0

System:

OS - Ubuntu 16.04 LTS // RAM - 8 GB // GPU - 2 x GTX 1080 // Software - Tensorflow / Keras

Problem Statement:

I have large images (6500 x 4308 px) on which I want to run a training process to train a DL-algorithm. The problem is this size of images would not fit in the Memory of my graphic Cards (2 x GTX 1080). I planned to divide the Images into smaller tiles to fit into the Memory and at the end, the tiles would be put together to show the result Image (in the test Phase to see the prediction on the Images, example Surface defects).

But for every large Image, I have an Annotation file which has Information about the bounding boxes in the Image. So now I am out of tricks or idea about how to proceed. I'd be really happy if you guys could suggest me method(s) out of your experience.

Thank you in advance!

edit: Image re-sizing is not an Option because the ROI's are as small as 5 x5 px. They would be lost if the Image is re-sized.

LK17
  • 1
  • 3

1 Answers1

0

You could crop small parts from the images containing regions of interest in pre-processing. A small neural network could then be trained on these cropped regions of interest. You could then either scan the small neural network across the full-sized images (inefficient) or use a reinforcement learning strategy (efficient) to identify the regions of interest in the full-size images.

Jeffrey Ede
  • 601
  • 5
  • 11