[reconstitution] Improve synthesize output quality

[tzktz]

          @felixdittrich92 i have face result image is not upto quality...fonts are breaks in result image..

model = ocr_predictor(pretrained=True)
# PDF
doc = DocumentFile.from_pdf("bankstatement.pdf")
# Analyze
result = model(doc)
import matplotlib.pyplot as plt
plt.imshow(result.synthesize()[0]); plt.axis('off'); plt.show()

see the result image..

Originally posted by @tzktz in #1525 (comment)

[felixdittrich92]

Yeah we can maybe align the y-coords between line elements (words) and add some small horizontal default padding between detections
CC @odulcy-mindee

[tzktz]

Yeah we can maybe align the y-coords between line elements (words) and add some small horizontal default padding between detections CC @odulcy-mindee

how to change the font_family ? @felixdittrich92

[felixdittrich92]

result.synthesize(font_family="XYZ")

under the hood calls PIL:
font = ImageFont.truetype(font_family, font_size)

[tzktz]

result.synthesize(font_family="XYZ")

under the hood calls PIL:
font = ImageFont.truetype(font_family, font_size)

synthetic_pages = result.synthesize(font_family='Arial.ttf', font_size=13)
plt.imshow(synthetic_pages[0]); plt.axis('off'); plt.show()

same warning even i pass the font_family.. @felixdittrich92

WARNING:root:unable to load recommended font family. Loading default PIL font,font size issues may be expected.To prevent this, it is recommended to specify the value of 'font_family'.

[felixdittrich92]

result.synthesize(font_family="XYZ")

under the hood calls PIL:
font = ImageFont.truetype(font_family, font_size)

synthetic_pages = result.synthesize(font_family='Arial.ttf', font_size=13)
plt.imshow(synthetic_pages[0]); plt.axis('off'); plt.show()

same warning even i pass the font_family.. @felixdittrich92

WARNING:root:unable to load recommended font family. Loading default PIL font,font size issues may be expected.To prevent this, it is recommended to specify the value of 'font_family'.

The font is installed on your system ?

[tzktz]

result.synthesize(font_family="XYZ")

under the hood calls PIL:
font = ImageFont.truetype(font_family, font_size)

synthetic_pages = result.synthesize(font_family='Arial.ttf', font_size=13)
plt.imshow(synthetic_pages[0]); plt.axis('off'); plt.show()

same warning even i pass the font_family.. @felixdittrich92

WARNING:root:unable to load recommended font family. Loading default PIL font,font size issues may be expected.To prevent this, it is recommended to specify the value of 'font_family'.

The font is installed on your system ?

yes i have that font in my project folder.. @felixdittrich92

[felixdittrich92]

result.synthesize(font_family="XYZ")

under the hood calls PIL:
font = ImageFont.truetype(font_family, font_size)

synthetic_pages = result.synthesize(font_family='Arial.ttf', font_size=13)
plt.imshow(synthetic_pages[0]); plt.axis('off'); plt.show()

same warning even i pass the font_family.. @felixdittrich92

WARNING:root:unable to load recommended font family. Loading default PIL font,font size issues may be expected.To prevent this, it is recommended to specify the value of 'font_family'.

The font is installed on your system ?

yes i have that font in my project folder.. @felixdittrich92

Ah ok got it that's not enough you need to install the font on your system : https://linuxiac.com/how-to-install-fonts-on-linux/#:~:text=Go%20to%20%E2%80%9CSystem%20Settings%E2%80%9D%20%3E,%E2%80%9CInstall%20from%20File%E2%80%9D%20button.&text=Then%20select%20the%20font%20files,%2Dwide%20or%20per%2Duser.

[tzktz]

see the below input and output results.. result image quality is very poor.. pixels were broken @felixdittrich92
input image..(1240 x 1754) 158.44kb

result image..(1907 x 965) 46kb

[tzktz]

@felixdittrich92 any update?

[felixdittrich92]

Hi @tzktz 👋,

Unfortunately i don't have the time to work on that at the moment, so we need to address this later on or you work on that if you want (feel free to open a PR)

related code can be found at:

doctr/doctr/utils/visualization.py

Line 291 in 0d84915

def synthesize_page(

Best regards,
Felix

[SkaarFacee]

@felixdittrich92 Hey, sorry for being MIA. I needed to take some time off. I am back now and I was hoping I could take up this issue? Let me know regarding this :)

[felixdittrich92]

@felixdittrich92 Hey, sorry for being MIA. I needed to take some time off. I am back now and I was hoping I could take up this issue? Let me know regarding this :)

Hey @SkaarFacee 👋
Sure feel free to work on it 😊
The code moved a bit it is now in:
https://github.com/mindee/doctr/blob/main/doctr/utils/reconstitution.py

[SkaarFacee]

Okay thanks. Let me take a look on what I can do

[SkaarFacee]

@felixdittrich92 Do you have any suggestions on how I can improve the quality of the image ?

