Recognition of facial expression by using neural-network system with fuzzified characteristic distances weights

A neural-network with fuzzified characteristic distances weights (NNFCDW) is proposed in this paper to recognize the facial expressions effectively. During the recognition process, the characteristic distances that represent the relationship between the facial expressions and the muscle movement are used to be the major basis for recognition. The different expressions will somehow dominate the characteristic distances defined from different feature-area (mouth, eye or eyebrow). Therefore, the weights of the characteristic distances will be an important factor to determine the recognition rate. In this paper, a reasonable method of tuning the weights without trial-and-error is proposed. A fuzzy system based on the recognition results is developed to generate the weights rationally. The characteristic distances are multiplied with the fuzzified weights and sent to a neural-network system for recognition of the facial expressions. The proposed neural-network system is composed of the self-organizing map (SOM) neural network and back-propagation neural network (BPNN). When BPNN used the pre-classified data as its training data, the training cycles can be obviously reduced. The experimental results demonstrate that the recognition rate of using the proposed NNFCDW obviously increased about 10% ~ 13% as comparing with the results obtained by using pure BPNN. The computational time of using the proposed NNFCDW is also effectively decreased about 60% as comparing with the results obtained by using pure BPNN.