I'm trying to do some research about semantic segmentation for webpages, in particular e-commerce webpages. I found some articles which provide some solutions based on very old dataset and those solutions in my opinion can't be effective for modern websites, in particular e-commerce. I would like to semantically infer the images bounding box, text, price etc..
Another problem is related with the size of webpage screenshot which are huge, I resized to 1024x512, but I think that I can't resize the image more otherwise I loose quality.
I built a very complex neural network in order to semantically infer text, images and background, (not classification but just segmentation), and the results are not so bad, but they are far from my expectations which seems strange to me, as we have many DNN able to do semantic segmentation of road, building, car etc for example. One problem is for sure the lack of a dataset with detailed labels. I didn't find any dataset that can satisfy my requests.
QUESTION: Any idea to help the network learn better the structure of a webpage just with a screenshot?
My DNN essentially is built as an auto-encoder architecture based on Segnet, with some modifications, skip connections, unpooling etc, I think that it is a good network.