Check out a new model for face recognition software from SPECS!
Many mammals have an amazing ability to recognize objects under distinct conditions. Tasks that may seem simple are in fact only possible thanks to the great complexity of the mammalian cerebral cortex. One of the most complex stimuli mammals are challenged by are faces.
Natural vision systems remain invariant to things like shifts in position, rotation, and scaling so that if you were to look at a picture of your friend’s face upside down, it wouldn’t cause you to mistake them for somebody else!
Due to the increasing use of social robots, scientists today are keen on advancing facial recognition software. In order to do this, researchers have been working on models of natural visual systems; however, turning the effortless intricacies of the cerebral cortex into hard wires and code is tricky business. A lot of facial recognition software systems are based on a large dictionary of features stored in memory that must be filtered by artificial neurons through out various layers. Through these means, invariance to factors such as position and scale can be achieved but it’s often at the high cost of increasing the number of connections between the layers of the network which results in bulkier hardware.
To tackle this issue, researchers from the SPECS lab at Pompeu Fabra University employed a model of an encoding scheme called the Temporal Population Code (TPC) in an artificial system. Applied as part of the facial recognition feature of the iCub robot, the TPC model has important advantages over many other models used in the past. It is wire-independent, allowing the system to accommodate various streams of complex information and it can also be used as a generic framework for object recognition tasks regardless of input. Beyond facial recognition, the TPC model can be generalized to other tasks such as hand writing recognition.
The iCub seems to be a fan of the TPC model. In the face recognition tasks, its speed of encoding was compatible with the human visual system with a correct ratio of 97 % ! These results are included in a paper titled: The encoding of complex visual stimuli by a canonical model of the primary visual cortex: temporal population coding for face recognition on the iCub robot by: Andre Luvizotto, Cesar Rennó-Costa, Ugo Pattacini and Paul Verschure. It was one of the five finalists for best conference paper in the 2011 International Conference on Robotics and Biomimetics
