Author Affiliations
Physics and Optoelectronic Engineering College, Xiangtan University, Xiangtan, Hunan 411105, Chinashow less
Fig. 1. Fusion images at four scales. (a) Scale 1; (b) scale 2; (c) scale 3; (d) scale 4
Fig. 2. Flow of AWCHOG algorithm
Fig. 3. ORL face dataset. (a) Image 1; (b) image 2; (c) image 3; (d) image 4; (e) image 5; (f) image 6
Fig. 4. AR face dataset. (a) Training sample; (b) facial express subset; (c) illumination subset; (d) partial occlusion subset A; (e) partial occlusion subset B
Fig. 5. CAS-PEAL face dataset. (a) Training sample; (b) Express subset; (c) Background subset; (d) Accessory subset
Fig. 6. Recognition rate of ORL face dataset at different gradient directions
Fig. 7. Recognition rate of ORL face dataset in different block modes. (a) Coarse floor; (b) detail 1 floor; (c) detail 2 floor; (d) fine floor
Coefficient level | Coefficient matrix form | Dimension |
---|
Coarse | 19×15 matrix | 285 | Detail 1 | 1×8 cell consisting of the matrix of size 17×31 or 37×15 | 4328 | Detail 2 | 1×16 cell consisting of the matrix of size 32×31 or 28×30 or 38×26 or 38×23 | 14776 | Fine | 112×92 matrix | 10304 |
|
Table 1. Form of Curvelet coefficients after Curvelet transform for 112 pixel×92 pixel face image
Scale | Recognition rate /% |
---|
ORL | AR | CAS-PEAL |
---|
1 | 90.00 | 91.00 | 98.14 | 2 | 92.50 | 92.33 | 98.54 | 3 | 93.00 | 94.67 | 98.89 | 4 | 95.50 | 95.67 | 99.20 | 5 | 89.50 | 90.67 | 95.94 |
|
Table 2. Recognition rate of Curvelet transform at different scales
Algorithm | Recognition rate /% |
---|
Image 2 | Image 3 | Image 4 | Image 5 | Image 6 |
---|
Wavelet | 83.13 | 88.21 | 87.92 | 93.74 | 96.50 | HOG | 85.00 | 89.29 | 92.50 | 97.00 | 96.88 | Curvelet+PCA+SRC | 88.21 | 93.33 | 94.00 | 95.00 | 95.83 | Gabor+HOG | 90.41 | 94.61 | 97.21 | 98.20 | 98.69 | NSCT+LBP | 89.14 | 94.79 | 96.20 | 98.11 | 98.69 | AWNHOG | 82.25 | 87.21 | 91.25 | 95.00 | 96.88 | AWCHOG+CRC | 90.12 | 86.79 | 92.50 | 96.50 | 97.33 | AWCHOG+KNN | 90.71 | 95.50 | 96.88 | 98.75 | 99.27 |
|
Table 3. Recognition rate on ORL database
Algorithm | Recognition rate /% |
---|
Expression subset | Accessory subset | Background subset |
---|
Wavelet | 96.45 | 78.61 | 78.89 | HOG | 99.20 | 69.00 | 93.00 | Curvelet+PCA+SRC | 97.34 | 78.89 | 80.83 | Gabor+HOG | 99.73 | 87.22 | 95.56 | NSCT+LBP | 99.38 | 74.25 | 92.75 | AWNHOG | 99.63 | 80.50 | 99.00 | AWCHOG+CRC | 99.33 | 98.67 | 95.00 | AWCHOG+KNN | 99.75 | 97.50 | 98.50 |
|
Table 4. Recognition rate on CAS-PEAL database
Algorithm | Recognition rate /% |
---|
Facial expresssubset | Illuminationsubset | Partial occlusionsubset A | Partial occlusionsubset B |
---|
Wavelet | 92.67 | 80.50 | 79.50 | 62.00 | HOG | 91.33 | 90.33 | 69.00 | 49.00 | Curvelet+PCA+SRC | 94.67 | 85.50 | 91.33 | 73.67 | Gabor+HOG | 96.33 | 97.50 | 81.00 | 76.50 | NSCT+LBP | 96.67 | 98.33 | 96.67 | 76.00 | AWNHOG | 98.33 | 99.67 | 96.67 | 82.67 | AWCHOG+CRC | 95.33 | 96.67 | 93.67 | 80.33 | AWCHOG+KNN | 99.50 | 98.89 | 96.00 | 90.00 |
|
Table 5. Recognition rate on AR database
Method | Normalized variance of Gaussion white noise | | φ |
---|
0 | | 0.0001 | 0.0002 | 0.0003 | 0.0004 | |
---|
Wavelet | 80.50 | 39.00 | 21.50 | 16.67 | 14.67 | 81.78 | HOG | 90.33 | 68.33 | 58.67 | 53.67 | 52.33 | 42.06 | Curvelet+PCA+SRC | 85.50 | 71.67 | 68.33 | 67.67 | 65.67 | 23.19 | Gabor+HOG | 97.50 | 86.00 | 84.67 | 81.00 | 79.33 | 18.63 | NSCT+LBP | 98.33 | 91.50 | 85.50 | 74.00 | 70.50 | 28.30 | AWNHOG | 99.67 | 93.67 | 90.33 | 83.67 | 79.67 | 20.06 | AWCHOG+KNN | 98.89 | 98.00 | 96.67 | 91.67 | 90.67 | 8.31 |
|
Table 6. Results of different algorithms on AR illumination subset after adding Gaussian noise
Method | Featuredimensionality | T1 /ms | T2 /ms |
---|
Wavelet | 1920 | 16.4 | 14.2 | HOG | 320 | 4.5 | 38.1 | Curvelet+PCA+SRC | 2850 | 33.2 | 41.4 | Gabor+HOG | 9600 | 163.3 | 98.9 | NSCT+LBP | 3696 | 856.3 | 30.1 | AWNHOG | 2048 | 15.9 | 38.3 | AWCHOG+KNN | 2272 | 14.5 | 35.7 |
|
Table 7. Dimensionality and time of different algorithms on ORL face dataset