【作者】 陈进；

【导师】 胡敏； 王钊；

【作者基本信息】 合肥工业大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 抑郁症作为一种常见的精神障碍,受到广泛关注。目前,自动抑郁症检测(Au-tomatic Depression Detection,ADD)因其高效、客观、非侵入以及可部署性强的优势,逐渐成为心理健康研究领域的热点问题。其中,基于面部图像的自动抑郁症检测方法因面部图像蕴含丰富情绪线索,能有效反映个体心理状态与潜在抑郁倾向,而备受青睐。然而,现有方法在特征提取精细度、局部信息建模能力及特征交互融合方面存在不足,难以充分挖掘面部图像中与抑郁症相关的关键特征。为解决这些问题,本文深入研究基于多层次特征交互的抑郁症检测模型,主要研究工作如下:(1)为了充分挖掘面部图像中的抑郁情感信息,本文提出了一个双分支表征增强网络(Dual-branch Representation Enhancement Network,DBRE-Net)用于ADD。该网络通过多层次特征提取与语义增强,有效融合全局与局部面部特征。首先,为模拟人类视觉感知中对整体与细节的协同处理机制,设计了双分支结构,分别从全局和局部层面对面部图像进行特征提取。其中,全局分支通过特征复用机制,整合不同感受野下的全局特征;局部分支则对特征图进行区域划分,并结合精心设计的局部特征提取模块,挖掘与抑郁症相关的关键细节特征。然后,特征增强模块通过建模通道间的依赖关系,提升特征的表达能力;而特征关系学习模块则进一步挖掘不同特征之间的互补性与关联性,增强模型对抑郁症潜在表征的建模能力。在AVEC2014数据集上的大量实验结果表明:与现有方法相比,DBRE-Net取得了有竞争力的效果。(2)部分抑郁症患者的面部表情等非语言信号与正常人差别不大,尤其在轻度和中度患者中,这种差异更为细微。因此,如何从细节中提取与抑郁症相关的信息,是提升自动检测准确性与鲁棒性的关键。为进一步优化DBRE-Net,本文提出一种表情感知增强与频域特征融合的抑郁症检测网络(Expression-Sensitive and Frequency Fusion Network for Automatic Depression Detection,ESFF-Net)。具体地,引入细节增强卷积模块,通过集成多种差分卷积,挖掘图像中的细微情感变化,从而提升对轻中度抑郁症的识别能力。此外,设计频域卷积融合器与全局先验引导模块,前者在空间域与频域中进行双阶段融合,增强局部特征间的关联性,后者则利用全局信息引导局部特征学习,帮助模型更准确地理解抑郁症的整体表现。在AVEC2014数据集上进行了大量实验,实验结果表明了ESFF-Net中所提模块的有效性。更多还原

【Abstract】 Depression,as a common mental disorder,has attracted widespread attention.Cur-rently,Automatic Depression Detection(ADD)has emerged as a hot topic in the field of mental health research due to its advantages of high efficiency,objectivity,non-invasiveness,and strong deployability.Among them,facial image-based ADD meth-ods have garnered significant interest,as facial images carry rich emotional cues that effectively reflect an individual’s psychological state and potential depressive tendencies.However,existing approaches suffer from limitations in the granularity of feature extrac-tion,the modeling of local information,and the interaction and fusion of features,which hinder their ability to fully capture depression-related key features from facial images.To tackle these challenges,this dissertation conducts an in-depth investigation into a depres-sion detection model based on multi-level feature interaction.The main contributions are as follows:(1)To fully exploit depression-related emotional information embedded in facial im-ages,this dissertation proposes a Dual-branch Representation Enhancement Network(DBRE-Net)for automatic depression detection.The proposed method integrates global and local facial features through multi-level feature extraction and semantic enhancement.Inspired by the human visual perception mechanism that simultaneously processes holis-tic and detailed information,a dual-branch architecture is designed to extract features from both global and local perspectives.Specifically,the global branch employs a fea-ture reuse mechanism to integrate global features under different receptive fields,while the local branch partitions the feature maps into regions and leverages a carefully designed local feature extraction module to capture key depression-related details.Subsequently,a feature enhancement module models inter-channel dependencies to improve the expres-siveness of the features,and a feature relationship learning module further explores the complementarity and correlations among different features,enhancing the model’s ca-pability to represent potential depressive cues.Extensive experiments conducted on the AVEC2014 dataset demonstrate that DBRE-Net achieves competitive performance com-pared to existing state-of-the-art methods.(2)The non-verbal cues of some individuals with depression,such as facial expres-sions,show minimal differences from those of healthy individuals—particularly in mild and moderate cases,where such distinctions are often subtle and easily overlooked.Ef-fectively capturing depression-related information from these fine-grained facial features is thus essential for improving the accuracy and robustness of automatic depression detec-tion.To further enhance DBRE-Net,this dissertation proposes the Expression-Sensitive and Frequency Fusion Network for Automatic Depression Detection(ESFF-Net).Specif-ically,a detail-enhanced convolution module is introduced to extract subtle variations in facial images by integrating multiple types of differential convolutions,thereby improving the recognition of mild and moderate depression.In addition,a frequency-domain convo-lution fusion module and a global prior guidance module are designed to fuse key features across facial regions.The former performs two-stage fusion in both spatial and frequency domains to strengthen local feature correlations,while the latter leverages global informa-tion to guide local feature learning,enabling a more accurate understanding of depressive manifestations.Experimental results on the AVEC dataset confirm the effectiveness of the proposed modules.更多还原