I have been training my model several times with Unet. Unfortunately, it seemed like the model can only make good prediction on images with certain histogram even after i added both type of images.
I made sure I have 50:50 dataset of both histogram. I applied also dropout layer to the architecture.
Hi, yes I did with cv2.convertScaleAbs(), and I also played on brightness and contrast. But it seemed that the algorithm can not work well on dark and light images. So now we have one model that can work well on dark images and another model that can work well on bright images.