TY - GEN
T1 - Performance comparison of laplacian-based features
AU - Khabou, Mohamed A.
AU - Rhouma, Mohamed B.H.
AU - Hermi, Lotfi
PY - 2008/12/1
Y1 - 2008/12/1
N2 - Three feature sets based on the eigenvalues of the Laplacian operator with different boundary conditions were generated and their computational complexity and effectiveness in a pattern recognition application were compared. The first feature set is based on the eigenvalues of the Laplacian operator with Dirichlet boundary condition, the second feature set uses Neumann boundary condition, and the third set uses Stekloff boundary condition. All feature sets are rotation, translation, and size invariant. The effectiveness of these features is demonstrated by using them in the classification of 5 types of binary hand-drawn shapes. The classification was done using 4 to 20 features fed to a simple feed-forward neural network. Even though the Dirichlet and Neumann feature sets were computationally more complex than the Stekloff features, they were more tolerant of input variations and clearly outperformed the Stekloff feature sets in the pattern classification application. The correct classification rates of the Stekloff feature sets ranged from 34.0% to 61.0% while those of the Dirichlet and Neumann feature sets were 60.0%-95.5% and 87.5-95.5%, respectively.
AB - Three feature sets based on the eigenvalues of the Laplacian operator with different boundary conditions were generated and their computational complexity and effectiveness in a pattern recognition application were compared. The first feature set is based on the eigenvalues of the Laplacian operator with Dirichlet boundary condition, the second feature set uses Neumann boundary condition, and the third set uses Stekloff boundary condition. All feature sets are rotation, translation, and size invariant. The effectiveness of these features is demonstrated by using them in the classification of 5 types of binary hand-drawn shapes. The classification was done using 4 to 20 features fed to a simple feed-forward neural network. Even though the Dirichlet and Neumann feature sets were computationally more complex than the Stekloff features, they were more tolerant of input variations and clearly outperformed the Stekloff feature sets in the pattern classification application. The correct classification rates of the Stekloff feature sets ranged from 34.0% to 61.0% while those of the Dirichlet and Neumann feature sets were 60.0%-95.5% and 87.5-95.5%, respectively.
KW - Dirichlet
KW - Laplacian egenvalues
KW - Neumann
KW - Stekloff
UR - http://www.scopus.com/inward/record.url?scp=62749100021&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=62749100021&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:62749100021
SN - 1601320787
SN - 9781601320780
T3 - Proceedings of the 2008 International Conference on Image Processing, Computer Vision, and Pattern Recognition, IPCV 2008
SP - 39
EP - 43
BT - Proceedings of the 2008 International Conference on Image Processing, Computer Vision, and Pattern Recognition, IPCV 2008
T2 - 2008 International Conference on Image Processing, Computer Vision, and Pattern Recognition, IPCV 2008
Y2 - 14 July 2008 through 17 July 2008
ER -