Restoration of Optically Butted Satellite Images With Spatially Variant Blur Using Multi-Scale Vision Transformer

Optical butting scales existing sensor technology to expand the field of view (FOV) in remote imaging systems, meeting demands for higher resolution and coverage. A butting mirror splits the light beam into two, reflecting it to be captured by two sensors positioned in different planes. The FOV of two adjacent sensors contains overlap with each other, imaging the same area twice. The butting region includes the overlap and adjacent areas affected by the splitting of the light beam. Due to the physics of the imaging process, the characteristics of the blur and noise of the butting region is different. In this paper, we introduce a mathematical model to characterize the blur in the butting region. Our experiments demonstrate that the blur kernel for each unique location in the butting region is a truncated variant of the point spread function of the device. We present a two-step restoration framework using a multi-scale vision transformer, that requires no ground truth for the training process. We compare our proposed framework with deep learning based deblurring methods to prove its restoration capabilities.