image preprocessing

[IceBear-huang]

Hello, why is it that some images are not accurately recognized, but when I re-screenshot them and make the screenshot larger, they can be recognized correctly? What's the reason for this? Could it be that image preprocessing is needed to improve accuracy? Do you have any good suggestions for image preprocessing?

[mandalsouvik3333]

VLMs generally work well with high resolutions. Docext, by default resizes the images to 2048. But high resolution will increase processing time. You can use this to resize directlyhttps://github.com/NanoNets/docext/blob/668eee618eb48bc81d97af65d770ea8c8e61e781/docext/app/args.py#L72

[IceBear-huang]

I deployed the model using vllm, but the images used have low resolution, with a size of only 2kb. In fact, the size of the images I directly screenshot is 29kb. Therefore, I suspect that these 2kb images have undergone compression. Visually, these 2kb images don't look any different from normal ones, but the model just can't recognize them accurately. However, the ones I manually screenshot and save can be recognized correctly.