. To achieve this goal, several elements like fusion techniques and registration strategies will be studied and compared. Moreover, a way to synthetically generate a burst image dataset is also exploited, in order to evaluate PSNR performance and compare the results with those of the original neural network. This work includes an introduction to the problem statement, a theoretical background on the used tools, a literature review of methods for both single and multiimage super-resolution, as well as a description of the developed methodology and the results achieved.