tripathy@hpscdc.scd.hp.com (Aurobindo Tripathy) (06/29/90)
/ hpscdc:comp.ai.neural-nets / SESHADRI@rcgl1.eng.ohio-state.edu (M D Seshadri) / 9:28 am Jun 25, 1990 / |Hello Neural netters, | I am relatively new to this field, yes I have read Lippman's paper, |but I need to do some serious pattern recognition, say 400 x 400 arrays, can |anybody out there suggest a good algorithm or method to achieve my objective. |This is part of my thesis, and any help would be greatly appreciated. I am |also familiar with Hough transform and edge detection ref Dana Ballard. If |this is not the proper group can any gentleman/lady suggest a cross posting. |Thanks people and I read this news group every day and also my mail. |---------- | I have not read Lippman's paper (I just borrowed it from our library) and my recommendation is to try some of the 'traditional methods'. By this I mean (1) Normalized Correlation. (2) Feature based matching. Normalized Correlation: Select a template (the same image in the case of auto-correlation) and then perform correlation either in binary or gray-scale. Binary would be faster if you can consistently create a global threshold or use a scheme for doing local thresholds. If you are doing OCR, then the template that gave you highest coefficient of correlation is the best match character. The key is to come up with a good and fast scheme for normalization. Feature based: Extract some features from the image like area, perimeter, x-projection y-projection ... First do some training in the n-dimensional feature space. Then do a nearest neighbour computation for the newly arriving image. Now its time to get back to dreary, mundane work. If you have further question about any of the this stuff, we can discuss further. If you come across a neural net scheme, let me know...meanwhile I will read the paper and give the problem some thought. /aurobindo tripathy