 |
About AMP Lab Projects Downloads Publications People Links
Project - Illumination Normalization For Face Recognition
Top of this page
| Motivation and Goal |
General face recognition algorithms work well under normal, or good, illuminations. However, if the illumination gets "bad", then the algorithm gets confused and recognizes the image as the wrong person. To solve this problem, an illumination normalization algorithm is proposed as a pre-processing stage before face recognition. By doing the pre-processing, we can guarantee all images fed into face recognition are images under "good" illumination and therefore looking to increasing the recognition rate.
Statistical PCA-based shape-from-shading
The proposed system is shown as follows in Figure 1.
|
|
Figure 1. System overview of the illumination normalization for face recognition
At the upper stage, a set of images under known illuminations are trained to learn what human face shape should be like. A reflectance model is used here to relate the 3D face shape, the 3D illumination vector, and the resultant 2D image intensities to each other. At the lower stage, again using the reflectance model and the knowledge learned above, we estimate the most probable face shape lying in this 2D image, and then synthesize a new image under the known, desired new illumination.
Experiment
Figure 2 shows the images before and after illumination normalization. It is clear that the illumination variation has been compensated and the personal identity can be better recognized.
|
Original Image |
 |
 |
 |
 |
| Synthesized Image |  |
 |
 |
 |
Figure 2. Sample images. Upper: Original poor-illuminated images, Bottom: Synthesized images
Any suggestions or comments are welcome. Please send them to Yufeng Jessie Hsu.