Reading #14: Shape vs. Text. The Ultimate Showdown

Comments:
Sam

Summary:
Text vs. Shape returns in this paper, with an all star ink feature paving the way to high accuracy.
Entropy is the measure of the degree of randomness in a source (that is from the paper, I swear). Text strokes are normally more dense than shape strokes, thus giving them a higher entropy. And from here, the authors go beast mode.
The authors created an entropy model 'alphabet' that is used to assign a symbol to the angle a points makes with its two adjacent points in a given stroke. Printing out the entrobet (now that one I made up) provides a visual cue as to the changes that a stroke undergoes in terms of curvature. The points are measured 4 pixels apart and the assigned values are averaged over the bounding box of the stroke in order to ensure scale independence. Testing data was measured for the percentage of strokes that were classified as shape or text, and the accuracy of said classifications. Overall, the entrobet had an accuracy of 92.06%.

Discussion:
I did not know this was an SRL paper until they began talking about SOUSA in the data collection section (I skipped over the author's names somehow). Good thing I didn't say anything bad about it! But seriously, entropy as a measure for shape vs. text can be deemed a goto option based on the research presented in this paper. The issue of dashed lines in the COA data set accounted for a high level of the errors, so including a system that can pre-process out these dashes lines would lead to even greater accuracy. How would you remove shapes made up of dashes?