Sparse Representation-based Object Recognition - Sparse Representations and Compressive Sensing for Imaging and Vision - page 66

Tabl e 5. 1 Recognition Rates
(in%)ofSRCAlgorithm
[156] on the Extended Yale B
Database

Dimension 30 56 120 504

Eigen 86.5 91.63 93.95 96.77

Laplacian 87.49 91.72 93.95 96.52

Random 82.60 91.47 95.53 98.09

Downsample 74.57 86.16 92.13 97.10

Fisher 86.91 - - -

Fig. 5.2 Examples of partial facial features. (a) Eye (b) Nose (c) Mouth

Tabl e 5. 2 Recognition
results with partial facial
features [156]

Right Eye Nose Mouth

Dimension 5,040 4,270 12,936

SRC 93.7% 87.3% 98.3%

NN 68.8% 49.2% 72.7%

NS 78.6% 83.7% 94.4%

SVM 85.8% 70.8% 95.3%

Partial face features have been very popular in recovering the identity of human
face [135], [156]. The recognition results on partial facial features such as an eye,
nose, and mouth are summarized in Table 5.2 on the same dataset. Examples of
partial facial features are shown in Fig. 5.2 . The SRC algorithm achieves the best
recognition performance of 93

3% on eye, nose and mouth features,
respectively and it outperforms the other competitive methods such as Nearest
Neighbor (NN), Nearest Subspace (NS) and Support Vector Machines (SVM).
These results show that SRC can provide good recognition performance even in
the case when partial face features are provided.
One of the main difficulties in iris biometric is that iris images acquired from
a partially cooperating subject often suffer from blur, occlusion due to eyelids, and
specular reflections. As a result, the performance of existing iris recognition systems
degrade significantly on these images. Hence, it is essential to select good images
before they are input to the recognition algorithm. To this end, one such algorithm
based on SR for iris biometric was proposed in [112] that can select and recognize
iris images in a single step. The block diagram of the method based on SR for iris
recognition is shown in Fig. 5.3 .
In Fig. 5.4 , we display the iris images having the least SCI value for the blur,
occlusion and segmentation error experiments performed on the real iris images in
the University of Notre Dame ND dataset [14]. As it can be observed, the low SCI

. 7% , 87 . 3% , 98 .