views:

177

answers:

1

Hello,

I am new into Image Processing, and I have started simple project for recognizing text on images with complex background. I want to use stroke filter as first step, but I cannot find enough data about stroke filters to implement this.

I have only managed to find pure definition of Stroke Filters. So does anyone know something about Stroke filters that can be helpfull to me for my implementation.

I found one interesting article here, but it doesnt explain Stroke Filters in depth http://www.google.com/url?sa=t&source=web&ct=res&cd=4&ved=0CBoQFjAD&url=http%3A%2F%2Fwww-video.eecs.berkeley.edu%2FProceedings%2FICIP2006%2Fpdfs%2F0001473.pdf&ei=qtxaS5ubEtHB4gbFj8HqBA&usg=AFQjCNEnXQCMAFnqPRHe2kNZ6JEidR1sQg&sig2=wpaIDIQmNn739aF0cYWbsg

+2  A: 

"Stroke Filters" are not a standard idea in image-processing. Instead, this is a term created for the paper that you linked to, where they define what they mean by a stroke filter (sect. 2) and how you can implement one computationally (sect. 4).

Because of this, you are unlikely to find an implementation in a standard toolkit; though you might want to look around for someone who's posted an implementation, or contact the authors.

Basically, though, they are saying that you can identify text in an image based on typical properties of text, in particular, that text has many stroke-like structures with a specific spatial distribution.

tom10
Thank you very much for your answer. I have searhed a lot and there was only few resources about Stroke Filters, now I see why.It says that when we find pixel with strong response all of its neighbours should be set at same value. But as stated in this paper one "stroke" have 4 orientations, do I have to iterate 4 times through one pixel neighbourhoud(becayse of different orientation) or is there some simpler way to do that?If I have larger image, it would be a lot of time visiting same pixels.Can you suggest me maybe better way to acomplish my goal?
svlada
I have read 20-30 papers on this subject, and all of authors have just propesed methods ... a lot of words but it is question if that methods actually works. Most of them have even same pictures as final result.This article seemed well because I have text with very specific sizes to filter out.
svlada
I've never personally implemented this method so I can't easily answer your questions. They are good questions though, and not having them completely answered is why it's difficult to do work on the cutting edge. It seems like they have some good ideas, and figure 7 in the linked paper is convincing; but it's a question of what they are not telling you, like do they only show the 1% of images where the method actually worked? It's impossible to know the details without contacting the authors or finding someone who's spent some real time with this.
tom10
You are right about that. I am definitely going to contact authors to see, if they are willing to help in any way :) Thank you for your answer again.
svlada