How to implement an intelligent cropper (remove frames from an image) using deep learning (keras, tensorflow, theano)?

Question

Python

Hello. There is a need for an intelligent image cropper (if the image has a frame, it must be removed and the original image without a frame returned, as if cropped).
I suppose that you need to submit a variety of images to the input (with and without frames), and to the output (match) exactly the same images, but cropped (and possibly stretched to the original size) in case the original image had a frame .
For training, I want to use Keras / Tensorflow and multilayer deep networks (convolution / pooling), but I still don’t fully understand (due to my little experience with neural networks and Keras) how to organize the network structure and whether I chose the right philosophy regarding the dataset.
Perhaps it is worth making a binary classifier (determine whether there is / is not a frame on the image), and only then somehow crop (possibly with another tool, for example, using openCV or other image manipulation tools)
If anyone has dealt with a similar issue, please advise , in which direction to dig. Many thanks in advance for the time spent on reading and possible help.

Answer 1

I would try to look in the direction of segmentation (frame or image), namely U-net .

Answer 2

opencv remove border