I have only a limited dataset (<25) with large-sized images (>1500x2000) and their pixelwise labels. The aim is to find unusual patterns in this industry dataset and highlight them.
To generate training images I crop 256x256 grids out of every image and do some data augmentation and use these images to train my U-Net.
For my prediction I split my image with numpy again into 256x256px grids and predict every grid separately and put them together to an image. But this will take some time, like >10 Minutes. But it has a quite good accuracy. How can I optimize my prediction to be faster?
Is it faster to create this with a Tensorflow Pipeline? When I want to predict the full image with giving shape(None,None,3), I get "ConcatOp : Dimensions of inputs should match" after some time.