自适应重加权池化深度多任务学习的表情识别

Expression recognition based on adaptive reweighting pooling deep multi task learning

为实现表情准确且快速的识别,提出一种自适应重加权池化深度多任务学习(DMTL)的表情识别。设计孪生神经网络,通过自适应重加权模块动态调整缩放概率参数,得到具有不同置信度的类别标签信息。改进自适应池化方法,根据样本及训练过程变化情况选取合适参数,提高特征提取的灵活性。结合类别标签信息和样本局部空间分布信息,利用改进型判别式DMTL进行人脸表情识别。基于CK+、MMI和FER2013数据集对所提方法进行实验论证,实验结果表明,其识别率在3个数据集上的识别率分别是95.2%、84.1%和73.6%,执行时间为0.082 s,均优于其它对比方法。

To achieve accurate and fast expression recognition,an adaptive reweighted pooling deep multi task learning(DMTL)expression recognition was proposed.A twin neural network was designed to dynamically adjust the scaling probability parameters through the adaptive reweighting module to obtain the category label information with different confidence levels.The adaptive pooling method was improved to select the appropriate parameters according to the changes of samples and training process to improve the flexibility of feature extraction.The improved discriminant DMTL was used to recognize facial expressions by combining the information of category labels and local spatial distribution of samples.Based on CK+,MMI and FER2013 data sets,experimental results show that the recognition rate of the proposed method is 95.2%,84.1%and 73.6%on the three data sets,and the execution time is 0.082 s,which are better than that of other comparison methods.