基于新双模融合算法的情感检测系统

Emotion Detection System Based on New Double-Mode Fusion Algorithm

针对传统的分类器融合存在的诸多问题,提高情感检测正确率,采用双模态(音频、视频)参数提取,选择差异性强的组合小波神经网络(MWNN)与混合高斯模型(GMMs)分类器.在语音韵律、音质特征与人脸几何特征提取后,对提取后的特征用主元分析法(PCA)进行降维,对分类器进行匹配化输出,最后引入GA算法来搜索最优的融合系数向量,充分发挥各分类器本身对特定情感的敏感特性.实验证明,与传统的融合算法相比,经匹配化的GA融合算法将识别率提高了4%～10%,具有更高的识别率与更强的泛化能力.

Aiming to solve the problems existing in traditional classifier fusion and improve the correct rate of emotion detection system, feature parameters were extracted from both video and audio, and module wavelet neural network （MWNN） and Gaussian mixture models （GMMs） were chosen as classifiers to maintain diversity for effective fusion. Prosodic features of speech signal and geometric features of facial expression image were extracted after preprocessing, and then principle component analysis （PCA） was adopted to reduce the dimensions of eigenvectors. After the outputs of MWNN and GMMs were matched,genetic algorithm （GA） was introduced to search for the optimal fusion coefficient vector, which would give full play of the two different classifiers＇ susceptibility to specific emotion. The experinental results show that, compared with the traditional fusion algorithm, the output-matched GA fusion algorithm has higher recognition rate and stronger generation ability, improving recognition rate by about 4% to 10%.