Some thoughts on Mobile Face Recognition (part 4 - Embedded processing)

Being able to embed the biometric processing in the mobile devices has gained a lot of interest for face recognition systems. Some of the advantages of the algorithms being embedded in the device are:

  • Reduced volume of data exchanged over the network, since sending the video stream to a server is not needed. This allows an off-line use of the system and also response time is reduced.
  • The privacy of the users is better preserved, since their biometric data stay in their own devices.
  • Scalability: the computational power of the server (for biometric template extraction) does not need to grow with the number of users as much as if the processing were performed in the server.
  • Some recent interoperability standards for online identification, like the one proposed by the FIDO Alliance, require a secure unlock operation (biometric or not) to release the cryptographic keys. This is accomplished through a safe action, such as the use of biometrics, but the biometric information is required to never leave the user device, so embedded biometric processing is mandatory.

However, an essential challenge remains on mobile face recognition scenario. The duration of the battery remains to be one the biggest weaknesses in mobile devices. Since energy efficiency is not a problem for traditional server-based face recognition systems, it is usually overlooked (at least, more than it would be desired for mobile scenarios). Nevertheless, it is a key issue in mobile embedded face recognition systems, so a broad study on more efficient algorithms, parallel computing optimization and exploitation of the hardware resources need to be done, as recent works point out.

An interesting topic related to the above point is the implementation of the recent Deep Neural Network paradigms for face recognition into mobile devices, taking advantage of the embedded GPU and exploiting its capabilities for energy-optimized real-time processing. Some questions to solve are how to design a proper net architecture for mobile computing or the effects of feature representation and net dimensionality on mobile face recognition accuracy.