Diploma work: Monthly report 8/97

Review August 1997

I did the following things during the last month:

• To improve the detection quality I had to change the correction of the delta images. Before I used only a look-ahead of 1 delta image to correct moving black parts. But if a moving part is large (e.g. 3 to 5 pixels thick, like a pen tip) it covers itself in a movement. Because I only had a look-ahead of 1 frame the wrongly black pixels still were black. This pixel would only turn white whitin 2 or 3 frames.
  

  At the moment I use a look-ahead of 3 delta images and the results got OK.

• I could now get sense out of the return parameters of Otsu. My tests and comparisions showed that the variance is the best parameter to decide whether the resulting threshold is useful or not. If the analysed image contains dark and white parts the variance is high (mostly over 100). If the analysed image contains only bright pixels the varinace is small (mostly below 50). Actually I calculate the Otsu threshold for every delta frame and based on this value I turn the image black and white to apply one of the detection routines. With this approach most of the misdetections could be avoided.
• To improve the Otsu algorithm I patch the histogram of the delta image. Because most of those delta frames consist of bright or even white pixels I remove the brightest part of the Histogram. With this approach the histogram gets more weight on the darker pixels.
• To improve the testing and debugging of the detection process I implemented an interface to write HTML output with debug informations. I get there a huge table with the Otsu parameters, the delta image, the Histogram and the resulting pen position. Some of those reports can be seen under http://www.siebenthal.ch/diplom/examples/.

Two examples:	Detected Linestrokes in July:
	Original image masked:
	Resulting Linestrokes:
	Detected Dark Pixels:

Preview September 1997

My goals for September 1997 are:

• The treating of Otsu still needs some improvement. The problem is at the moment sometimes the limit of the allowed variance and sometimes the resulting threshold is too bright.
• Some character show problem to be detected well. E.g. the letter "w" is problematic. I write this letter like two combined "v". But after the first downstroke of the "v" I start the upstroke to finish the "v". But during this upstroke the pen covers the new written ink. This ink is only detected during the downstroke of the second "v" and is nearly the same as the downstroke of the 2^nd "v". This results in a wrong order of the written line segments of this letter. I don’t see a solution for this problem at the moment, because I think I’m missing data to detect and solve this problem. I’ll have to investigate it.
• The interpolation routine needs some improvements. At the moment only gaps of 1 or 2 positions can be fixed. I will try to fix gaps that can be larger. But there will be a limit of the interpolated distance per missing position.
• My program has still a memory leak. It’s smaller than before but still has to be fixed.
• As soon as the memory leak is fixed I’ll check sequences of frames that represent a full line of written text.
• If the above showes satisfiying results I’d like to check a slightly new approach. I’d like to analyse the greylevels of a single pixel or a set of pixels over all the frames. In the optimal the greylevels should be white until the ink covers this pixel and then stay black. If this pixel is coverd by the pen in the meantime there will be a short disturbance of the graph.
  

Longer time goals with less priority:

• I will capture some more sequencies with pens that have different ink colors (red, green, grey) and with other writers than myself.
• Analyse the influence of the ETHZ input interface to my detection process and my output interface.
• Think about the automatic getting of the mask after the writer completed a line of text.
• Read about projective geometry to be able to reproject the captured image into a normalized view of the paper. With this I should be able to detect the position of the paper and normalize into the point of view of a scanner or a graphics tablet.

zurück zur Uebersicht der Diplomarbeit

Last modified 23.12.1999 00:18:51 by Thomas von Siebenthal