基于改进轻量级秩扩展网络的人脸表情识别方法

人脸表情识别作为人机交互的一种重要方法，广泛应用于智能医疗、公安测谎系统、车载安全系统等领域。现有人脸表情识别方法多数存在参数量冗余、计算成本高、特征表达瓶颈等问题。提出一种基于改进轻量级秩扩展网络ReXNet的人脸表情识别方法。通过构建改进的ReXNet以提取人脸表情特征，在参数量较少的条件下解决特征表达瓶颈的问题，增强对表情局部特征的关注，获得高层次的表情特征，同时融合坐标注意力模块，将位置信息嵌入到通道注意力中，精准地定位和识别感兴趣的特征，建立位置信息与局部特征之间的长依赖关系，减少计算开销。在此基础上，将细化模块引入到改进的网络架构中，利用类别上下文信息细化分类结果，增强类间的分化效果，从而提高人脸表情识别的准确率。实验结果表明，该方法在RAF-DB和FERPlus数据集上的人脸表情识别准确率分别达到88.43%和88.8%，相比VGG16-PLD、SHCNN、ResNet+VGG等方法，具有较高的准确率和较优的鲁棒性。

Facial Expression Recognition Method Based on Improved Lightweight Rank Expansion Network

Facial Expression Recognition(FER), an important method in the field of human-computer interaction, is widely used for intelligent medical treatment and in public security lie detection systems, vehicle safety systems, and other applications.However, existing FER methods are problematic in that they are plagued by parameter redundancy, high computational cost, feature expression bottlenecks, and so on.This study proposes an FER method based on an improved lightweight Rank Expansion Network(ReXNet).By constructing improved ReXNet to exact facial expression features, solves the feature expression bottleneck by decreasing the number of parameters and enhancing attention to local features of expression to produce high-level features of expression.At the same time, a Coordinate Attention(CA) module is integrated, and the location information is embedded into the channel attention.These improvements enable the model to accurately locate and identify the features of interest, and to establish a long-term relationship between location information and local features, thereby reducing the computational overhead.On this basis, a refinement module is introduced to the improved model.The module refines the classification results by using the category context information to enhance the differentiation between classes to ultimately improve the accuracy of FER.The experimental results show that the accuracy of FER on the RAF-DB and FERPlus datasets reaches 88.43% and 88.8%, respectively.Compared with VGG16-PLD, SHCNN, ResNet+VGG, and other methods, the proposed method has higher accuracy and superior robustness.