muthuspark
commented
4 years ago Hello @SB2020-eye, My Apologies for the delayed response, I just saw the issue you opened. I am not sure if you are still looking for a solution. I am adding this comment to help anyone else with a similar query.
- Your sample image was too big which is the reason why thresholding did not work well. I am using an Adaptive Thresholding with a window size of 3. Small window sizes will only work for small images which was my case. I increased the window size to about 21 for your large image and I was able to get a proper separation of background and foreground.
A good window size is an important factor that can make or break any text recognition algorithm. There are many ways to choose this window size but it depends a lot on the size of the characters in the image. Building a generic window size predictor is quite difficult. It's usually selected based on heuristics.
- The algorithm runs very slow on my machine too, its mainly because of the size of the image it is processing. Faster machines may speed up the processing speed a bit. Also there is one step where I am discarding connected components having more than 2 children because these are likely to be non characters. I haven't been able to find a faster algorithm to compute this. Usually I skip this part when I need speed.
Also , I am discarding regions smaller than 15, this number will be much larger in your case. Basically trying to get rid of regions which are too small compared to the full image. I go by the general rule of 15*(image_width/1000). So in you case I changed it to 101.
So after discarding the likely non-text regions, this is what I get.
- Finally after combining all the other overlapping boxes, I get the below output for the proposed algorithm
- Extracting only the found regions, I get the below image.
You can find the running sample in this notebook here: https://github.com/9sphere/text-detector/blob/master/notebooks/User_sample_Segmentation_in_Image.ipynb
I see that you need individual characters, you may have to run some character segmentation algorithm further on the identified boxes. I will add this as an enhancement to my algorithm. Let me know if you need any more help, I would be happy to guide you.
Hi, muthuspark.
I'm new to all of this (including Python). I'm wondering if you can help me with getting this to work correctly.
I first came upon https://muthu.co/extracting-text-regions-from-an-image-using-geometric-properties/. I tried some of the code, but I couldn't get it to give me any output - but I'm sure the problem is me.
Then I came here to github. I couldn't get it to work; but again, probably it's me. I don't really understand how folders relate to each other in general (yet I realize this MUST be extremely basic). (I only have success when someone on github has very, very elementary steps guiding me through every piece. :))
But I figured I was close. So I went to https://nbviewer.jupyter.org/github/muthuspark/text-detector/blob/master/notebooks/Text%20Segmentation%20in%20Image.ipynb. With this info, I worked through each step [1], [2], etc., adding
plt.show()along the way to see if I was getting results. I managed to get some output (Yay!) - up through step [6]. (But it took my computer a very long time - especially step [6].)But I am not entirely confident about my output. The images look a little different than your examples. I've attached my original (greatly reduced in size here) plus my results from steps [2], [3], [4], and [6].
Ultimately, my goal is really to get a box around every individual character - and then saved each one in as an image file. (Ie, I don't need boxes around whole words or paragraphs. But the saved letter images need to be lossless - totally original color and quality.)
If you are able and willing to offer me any guidance, it would be greatly appreciated!