I have tried using RCNN and faster RCNN to detect tables and paragraphs. However, I keep encountering the same issue, which I guess is because of region suggestion algorithm; Parts of the first or last lines get left out of the bounding box. So, the words in those sentences are illegible, and in some cases, the extraction of that paragraph or table is pointless (I need to extract very specific info with an OCR).
Which architecture should I go for to maximize IoU? Should I look for ML techniques from outside of computer vision?