双流增强融合网络微表情识别

为解决微表情识别领域数据集样本数量少，样本类型分布不均导致识别率鲁棒性差的问题，提出了一种基于双流增强网络的微表情识别模型。该模型基于单帧RGB图像流及光流图像流的双流卷积神经网络，以权威数据集为基础，数据增强为基准，构建微表情识别模型。通过在SoftMax逻辑回归层融合单帧空域信息和光流时域信息，对两个独立流的网络性能进行提升，并通过引入基于带循环约束的生成对抗网络的图像生成方式对数据集进行扩充。通过将输入微表情视频帧序列进行分解，将其分割为双流网络的灰度单帧序列与光流单帧序列，对两类序列图进行数据增强，再进行微表情识别模型构建的方法，有效提高了微表情识别率。基于双流增强网络的微表情识别模型可以较好提升微表情识别准确度，鲁棒性较好，泛化状态较稳定。

Micro-expression recognition based on a dual-stream enhanced fusion network

We propose a micro-expression recognition model based on a dual-stream enhanced network in order to address the issues of insufficient samples from the dataset of micro-expression and uneven distribution of sample types leading to a low rate of robustness. Targeting a dual-stream convolutional neural network of single-frame RGB image flow and optical image flow, a micro-expression recognition model is built based on a fundamental authoritative dataset and data enhancement. Single-frame airspace information and optical time flow domain information are incorporated in the SoftMax logistic regression layer to improve the network performance for two independent streams. The dataset is augmented by introducing a method for image generation based on a generative adversarial network with loop constraints. After segmenting the input micro-expression video frame sequence into greyscale single frame sequences and optical flow single frame sequences of a dual-stream sequence diagram, augmenting the data of the two sequences, and constructing the micro-expression recognition model, the input micro-expression video frame sequence is subdivided. The rate of micro-expression recognition has been significantly enhanced by this method. The micro-expression recognition model based on dual-stream enhanced networks can effectively improve the recognition accuracy of micro-expressions with improved robustness and generalization state.