基于特征解码的表情识别方法研究

Research on Expression Recognition Method Based on Feature Decoding

针对目前大多数面部表情提取未充分考虑语义特征以及个人独特面部特征导致面部表情识别准确性低的问题,提出一种基于特征解码的高效表情识别方法,称为FER-FD方法。该方法由两个模块组成,即特征解耦模块(FFD)和语义强化模块(VTS)。首先,FFD模块使用两个深度二维卷积神经网络从输入图像中提取面部和表情特征,面部特征解耦器将面部特征与表情特征解耦,以最大限度地减少个人独特面部特征的影响;其次,VTS模块采用两个关键思想以无监督的方式自动捕获面部运动,从而建立全局面部区域的深层语义信息;最后,将两个模块的特征串联起来,以更准确地预测样本的面部表情。实验结果表明,本文提出的特征解码方法在CK+数据集上获得了98.78%的准确率,对不同场景具有可扩展性和适应性,具有较好的泛化能力。

To address the issue of low accuracy in facial expression recognition due to insufficient consideration of semantic features and individual facial characteristics in most current facial expression extraction methods,a highly efficient facial expression recognition method based on feature decoding(FER-FD)is proposed.This method consists of two modules,namely the feature decoupling module(FFD)and the semantic enhancement module(VTS).Firstly,the FFD module employs two deep 2D convolutional neural networks to extract facial and expression features from input images,where the facial feature decoupler disentangles facial features from expression features to minimize the influence of individual facial characteristics.Secondly,the VTS module adopts two key ideas to automatically capture facial movements in an unsupervised manner,thereby acquiring deep semantic information of the global facial region.Finally,concatenating the features from both modules enables more accurate prediction of facial expressions of samples.Experimental results demonstrate that the proposed feature decoding method achieves 98.78%accuracy on the CK+dataset,exhibiting scalability,adaptability to different scenarios,and good generalization capability.