How can I feed in a visual attention map to a network?

Question

I have a method that shows where to focus roughly on a given image, that is, it is not localization or segmentation labels but just a heat map that shows where what I look for in the image likely are. I am only hoping to increase classification accuracy as I don't have any other source of labels. I don't want to preprocess and crop the image according to the map and input in the network but I want to give higher prior to regions that my attention map shows. I have come across a lot of papers on self-guiding attention, soft attention learning, or using earlier activations as attention map, but I can not find more details on how I can implement a simple approach like I have mentioned.