A wavelet-based approach to face verification, recognition for mobile phones
The rise of international terrorism and the increase in fraud and identity theft have added urgency to the task of developing biometric-based personal identification as a reliable alternative to conventional authentication methods. Face recognition schemes have been developed and deployed to control access to sensitive sites.1 Their performance is affected by variations in facial images related to illumination, pose, occlusion, facial expressions, and scale. Other factors influencing the accuracy of face recognition include changing facial features as a result of age progression or facial hair. Iris, and to some extent, fingerprint-based identification have superior accuracy rates, yet because of their unobtrusive nature, face and voice recognition are the preferred and more acceptable methods of identification for security applications.
On top of this, biometric systems of the future will have to be capable of serving a large and mobile user population. Deciding where to deploy such systems will be based on cost-effectiveness, user acceptability, and the presence of supporting infrastructure. Mobile communication devices can provide an adequate solution: current 3G smart phones and PDAs are equipped with video and touch-sensitive display panels. In addition to convenience, such devices provide cost-effective support for financial or other sensitive online transactions. Biometric authentication systems on mobile devices can complement cryptographic functions and provide added protection against fraud and repudiation of online financial transactions while ensuring accountability. Such systems also have advantages in conflict scenarios when soldiers must act from behind enemy lines and cannot rely on traditional secure communication infrastructure.
Most existing biometrics are not particularly well-suited for use in platforms that are constrained in memory, computational power, battery power, and transmission rate. Currently available PDAs have modest processing speeds of up to 400MHz and just 128MB of RAM, and SIM cards have a small memory (up to 256K). Therefore it is essential to develop efficient, scalable, and highly accurate algorithms in this area.
The European-funded SecurePhone project, in which we participate, is focused on developing and implementing face and voice biometrics on mobile phones and PDAs. SecurePhone's multi-modal, biometric recognizer is based on fusing audio-visual and handwritten signature identification schemes, and facilitates access to the PDA's e-signing facilities to e-sign and send data during phone calls. Face recognition requires efficient algorithms to compute and analyze digital representation of facial features.
A typical face image is represented by a high dimensional array (192 × 224 pixels). To reduce the so-called ‘curse of dimensionality’, dimension-reducing techniques have been developed to represent images in much lower dimensional space without losing information. One of the first and most commonly used of these techniques is principle component analysis (PCA),2 which is based on the idea that face vectors do not scatter randomly in their infinite high-dimensional universe, but rather appear in clusters. While PCA is good at compactly representing face feature vectors, it does not have good discrimination capability.
Recently, wavelet-transforms have been used in face recognition, whereby a face image is represented by the wavelet coefficients in some subbands of the transformed images.3,4 Wavelet transforms are multi-resolution image decomposition tools that provide a variety of channels representing the image features by different frequency subbands at different scales (see Figure 1). Wavelet-transforms have been used extensively for image compression, and a large number of wavelet filters have been developed. The size of a wavelet subband is a fraction of the original face image size. For any wavelet filter, depth of decomposition and a frequency subband can give rise to a feature vector representation of a face. The size of a wavelet subband channel at depth k is 4-k of the original image size, making such features suitable for implementation on PDAs.
We have developed wavelet-based face verification schemes, using three different wavelet filters and up to three different subband channels at wavelet decomposition depths of three or four. The performance of these schemes is comparable to sophisticated state-of-the-art face verification schemes that require computational resources far beyond the capabilities of constrained platforms such as PDAs. SecurePhone's biometric recogniser is based on fusing audio-visual and handwritten signature identification schemes. Biometrically authenticated users can get direct access to the PDA's built-in facilities to e-sign and send data concurrently with voice during a normal phone call.
The work on SecurePhone has progressed very well and many technical challenges have been solved to enable real-time voice and face verification. The audio and video sampling rates have been significantly increased to 44KHz and approximately 20 frames per second (fps) respectively, compared to the rates possible on an adapted PDA (8KHz and 10fps). Online synchronization of audio and visual signals has been achieved as well. The performance of each individual modality and the fused system is encouraging, and a new audio-video database of indoor and outdoor recordings has been recorded using a currently available PDA.
