From Canonical Face to Synthesis - an Illumination Invariant Face Recognition Approach
Tele Tan · 2007
The results obtained using the canonical face recovery algorithm is very encouraging. We have shown that the minimum number of illumination specified bootstrap images (i.e. N) needed to generate a stable canonical face ranges between 3 and 5. This makes good hardware design sense as an elaborate light stage setup (Sim et al., 2003) (Debevec et al., 2000) becomes unnecessary. Currently we are fabricating a low-cost portable light array module used to implement the canonical face recovery. Depending on the role, this lighting module can be embedded into the different stages of the face recognition system. For example, the module can be introduced during the registration stage where the objective is to capture a good quality neutral image of the subject (i.e. its canonical representation) to be registered into the system irregardless of the ambient lighting condition. Another possible use is to incorporate it into the image capture system at the front end of the recognition system. This will ensure that the picture taken and used for matching will not be affected again by external light sources. Besides using the images captured by the lighting module as described here, we can explore using shape-from-shading techniques to recover the 3D shape of the face (Zhao and Chellappa, 1999). The range information will be an added boost to improve on the illumination rendering quality as well as for recognition. Although the illumination models recovered using the CMU PIE database generates 21 different variations they are inadequate as some important lighting directions (i.e. especially those coming from the top) are glaringly missing. We will next consider using computer graphics tools to develop a virtual light stage that has the ability to render any arbitrary lighting conditions on a 3D face. These new variations can then be used to extract finer quality illumination models which in turn can be use to synthesis more realistic novel appearance views. Finally, we will explore how the systems of canonical face recovery and appearance synthesis can play a further role in enhancing the performances of illumination challenged real world analysis systems. One possible use of this would be in the area of improving data acquisition for dermatology-based analysis where maintaining colour integrity of the image taken is extremely important.