I'm currently trying to automate the testing of the visual appearance of a web application. I'm using a combination of openCV and tesseract OCR to find text intersection with Element borders with quite some success but the second Text overlaps with text tesseract fails to recognise characters and therefor fails to properly outline text boxes. I'm using seleniums WebElement.getRectangle() function to try and look for Elements intersecting into each other but not only is it slow but also shows a lot of false positives because of hidden text nodes.
Since the problem boils down to a binary Image classification (overlapping text or not overlapping text) I've created a simple classification model in Tensorflow which doesn't produce good results though since even after I mined all the data our bug tracker / google gave I'm still far away from having a solid set to train it on.
Now to the heart of my question: Since I do not believe that I'm the first person to try to get rid of the task to look through hundreds upon hundreds of screenshots only to make sure everything looks right does anyone know of a huge enough data set of common visual errors to efficiently train an convoluted neural network with it?