I’m currently working on a binary image classification problem using high resolution (up to 6000x4000 pixels) images with complex backgrounds, and CNN transfer learning.
In order to reduce Images size and complexity, I applied an offline crop to all images in a way that keeps the Region Of Interest (ROI) distinguishable. This resulted with three types of cropped images:
- images with background only,
- images with ROI only, and
- images with both ROI and background.
Only images of type 1 and 2 were used to train validate and test the CNN model.
I want to use the background-only images (type 1) as a background class in order to make my model fucus only on ROI and ignore backgrounds. How can I deal with the background class (number of images, loss function…) ?