Segmentation and inpainting of sensor failures in remote sensing images

Abstract

Although Remote Sensing (RS) images have been successfully used in several applications, the quality of the processed images may be dramatically reduced due to particular issues occurred in the data acquisition step, such as changes in signal capture or sensor failures, leading to the emergence of occlusions within the images. These occlusions make post-processing tasks less effective in practice, which includes image classification applications or the analysis of the original data which has been damaged. Therefore, this project focuses on restoring RS images containing failures from capturing sensors, specifically the Landsat-7 sensor. The proposed solution combines the prior segmentation of the occluded targets, as it was specially designed to deal with degenerations caused by the Landsat-7, and the inpainting of segmented failures, which will be performed by exploiting two digital inpainting approaches: the first one based on PDEs, and the another one relying on patch replication. As a result, it is expected that the detected occlusions, i.e., the regions that do not correspond to the original image scenario, will be fully inpainted by the computational solutions to be implemented in this project. (AU)