I know an imbalance in the classification of images, such as the classification of cats and dogs, if there are too many images of cats and too few images of dogs. But I do not know how to address the imbalance in the problem of segmentation.
For example, my task is to mask the cloud cover from satellite images, so I convert the problem into two classes of segmentation, one is the cloud, the other is the background. The data set has 5800 4-way 16-bit images of 256 * 256 size. Segnet architecture, loss function - binary cross-entropy.
Two cases are suggested:
- Half of all samples are completely covered with clouds, half without cloud.
- In each image, half is cloudy, half is not.
So, case 2 is balanced, I think, but what about case 1?
In fact, my task, two cases are impossible in the original satellite image, since the cloud cover is always relatively small against the background, but if the image samples are cropped from the original images due to their large size, new cases appear.
So, samples always contain three types of images:
- completely covered with clouds (254 out of 5800 samples).
- without cloud (1241 out of 5800 samples).
- some areas covered by the cloud, in some areas not. (4305 at 5800, but I do not know the percentage of clouds, perhaps very high in some samples, maybe a little in other samples).
My question is:
Are the samples unbalanced and what should I do?
Thanks in advance.