Model trained on patent data + the original training data (starting point: original DECIMER Segmentation model)
The seed pixel determination now works differently: All pixels that are black in the binarised, dilated image and that are covered by the "inner 80%" of the mask are considered seed pixels.
We are now using scipy to detect connected objects. The seed pixels are then just used to determine which connected objects to add to the refined, expanded mask. This speeds up the whole procedure