Detect Contours

edit

The problem is I don't need the complete text. If you see the image, I only need specific text blocks. I have used EAST in other approach where I was clustering text first. Here, I want contours first and then use EAST on them.

I only need specific text blocks --> knowing the process inside those algorithms, to me it seems better to detect all text first, then remove blocks you do not need afterwards. That goes faster than cutting up the process beforehand. At least thats my experience for the stroke width transform.

Thanks for the help. But I would still like to try this approach. I had a different pipeline with EAST before where I was clustering the text, but I want to try this approach as well. I need to get that text block (contour in this approach), and then want to do the elimination. Can you help me with that?

Can you somehow define a mask of regions you are interested in?

That's the problem. I'm not sure of the region because the images are not standard. I don't know where the area (area of interest) will be for any random image.

Okay but there is a contradiction, if you do not know beforehand where to look, than you cannot limit the text retrieval in any way ...

I can still remove noise based on the contour area or the dimensions of contour. But first I want to make sure I am detecting majority text block.

Then why not go for SWT, than in the returning binary map, apply opening, so you get white blocks. Find contours and retrieve the largest one. That will be your region of interest right?

I will give it a try and get back to you. Thanks for the suggestion.

The output is very similar to east, and selecting contours is not easy in this case. I want to select contours first then look for text in them, that also helps in results where I don't have to look for bounding boxes line wise and can print the results block wise. Most of the nutrition facts are getting detected by my current approach, it's the free text that's hard to comprehend.