Constructive Hermite polynomial feedforward neural networks with application to facial expression recognition
Liying Ma, K. Khorasani · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2001
Computer-based recognition of human facial expressions has been an active area of research since the 1970s. The ultimate goal is to realize intelligent man-machine interface. Recently, constructive One-Hidden-Layer Feedforward Neural Networks (OHL-FNNs) have been found promising for facial expression recognition. The hidden units in a FNN usually have the same activation functions typically selected as sigmoidal functions. However, it has not been proven that the use of the same activation functions for all the hidden units is the best or optimal choice in terms of network performance. In this paper, a new constructive polynomial OHL-FNN is proposed for pattern recognition. The well-known Hermite polynomials will be used as activation functions for the hidden units. Each time a new hidden unit is to be added to the network, a Hermite polynomial whose order is increased by one will be used as the activation function of the hidden unit. The proposed technique is applied to the facial expression recognition problem where the 2D DCT is performed over the entire face image before the resulting lower 2D DCT coefficients are fed to the constructive network training. The advantages and limitations of the constructive polynomial OHL-FNN for pattern recognition are also discussed.