The talk describes conception and evaluation algorithms for embedded computer vision applications on the next generation mobile telephones. The integration of image sensors and embedded image processing has profoundly changed the market for mobile telephones. Mobile telephones increasingly integrate two separate image analysis systems into their basic functionality. An outward looking camera is used to not only as a recording device for photographs and films, but increasingly as smart sensor for detection and recognition of text and objects for applications such as road sign translation, business card capture, document capture, image based internet search, and text to speech. A second back-facing camera, initially implemented for two-way video communications, increasingly used for human-computer interaction and emotion capture. Embedded computer vision is increasingly emerging as the key enabling technology to competitive devices for this market.
Within project MinImage, INRIA has demonstrated a novel O(N) binomial pyramid algorithm for real time view invariant description of images. In previous deliverables, this algorithm was shown to be suitable for implementation using embedded hardware on the image pipeline, providing real time image descriptors suitable for a variety of embedded vision applications. The resulting image descriptors where shown make it possible to trade memory for computation, leading to improved performance for applications such as face detection.
We report on results for use of this technique for Real time detection and description of scale and orientation invariant interest points,Face Detection, Face recognition, Age estimation and Detection of facial action codes for emotion recognition. The experimental performance evaluation demonstrates the suitability of the resulting software for use in embedded image analysis on mobile telephones.