【摘要】 在自然环境中各种因素的干扰下,人脸表情信息匹配的识别率受到严重影响,针对此问题,提出一种改进的基于VGGNet16(visual geometry group network16)的网络模型。在VGGNet16模型的侧方添加一系列的侧输出层,并在该侧输出层添加不同的卷积核,通过上采样和下采样方法连接侧输出层的上下2层,并通过训练使侧输出层能够对其上下2层的表情信息进行加权融合。在VGGNet16第5层的后方添加2种不同的卷积核。将侧输出层最终得到的特征图进行局部卷积操作,将VGGNet16输出的最终特征图进行全局特征卷积操作,使局部特征与全局特征融合得到最终要进行分类的特征。该模型在CK+(the extended cohn-kanade)数据集上的识别率为98.6%,在RAF-DB(real-world affective faces)数据集上的表情识别率为79.59%,通过对比常用模型在这2种数据集上的识别率发现该模型具有一定的优势。更多还原

【Abstract】 The recognition rate of facial expression information matching is seriously affected by various factors in the natural environment. Given this problem, the paper proposes an improved network model based on VGGNet16(visual geometry group network16). Firstly, a series of side output layers are added on the side of the VGGNet16 model, and the different convolution kernels are added in this side output layer, and then the upper and lower layers of the side output layer are connected by oversampling and undersampling, and through training, the side output layer can perform weighted fusion on the expression information of the upper and lower layers. Two different convolution kernels are added behind the fifth layer of VGGNet16, and the final feature map outputted by the side output layer executes local convolution operation, and the final feature map outputted by VGGNet16 executes global feature convolution operation. Finally, the local features are combined with the global features to be classified ultimately. The recognition rate of the model on CK+(the extended cohn-kanade) data set is 98.6%, and the expression recognition rate on the RAF-DB(real-world affective faces) data set is 79.59%. By comparing the recognition rates of common models on these two data sets, the model has certain advantages.更多还原