how can we detect a table which doesn't have strict tables

[Kotresh17]

Hi how can we recognize the table in the image which has not a strict table but table like pattern. for example please find the attachment in that image we have a table like structure we need to draw a bounding box around it. can you please help me with that how can we approach this.

[ulikoehler]

Sorry for the late reply - was quite busy in the previous months. In my opinion, the algorithm used by OTR is not fit to do this task. But it's absolutely possible to do it the same way I developed OTR. Here's my guess on what you'd want to do:

1. Detect lines that delimit tables (if there are lines such as in your example). This will allow you to estimate the area where the table is located. Then you might need to rotate your picture so it's absolutely straight (depends on your scan quality)
2. Detect "lines of white space", i.e. row separators, e.g. the entire horizontal line between TP-37 and TP-38. This will mark whereyour have to separate your columns
3. Detect "columns of whitespace". Same thing
4. Possibly skeletonize the lines and columns to get just vectors where your can draw your boxes. You might also want to ensure that all lines are 90° to each other.
5. After that, compute the corner coordinates of all the cells (do some math: intersections between the lines to compute the corners). That should be reasonably easy to do.
6. When you have the corner coordinates, you can reuse some of the algorithms in OTR. You do not need the "missing cell detection" cycle, see docs/Algorithm.pdf for an overview. This will give you rectangles for all the cells.

Note that your usecase is much much easier to implement than the original usecase of old naval logbooks I developed OTR for. Their scans look similar to this image

Thanks BTW for including an example image. This helped alot and I would hope more people would do that!

Please feel free to ask if you have any more questions.

[jewelcai]

User deep learning is a better idea. I just tested you image, use resnet-backboned faster rcnn can detect the table area.

But actually, the difficulty lies in the recognition part, and how to reconstruct the table for understanding.