Removing caption background for use in ORC (tessearct)

[gilberto_san]

I have some movies with embedded subtitles.
I'm doing a paper that tries to extract the text from these captions.

Examples:

Font colors may change

I would like a result like that.


I'm having a lot of trouble when the font color is gray
(as in the last 2 pictures)


Code used.

Code: Select all

 convert Screenshot_2.bmp  -channel rgba -alpha set -fuzz 10%% -fill none  -opaque "#786965" -opaque "#B9B1B9" -opaque "#D6D5D7" -opaque "#E8F4F7" -opaque "#19191C" -opaque "#39383A" -opaque "#8A8A8B" -opaque "#524C4E" -opaque "#3C4B4C" -opaque "#425353" -opaque "#425353" gil.bmp 

I'm new to IMAGEMAGICK
Sorry for bad English .
Thank you.

[gilberto_san]

Note: It is possible to obtain 2 or more images with the same text but with different background

[snibgo]

Note: It is possible to obtain 2 or more images with the same text but with different background.

That simplifies the problem. One possible method:

1. Take two images with the same text but with different background.

2. Find the best alignment between those images. Now the text appears in the same position on the two images.

3. Where corresponding pixels are different, this is background. Where they are the same, this is probably subtitle.

I show many methods for aligning images on my pages, linked below.

[gilberto_san]

Thank you for answering. Congratulations on the site. He helped me a lot.

I have a question.
1. ok.
2. ok.
3. Sorry, I do not know how to do this. Could you help me with an imagemagick code?

[snibgo]

For example, this finds the difference between two images. Where the difference is not zero (black), it changes the pixels to white.

Code: Select all

convert in1.png in2.png -compose Difference -composite -fill White +opaque Black out.png

[gilberto_san]

For example, this finds the difference between two images. Where the difference is not zero (black), it changes the pixels to white.

Code: Select all

convert in1.png in2.png -compose Difference -composite -fill White +opaque Black out.png

Look the images



i use

Code: Select all

convert in1.png in2.png -compose Difference -composite -fill White +opaque Black out.png

Result


Note that only the letter W was used.
They are aligned but the result was bad.

[snibgo]

The pixels in the text in the two images are not identical. Instead of "-fill White +opaque Black" you can use "-threshold 1%" or similar.

When the background pixels are also within 1% of each other, you may need to combine multiple images in this way.

When the height of the capital letters is less than 20 pixels, OCR with Tesseract may be unreliable.

[gilberto_san]

Thank you so much for your cooperation . I did the test with 4 identical texts and with different backgrounds.
Has improved a bit but is not yet readable

I was thinking of another solution.
It would be possible to normalize the colors of an image.
example:
An image has pixels: dark green, medium green, light green and green.
Would it be possible to turn them green?

Likewise transform: light gray, medium gray, dark gray to gray?

If so, please tell me how?

Look the figure


Remembering that green was just an example, the fonts can be any color.
And the Fund is always in variations of gray and black because the films are old.

[snibgo]

If the text is always coloured and the background is always gray, pixel saturation will separate them, eg:

Code: Select all

convert subtGrn.png -colorspace HSL -channel G -separate +channel -auto-level out.png

After "-auto-level", you could "-level 40%,60%" or "-sigmoidal-contrast 10,50%" to emphasise the difference.

[gilberto_san]

I did some tests, these comands solve 60% of the problems.

To solve it all I need something that does the following:
1) transform all shades of any color (not just green) in the actual color
Example: light red, dark red are transformed into red
2) all shades of gray are transformed into white (or other desired color)
Is there any command that does such a transformation?

See the example I did manually


Basically I would like similar colors to become a single color

Thank you one more time

[snibgo]

1) transform all shades of any color (not just green) in the actual color
Example: light red, dark red are transformed into red

This is vague. What are the possible "actual colors"? Just red green and blue? How about cyan yellow and magenta? Orange, brown, pink?

[gilberto_san]

Note the colors of the letter S. I have a variety of colors.
The fact is that the font seems to be the same and the background is in variations of gray and black

[fmw42]

Threshold on saturation might separate the colors from the grayish background.

[fmw42]

Try this Unix syntax:

Code: Select all

convert 9a6c443dfa76445fa95158948cc4e9bd.png \
\( -clone 0 -colorspace HCL -channel g -separate +channel -threshold 40% \) \
-alpha off -compose copy_opacity -composite result.png

Please always provide your IM version and platform, since syntax may differ.

[gilberto_san]

Could you help me by putting the code?

I got good medium results using the code.

Code: Select all

convert out.bmp +dither   -posterize 2  out1.bmp

However, in some cases, the following problem occurs:



I thank you for your attention.