基于时空差异网络的抑郁检测方法研究  

Research on Depression Detection Based on Spatiotemporal Difference Networks

作　　者：吴东升 徐鹏飞 林玉婷 WU Dongsheng;XU Pengfei;LIN Yuting(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,沈阳110159

出　　处：《沈阳理工大学学报》2025年第3期24-31,共8页Journal of Shenyang Ligong University

基　　金：辽宁省教育厅高等学校重点攻关项目(JYTZD2023006)。

摘　　要：在面部表情识别领域,针对不同抑郁水平的面部表情差异小以及模型计算复杂度高导致过度拟合的问题,提出一种由对角注意力机制(DMA)和时空差异模块(SDN)相结合的深度学习架构(DSN)进行抑郁检测,并根据面部表情的变化进行抑郁症等级评估。引入DMA使模型具备精确识别图像之间变化的能力;使用SDN模块捕获平滑与突然的面部表情变化,探索多尺度时间信息。实验结果表明,本文提出的DSN架构在保证模型性能的同时相对于传统架构降低了42%的运算复杂度,提高了模型整体容错水平,具有较好的泛化能力。In the field of facial expression recognition,to address the challenges of minimal facial expression differences at different depression levels and high computational complexity leading to overfitting,a deep learning architecture(DSN)is proposed for depression detection,which combins a diagonal attention mechanism(DMA)and a spatiotemporal differences network(SDN)to assess depression levels based on facial expression variations.The introduction of DMA enables the model to accurately identify changes between images.Meanwhile,the SDN module captures smooth and abrupt facial expression changes,exploring multiscale temporal information.Experimental results indicate that the proposed DSN architecture not only ensures model performance but also reduces computational complexity by 42%compared to traditional architectures.This enhancement improves the overall fault tolerance level of the model,demonstrating its superior generalization capabilities,

关 键 词：面部表情识别 抑郁检测 时空差异模块 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]