100 likes | 247 Views
Digit Recognition Using SIS Testbed. Mengjie Mao. Cycle 1: sequential component. Cycle 2: sequential components. Ten weight matrixes for next cycle. Identifier 0. AAM training. Identifier 9. Ten perfect digit image for training
E N D
Digit Recognition Using SIS Testbed Mengjie Mao
Cycle 1: sequential component Cycle 2: sequential components Ten weight matrixes for next cycle Identifier 0 AAM training Identifier 9 Ten perfect digit image for training Randomly generated digit images with defects for testing Overview ……
Cycle 1: AAM training • Training input • Hand-written pixel images for digit 0~9, the size is 20x15 • Implement a tool outside Testbed to extract the pixel data, by which a image can be represented by a binary matrix with size 20x15(1 for black, 0 for white) • Training output • Ten weight matrixes, each of which is for one digit • Algorithm • Hopfield network • Give up the MNIST dataset
Cycle 2: Testing • Testing input • Randomly generate the digit images with defect pixels (black<->white) • Testing procedure • Each digit identifier initializes its owe weight matrix which is output from cycle 1 • All identifier take a defected image as input • The best identifier is the one which converges fastest
Demo & Conclusion • With 1 defect pixel • With 15 defect pixels • With 45 defect pixels • The digit recognition prototype can be used for any recognition tasks: face, car, fingerprint…… • Using parallel components in cycle 2 is more preferable
Acknowledgment • Thanks for Prof. Chang’s lecture and consultancy • Thanks for Haifeng Xu and AngenZheng, for their helps on the Java program