Image extraction from pdfs can reduce resolution and alter aspect ratio

[apbard]

Hi,
If my understanding is correct have noticed that when dealing with pdf the process is roughly this:

• each page is converted to image at 200dpi (the default of pdf2image)
• it gets then resized to max_img_size x max_img_size

Wouldn't this process possibly greatly reduce the quality even if max_img_size is set very big?
also the resize also alters the aspect ratio of the images. Is this done on purpose?

thanks

Here the relevant code bits:

docext/docext/core/utils.py

Lines 51 to 52 in 668eee6

	img = Image.open(file_path)
	img = img.resize((max_img_size, max_img_size))

docext/docext/core/pdf2md/pdf2md.py

Lines 86 to 87 in 668eee6

	file_paths = convert_files_to_images(file_paths)
	resize_images(file_paths, max_img_size)

docext/docext/core/file_converters/pdf_converter.py

Line 14 in 668eee6

return convert_from_path(file_path)

[github-actions]

👋 Welcome! Thanks for opening your first issue. If you'd like to take a crack at fixing it, feel free to open a pull request — otherwise, we'll take a look as soon as we can!

[mandalsouvik3333]

Good points.

This should be fixed. Initially, the max_img_size was there mainly to make docext work with small GPUs. But we should keep the aspect ratio also. We can maybe take two parameters, max_size and min_size. When both are the same, then resize to a square; otherwise, resize using either max_size or min_size while keeping the aspect ratio the same.

We should take an argument for the DPI also.

Let me know if you want to contribute :).

[apbard]

@mandalsouvik3333 I am maybe missing a bit. why are you suggesting using max and min size and maybe not just a single parameter max_size where we preserve the aspect ratio while trying to match the target dpi?

[mandalsouvik3333]

So the minimum size will help people who want to get the best results out of the model. minimum size will help folks who want to process files at a smaller resolution because of VRAM constraints.

A single max size will reduce the max(width, height) to that value. Which is great for the second requirement. But for the best performance, you should scale min(width, height) to max_size.