Image stitching is a task in computer vision. The goal in this task is to create a mosaic image from some smaller images. The task is similar to the panorama image task. An example to an application of this task is aerial photo stitching. In this case, an airplane takes photos of different regions and the goal is to build one big aerial photo that contains the information from all the small photos. There are algorithms that solve this task, but their performance is not perfect and sometimes an irregularity can be found in the stitching area of two photos. As a result of the irregularity, some data can be lost.
In recent years, Deep Learning based systems solve many tasks in computer vision with impressive performances. Specifically, Generative Adversarial Networks (GANs) solve tasks of data generation in natural images (for example, filling holes in a natural images) with very high performances.
In this project, we build a system that gets as input a mosaic aerial photo that was stitched with a certain algorithm and try to fix the irregularities using GANs, when we focus on irregularities that cause loss of data. We train a network with an aerial photos database, then test the performance for different techniques. When it comes to completing details (for example, inpainting road with “hole”), results were good. However, the system had difficulties with color reconstruction.