摘要
针对面部不连续动作单元的关联特征提取困难,以及不同面部区域对表情识别影响程度不一可能引入无用信息的问题,提出了一种基于双分支注意力机制的多尺度自适应空洞卷积模型(dual branching attention mechanism-adaptive multi-scale dilated convolution,DAM-ADCNN)。模型通过双分支注意力机制生成特征映射,表征面部动作单元的局部和全局分布及关联关系;利用多尺度空洞卷积提取面部不连续动作单元的关键特征;采用自适应方式动态调整不同尺度关联特征的权重,以有效减少无用信息的干扰。结果表明,DAM-ADCNN模型在情感识别任务中的表现优于现有方法。在DEAP数据集的唤醒和效价维度上,模型的识别准确率分别提升了3.66%和3.99%。同时,在CK+数据集上,模型的识别准确率提高了3.93%。这些结果证明了DAM-ADCNN模型在面部表情情感识别中的有效性。
To overcome the difficulty of associated feature extraction for facial discontinuous action units and to address the problem that different facial regions may introduce useless information with different degrees of influence on emotion recognition,this paper proposes a multi-scale adaptive dilated convolution model based on a Dual Branching Attention Mechanism-Adaptive Multi-Scale Dilated Convolution(DAM-ADCNN).The model first generates feature mappings through a two-branch attention mechanism to characterize the local and global distributions and association relationships of facial action units.Then,the key features of the facial discontinuous action units are extracted using multi-scale cavity convolution.Finally,an adaptive approach is employed to dynamically adjust the weights of the associated features at different scales to effectively reduce the interference of useless information.Experimental results on the DEAP and CK+datasets show the DAM-ADCNN model outperforms existing methods in the emotion recognitions.On the arousal and validity dimensions of the DEAP dataset,the model improves the recognition accuracy by 3.66%and 3.99%respectively.On the CK+dataset,it enhances the recognition accuracy by 3.93%.These results demonstrate the effectiveness of the DAM-ADCNN model in emotion recognition through facial expressions.
作者
王春影
孟天宇
张震
葛雄心
杨继伟
WANG Chunying;MENG Tianyu;ZHANG Zhen;GE Xiongxin;YANG Jiwei(School of Applied Technology,Changchun University of Technology,Changchun 130012,China;School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)
出处
《重庆理工大学学报(自然科学)》
北大核心
2025年第5期90-97,共8页
Journal of Chongqing University of Technology:Natural Science
基金
吉林省科技发展计划项目(20190302114GX)。
关键词
面部情感识别
双分支注意力机制
空洞卷积
自适应权重
facial emotion recognition
two-branch attention mechanism
dilated convolution
adaptive weighting
作者简介
王春影,女,硕士,正高级实验师,主要从事人脸识别、虚拟现实研究,E-mail:405992879@qq.com。
