[Feature request] Discard out of bounds text

[Johndirr]

Describe the feature

When text detection is done for a cropped image the performance can be worse than when using an uncropped image. Here is an example:


In the cropped image only a part of the text is detected. To prevent this from happening it would be nice to have the option to discard text that is not detected inside the bounding box.
This will also be useful to get rid of text from TV station logos as you can see in the images.

[timminator]

I cannot quite follow you...
You make the crop box yourself in the GUI... If you dont include the TV station logo into the crop box it won't be detected.
And if the text is not detected in the crop box there is also nothing to discard because there was nothing detected in the first place?

[Johndirr]

Unfortunately I can't find the proper scenes but it happened to me that a text was not detected when the input image was cropped. It was only detected when the whole image was analyzed. What I try to highlight here is that the text detection performance for a cropped image can be way worse than for the whole image.
I have not taken a look at your implementation but if you crop the input before trying to detect text this will be a problem.
A tv station logo can be a problem if subtitles are reaching into the area of the bounding box of the logo. When the image is cropped already there is no way to remove the tv logo text (it will be mixed into the detected text) but if you discard text that is only partly in your area of interst (let's pretend the crop box defines your region of interest) you will be able to discern the full subtitles without the logo text.

I can draw you some examples when I have the opportunity again.

[Johndirr]

Here are the examples.
In the first image a cropping will result in a missing line:

The second image shows how the crop box can also be used to discard text that is not part of a subtitle:

So while it saves processing power to compute smaller images it can be a disadvantage. By having the option to use the crop box to discard text that is not inside it you can archive better text outputs.

[timminator]

I see your point, but I dont think im interested in reversing the functionality of the crop box.
If you leave some space around the subtitles and don't make the crop box too tight this also rarely happens from my testing...

In the advanced settings you have the option already to "Use Full Frame OCR", so the whole image is processed as you like.
If you now have the problem that a logo text is always recognized you could just open the subtitle file in vs code for example, search for it and replace it with nothing or write a simple script to do it. I think that should achieve exactly what you want

[Johndirr]

When using "Use Full Frame OCR" will text that is outside the crop box be discarded? If not I have to remove randomly detected text afterwards (on TV screens, Shirts, Advertising etc.).
About the logo text, you are right its fixable by hand or programmatically afterwards (if the logo text is detected correctly) but its still a step more to do :)

I can understand that you don't want to reverse the functionality of the crop box. But you could just add an option to discard text outside of the crop box when "Use Full Frame OCR" is used. If this isn't already the case. Please let me know how you think about this.