4 ms·

But zooming scales the fonts only. For pixel images you have the pixels that are in it and that's it. Scaling those either up or down does not produce good text

But zooming scales the fonts only. For pixel images you have the pixels that are in it and that's it. Scaling those either up or down does not produce good text. Now I'm just waiting for the inevitable "AI image scaler" that handles text inside images.

TeMPOraL 2y ago
> Now I'm just waiting for the inevitable "AI image scaler" that handles text inside images. I'm surprised this isn't a thing already, as it seems doable with what people called "AI" 20 years ago. I mean, unless some unusual/non-default font was used, upscaling text on an image should be almost trivial. Ligatures notwithstanding, "printed letters" have a fixed shape, so: 1. Identify the typeface, size, weight, etc. by looking at the pixels of the text; 2. OCR the text (which should be 100% reliable); 3. Blank out the original text pixels; re-render the content (from step 2.) at a larger size (using parameters from step 1.). I'm hedging here; it feels to me that OCR-ing normal text that never left the digital realm should be 100% reliable, but I'm not a specialist in that subfield so I surely must be missing something...
- ben_w 2y ago
  It's definitely possible, as Google Translate (and before it, Word Lens) does exactly that. Word Lens was demonstrated in late 2010.
- staplung 2y ago
  > I'm hedging here; it feels to me that OCR-ing normal text that never left the digital realm should be 100% reliable, but I'm not a specialist in that subfield so I surely must be missing something... A string set in a given font at a given size won't always render as a fixed pattern of pixels. The font describes the curves of the letter forms and how that's rasterized depends on lots of factors such as the zoom level, exactly how the font rendering engine is implemented, whether or not anti-aliasing is turned on which is further complicated by the fact that the text can be set in any color with any other color as a background, etc. And there are a LOT of fonts. Lastly, OCR is not just about recognizing letter shapes but has to contend with how the text flows. It has to understand line-breaks, multi-column layouts, captions, pulled quotes, page-numbers, hyphenation and all the other weird shit that we make text do.

TeMPOraL 2y ago

> Now I'm just waiting for the inevitable "AI image scaler" that handles text inside images. I'm surprised this isn't a thing already, as it seems doable with what people called "AI" 20 years ago. I mean, unless some unusual/non-default font was used, upscaling text on an image should be almost trivial. Ligatures notwithstanding, "printed letters" have a fixed shape, so: 1. Identify the typeface, size, weight, etc. by looking at the pixels of the text; 2. OCR the text (which should be 100% reliable); 3. Blank out the original text pixels; re-render the content (from step 2.) at a larger size (using parameters from step 1.). I'm hedging here; it feels to me that OCR-ing normal text that never left the digital realm should be 100% reliable, but I'm not a specialist in that subfield so I surely must be missing something...

ben_w 2y ago
It's definitely possible, as Google Translate (and before it, Word Lens) does exactly that. Word Lens was demonstrated in late 2010.
staplung 2y ago
> I'm hedging here; it feels to me that OCR-ing normal text that never left the digital realm should be 100% reliable, but I'm not a specialist in that subfield so I surely must be missing something... A string set in a given font at a given size won't always render as a fixed pattern of pixels. The font describes the curves of the letter forms and how that's rasterized depends on lots of factors such as the zoom level, exactly how the font rendering engine is implemented, whether or not anti-aliasing is turned on which is further complicated by the fact that the text can be set in any color with any other color as a background, etc. And there are a LOT of fonts. Lastly, OCR is not just about recognizing letter shapes but has to contend with how the text flows. It has to understand line-breaks, multi-column layouts, captions, pulled quotes, page-numbers, hyphenation and all the other weird shit that we make text do.