LogoScale – A method for vectorizing small, crappy logos
msprout.notion.site
msprout.notion.site
It seems like it doesn’t actually create a vector image by itself, which is what the title had me hoping. I’d be interested to see a process that links up these two parts of the problem.
I’ve been on the SVG using train for what feels like forever now, so long that I’ve tried to do the raster to vector conversion dance countless times and watched the recent progress with much satisfaction, but it’s still not perfect, two colour gradients (a common enough logo element) are frequently a problem last time I checked, which admittedly could also just be really troublesome set of examples since it’s not like I’ve got a library of images I use to regression test this stuff, I just use it.
Of particular note is when these kinds of issues crop up with overlapping elements, like having flowing ribbons or abstract line work, or other art with a colour gradient on it, which when converted to grayscale, produces the same tone or similar enough tone that the de-noising of the vectorisation decides to merge the regions as a single object. There are ways around it… but doing multiple conversions to vectorised images after performing colour space rotations, desaturations, grayscalings, and then having to stitch together a final vector in out of the correctly vectorised portions of multiple conversions… is not fun… and makes me deeply frustrated at the lack of available vectorisation software that makes full use of the colour information available.
It’s been a while since I dug down into the academic papers on the topic but when I last looked years ago, I didn’t find anything about vectorisation algorithms that didn’t basically rely on the same techniques as OCR algorithms, greyscale and contrast enhancements, before moving onto the parts of the algorithms where they began to make vectors from the raster data matrixes they had extracted.
Compared to the current state of the art in object detection in computer vision, which is admittedly something I have much more recently dug into the algorithms behind, it feels like a lot of vectorisation software is practically in the Stone Age.
I'm sure this is a solvable problem. For now though, a professional illustrator is still best for creating "parsimonious" logo designs
Parsimonious as in not an absolute nightmare to edit afterwards with nonsensical artifacts
I wonder how this compares with just directly dropping the logos into adobe converter or inkscape?
Also - does anyone know of a tool that does in fact vectorize logos with AI? As in, creates sort-of semantic svg elements instead of just paths for everything - given a circle-ish thing you get a <circle> out.
It’s not that hard to do automatically
Citation required?
The only satellite related thing I saw in my brief searching is some separate project[1] based on ESRGAN used for satellite imagery.
Some people release OSS out of passion. Some release it out of pain.
Tracing logos back will still give subpar results IMO – I think you're better off just using the resulting hi-res logo bitmap.
This way you can scale as you want and create bitmaps as needed at any size.
Years from now the client will be gone and someone else from the company will explain, “I didn’t have any special design programs installed on my computer so I couldn’t open that eps thing and I didn’t know it was important.”
Over time, I have grown to appreciate idiocy as a positive force of evolution. I now think it should be exploited whenever possible --- for the greater long term benefit of humanity.