Identify and remove text from graph

I'm in the process of designing a semi-automatic graph digitizer that can convert screenshots of a graph into a CSV file with the data.

I'm running into an issue when certain graphs contain characters and text boxes in the graph space, as in the example images below:

After removing the gridlines and producing a binary image "curve" (that's supposed to contain just the curve), my current code uses

[y, x] = find(curve)

to determine the x and y position of points on the curve. However, if there is any text on the graph, the code thinks it's part of the data curve and finds points in its location. As in the above examples, the code "finds" datapoints where the 'Conditions' and '3C96' text is located.

My question is: Is there any way to automatically detect and mask over this text, perhaps using OCR? Or is automating it a lost cause, and I should instead manually crop out any text initially?

Answer 1

[Image Analyst]

If the graphs are digital images, you can simply use bwareafilt() in the Image Processing Toolbox to take either the largest blob, or blobs within a certain size range. You might need to add some code options to handle cases where grid lines can be solid or dashed, whether text is in boxes or not, or whether you want the gridlines to remain in the output image or not.

By the way, there are File Exchange entries already:

https://www.mathworks.com/matlabcentral/fileexchange/?term=tag%3A%22digitize%22

[Carl Youel]

I'll give that a shot. Seems like there won't be an easy solution given the variability of graph styles.

I've seen the File Exchange scripts, and they are indeed helpful to a certain degree. However, it seems all of them require you to place clicks point by point. My script seeks to do this process semi-automatically, where the user need only upload their graph and select a few options. I hope to post it on the File Exchange when I end up finishing it.

Thanks again

[Image Analyst]

If you have a consistent type of graph you're dealing with then it's possible to do it automatically. If however the graphs are wildly different, then developing an automatic solution may be more trouble than it's worth and it's best just to have user-assisted processing.