The purpose of this project is to read solar panels scan photos and process them for evaluation.

The photos were taken with an infrared camera. Due to NDA aggreement the photos are not included.

The first step is to detect centralized cells inside each panel frame. We filter away any quadrilaterals that are not almost perfectly square.

Afterwards, we create a mask based on the first quadrilateral, which works as a trust region between the current and next frame. Then, we find matching interest points. The homography can be applied directly to the old quadrilateral, and now we have a matching quadrilaretal in the new frame. We convert the quadrilaterals to rectangular images/matrices and stuff them in to a tensor.