[felixT2K]

@SkaarFacee
One thing we could do is if we have the line box information we could align all boxes inside to the line y coordinate (to become a more straight view)
I found the following hf space the reconstitution looks not bad maybe you can use it as reference or to get some inspiration ^^ :
https://huggingface.co/spaces/SWHL/RapidOCRDemo/blob/main/utils.py

[SkaarFacee]

Okay, I will take a look and see what can be done over the weekend. This doesn't look that complex at quick glance :)

[felixdittrich92]

Okay, I will take a look and see what can be done over the weekend. This doesn't look that complex at quick glance :)

Yeah i think too :)

[SkaarFacee]

Hey, I am working on on this, sorry for the delay. Something came up at work and got me busy

[SkaarFacee]

@felixT2K
I was using the link you mentioned as reference (https://huggingface.co/spaces/SWHL/RapidOCRDemo/blob/main/utils.py)
I can't exactly pin point the place where the y coordinate was used to to align the line.
If the goal is to straighten the line, why don't we make the y coordinates of each box in the line the same using a mathematical approach ( using the mean or centroid of the boxes). If you could maybe give me more insights on the hf reference I would gladly implement that as too 😄

[felixdittrich92]

@felixT2K I was using the link you mentioned as reference (https://huggingface.co/spaces/SWHL/RapidOCRDemo/blob/main/utils.py) I can't exactly pin point the place where the y coordinate was used to to align the line. If the goal is to straighten the line, why don't we make the y coordinates of each box in the line the same using a mathematical approach ( using the mean or centroid of the boxes). If you could maybe give me more insights on the hf reference I would gladly implement that as too 😄

Correct that was what i have had in mind you know which boxes are in one line (of `resolve_lines=True otherwise if only one line element available keep the y coords of each box) then take the lines y coordinate for each box to straighten the boxes on the line :)

[felixdittrich92]

@SkaarFacee any updates ? 🤗

[SkaarFacee]

Hey, I am soo sorry. I had some personal issues that needed some time to be resolved. Everything is fine now and back to normal. I shall give you an update asap now 😓 😞

[felixdittrich92]

Hey, I am soo sorry. I had some personal issues that needed some time to be resolved. Everything is fine now and back to normal. I shall give you an update asap now 😓 😞

No problem :)

In the meanwhile i identified the root issue.

• Code itself works as expected
• Align to line geometry can be added easily
• issue: font size computation isn't correct (it should be computed depending on the geometry size in respect to the page size) - that's the reason why it looks sometimes gibberish

[SkaarFacee]

Regarding the font size computation, should we maybe add a limit to the font to page size ratio so that the issue becomes less frequent ? Also regarding the line geometry alignment, I think maybe we can also add an tolerance value of some sort so that words that have y pixel values as 40 and say 43 come together in a straight line. ( Here I assume the tolerance is of +-5)

[felixdittrich92]

Hey @SkaarFacee 👋,

• point for the line alignment sounds good to me 👍
• about the font size i think we need some logic to get the ratio of a geometry to the page size and then find some calculation to compute the font size or we do a mapping depending on the ratio to some fixed font size values which fits well wdyt ?

[SkaarFacee]

about the font size i think we need some logic to get the ratio of a geometry to the page size and then find some calculation to compute the font size or we do a mapping depending on the ratio to some fixed font size values which fits well wdyt ?

Yes, do we have some sample images so to get started with in this case ?

[felixdittrich92]

about the font size i think we need some logic to get the ratio of a geometry to the page size and then find some calculation to compute the font size or we do a mapping depending on the ratio to some fixed font size values which fits well wdyt ?

Yes, do we have some sample images so to get started with in this case ?

Yes give me few minutes i can attach some

[felixdittrich92]

Here you go :)
samples.zip

[SkaarFacee]

Are these samples that need to be improved or a combination of good outputs and bad outputs ?

[felixdittrich92]

Are these samples that need to be improved or a combination of good outputs and bad outputs ?

Only some samples for testing :)