In this thesis, Facial Emotional Expression Recognition system based on Gabor Wavelet Neural Network, considering six points of face image to extract specific features of facial expression, is proposed. At the same time, novel feature separability criterion as an objective function and its learning methods based on Levenberg-Marquardt methods are proposed.
In learning stage, objective function for learning is a determinant of novel feature separability, which improves conventional concept of feature separability validity index. Through this process, Gabor filters’ parameters for obtaining separable features are learned, and it enables us to use any kinds of recognizer. In this paper, we use Fuzzy Neural Network Model to make an unsupervised and adaptive recognizer.
Learning Gabor Filters’ parameters with feature separability scheme means that heuristics in the stage of feature extraction can be excluded, and enables us to get more separable feature vector. Therefore, because features that are more separable are used, the recognition rate is high with Fuzzy Neural Network Model. This simplified and integrated Gabor Wavelet Neural Network has good performance and adaptation capability, and enables it to recognize facial expressions efficiently. The simulation result shows that recognition rate is 89% as general classifier and 97% as an adapted classifier.
본 논문에서는 얼굴 영상에서 6개의 특징점을 대상으로 하는 얼굴 표정 인식을 위한 가버 웨이블릿 신경망이 제안되었다. 그리고, 새로운 특징 분리 척도와 이를 목적함수로 하는 학습방법이 제안되었다.
학습은 기존의 특징 분리도의 단점을 개선하는 특징 분리 척도의 행렬식에 기반하여 이루어진다. 이러한 특징 추출부에 대한 학습과정을 통해서 분리도가 높은 특징을 추출하게 되며, 본 연구에서는 인식기로써 적응 성능을 가지는 퍼지 신경망 모델을 적용하였고, 89%의 일반화 성능과 97%의 개인화 성능이라는 높은 인식률을 얻었다.