基于人脸关键特征提取的表情识别

自然场景下人脸表情由于受遮挡、光照等因素影响，以及表情局部变化细微，导致现有人脸表情识别方法准确率较低。提出一种人脸表情识别的新方法，以ResNet18为主干网络，利用残差连接模块加深网络结构，以提取更多深层次的表情特征。通过引入裁剪掩码模块，在训练集图像上的某个区域进行掩码，向训练模型中增加遮挡等非线性因素，提升模型在遮挡情形下的鲁棒性。分别从特征图的通道和空间两个维度提取表情的关键特征，并分配更多的权重给表情变化明显的特征图，同时抑制非表情特征。在特征图输出前加入Dropout正则化策略，通过在训练中随机失活部分神经元，达到集成多个网络模型的训练效果，提升模型泛化能力。实验结果表明，与L2-SVMs、IcRL、DLP-CNN等方法相比，该方法有效提高了表情识别准确率，在2个公开表情数据集Fer2013和RAF-DB上的识别准确率分别为74.366%和86.115%。

Expression Recognition Based on the Extraction of Key Facial Features

The accuracy of existing facial expression recognition methods is typically low owing to the influence of occlusion, illumination, and other factors and to subtle local variations in facial expressions in natural scenes.This study presents a new method for facial expression recognition.A ResNet18 model is adopted as the backbone network, and a residual connection module is employed for a deeper network structure to extract deeper expression features.First, a cutout module is introduced.By masking a certain area of each image in the training set, the model learns to consider several nonlinear factors, such as occlusion, thus improving its robustness to these conditions.The key features of expressions are extracted from the channel and space of the graph, and more weights are assigned to feature maps with obvious expression changes to suppress non-expression features.Finally, prior to the output of the feature map, a Dropout regularization strategy is implemented to randomly deactivate some neurons and integrate multiple network models to improve the generalization ability of the model.The experimental results show that the proposed method exhibits improved accuracy in expression recognition tasks compared with L2-SVMs, IcRL, DLP-CNN, and other methods. Recognition accuracy values of 74.366% and 86.115% are achieved on two public expression datasets, Fer2013 and RAF-DB, respectively.