hi i have a question . i’m segmenting satellite images and for that i have to split images to smaller ones then apply the model . the problem is that when i’m merging the images together to make the whole image , the edge of my smaller images are visible . is there anyway to solve this problem
Hi Alireza!
Yes. Look at the “overlap-tile strategy” described in the original U-Net paper.
The basic idea is that you split your large image to be segmented
into overlapping smaller images, each of which is larger than its
corresponding predicted (segmented) sub-image and contains enough
context around its borders that there are no artifacts at the edges of
the predicted sub-images.
When the predicted sub-images are then combined together to produce
the final large predicted image, there are no edges where the sub-images
were glued back together.
How well this scheme will work for you will depend on the details of the
segmentation model you are using. If you use the original U-Net model
(properly), there will be no edge artifacts whatsoever interior to your
final large predicted image.
Best.
K. Frank
thanks for your reply . actually i think u are right . i only have this problem for two classes which my model cant separate them very well .