To distinguish the sharpness, the difference of the running speed along the inner and outer contours near the corner is utilized. For a point on the inner contour, the counterpart on the outer one is decided by the nearest neighbor and the correspondence between the two is made. Since the point on the inner contour runs more slowly, there will be more zero's in its difference sequence made from pixels on contour correspondence list. Thus the bit patterns representing the contour correspondence can be used as a feature vector for the corner sharpness.
Experiments using a trainable three-layered neural network were executed. In these experiments, three sets were used. The first set contained about 400 curves having sharp and round corners written by hand. About 300 samples were used in the training and the remaining 100 were used in the test. The recognition rate for the training and test sets were 100% and 96%, respectively. The second and third sets included actual handwritten characters, which were `U-V' and `(-<' pairs. A set consisting of 50 patterns for each pair was used in training, and another set consisting of 50 patterns was used for the test. All patterns in training and test sets were recognized correctly.
[Neural Network Applications] [Research Themes of Prof. Nakano.]
| mail address: | |
← お手数ですが打ち込んで下さい |
First Written Before June 17, 1998
Transplanted to KSU Before June 19, 2003
Transplanted to So-net April 22, 2007
Last Update April 22, 2007
© Yasuaki Nakano 1998-2007