Search engine :
Return to the menu
| : /
Vote:
Results:
212 Votes
JANUARY-DECEMBER 2019 - Volume: 6 - Pages: [17 p.]
Download pdf
ABSTRACTThis research modifies the training strategy in the first level of the neocognitron of 2003, using symmetric gabor filters, highlighting the excitatory and inhibitory zones that will allow extracting the characteristics more accurately as weights of the S cells of the first level, improving its recognition rate by 0.5%. To verify the results have been applied to the problem of recognition of handwritten digits, with a set of training of 5000 patterns and with tests of 8000 patterns extracted from the database of digits MNIST (Modified National Institute of Standars and Technology). The results obtained in the experiments show that the maximum recognition rate of the neocognitron with changes in the first level is 99.02%, while the maximum recognition rate for the neocognitron of 2003 is 98.5%. Also, the modified model has an efficiency of 83% with respect to the original model and in 35% of the speed of execution with respect to the original neocognitron, temporal complexity was also reduced by 64%.Keywords: Neocognitron, Gabor filter, oriented edges, handwitten digit recognition, recognition cup.
Share:
© DYNA New Technologies Journal
EDITORIAL: Publicaciones DYNA SL
Adress: Alameda Mazarredo 69 - 2º, 48009-Bilbao SPAIN
Email: info@dyna-newtech.com - Web: http://www.dyna-newtech.com
Regístrese en un paso con su email y podrá personalizar sus preferencias mediante su perfil
Name: *
Surname 1: *
Surname 2:
Email: *