Extract numerical data from plot images | Hacker News Reader

Extract numerical data from plot images

github.com

This is very nice.

I particularly like the automatic mode. However, the deafult 'distance' threshold of 120 seems to be too large: selecting the yellow foreground color picks points on both the yellow and red lines; selecting the green foreground color picks points on both the green curve and the axes.

A solution for the test image is to just reduce the threshold to 50.

The best threshold may depend on both the target color and the specific image being processed, but could potentially be automatically chosen: imagine constructing a greyscale image by replacing each pixel's color with its distance from the target color, and then picking a threshold for converting this grayscale image to a binary image (e.g. using Otsu's method [0]).

The export to plot.ly feature is also great - you can really quickly extract data from one graph and re-plot it in a different/clearer way.

[0]: https://en.wikipedia.org/wiki/Otsu%27s_method

Hey, I am the developer of this app. Thanks for the feedback on the color distance. Also, I wasn't aware of the Otsu method, but have thought of doing something very similar - I will give it a shot for sure. I have added an enhancement item in the issue tracker.

On further thought this probably won't work very well: the histogram of colour distances will be multimodal, with one peak for each color in the image (and with the peak corresponding to the white background being far taller than the others). The Otsu method assumes there is just a foreground and background that need to be separated.

A clustering method like k-means might work, but I've had a quick play in Matlab and the results weren't great.

Thanks for trying it out. I am also working on a grid removal algorithm which will run into similar issues and so it's good to get some ideas. The naive color distance approach might not be the best for that scenario.

That is fantastic. I actually wrote a one-off utility to do exactly this from polar plots several years back. This is WAAAAAYYYY better, in both features and UI. I don't have a specific need for this now, but it's actually kinda nifty to see someone else do this.

Hello everyone, I am the developer of this app. I wanted to thank you for the appreciation and I hope you find this tool useful in your work.

There are a lot of things that I want to add to this app in the future - e.g. working with heat maps, path and area measurement, simple image editing tools etc. Also while the auto extraction stuff works ok at the moment, I think there is still a lot of room for improvement. I hope I can get to all these things soon.

If you have any suggestions, then feel free to comment here, or use the issue tracker on GitHub. You can also just send me an email.

Awesome tool :)

One minor thing, it could make sense to decrease the default colour distance in the example, as with the current default setting the orange and red lines are selected together in automatic mode when using the orange from the "dominant colors". A distance of 78 or less separates them.

Neat. I've been using GraphClick [0] for this for decades.

[0]: http://www.arizona-software.ch/graphclick/

This brings back memories. I used to use these kind of programs during my study electrochemistry more than 20 years ago. Apparently it is still relevant.

Amazing work! Can it process plots with one data set (one line) with an all white background?

Reply on news.ycombinator.com