基于门控剪枝的轻量级表情识别方法  

Lightweight Facial Expression Recognition Method Using Gated Pruni

作　　者：孙培源 袁甲 孙玉宝 SUN Pei-yuan;YUAN Jia;SUN Yu-bao(School of Computer Science,Nanjing University of Information Science&Technology,Nanjing 210044,China;Zhongke Nanjing Intelligent Technology Research Institute,Nanjing 211135,China)

机构地区：[1]南京信息工程大学计算机学院,江苏南京210044 [2]中科南京智能技术研究院,江苏南京211135

出　　处：《中国电子科学研究院学报》2025年第1期75-82,共8页Journal of China Academy of Electronics and Information Technology

摘　　要：针对当前主流表情识别方法虽然具备较高精度但难以部署到边缘设备上的问题,文中提出了一种基于门控剪枝的轻量级表情识别模型的构建方法。首先,选用几种主流的轻量级模型进行人脸识别预训练并从中筛选出精度最高的模型进行迁移学习,以简化后续训练流程和提高模型精度;其次,通过基于门控的全局滤波器剪枝算法对迁移学习后的模型进行剪枝压缩,降低模型的计算复杂度和内存占用,为此,文中提出了一种Punish-Reward-Judge三阶段的迭代剪枝框架用以在剪枝过程中逐步恢复模型精度;最后,对模型进行微调以进一步地提高模型精度。模型的评估是基于当前主流的表情识别数据集AffectNet和RAF-DB上进行的。实验结果表明,在MobileNetV2模型内存压缩了23%的情况下,模型在AffectNet数据集上实现了63.92%的分类精度,超越了很多大体量模型。Addressing the issue that current mainstream facial expression recognition methods,although achieving high accuracy,are difficult to deploy on edge devices,this paper proposes a lightweight facial expression recognition model based on gated pruning.First,several lightweight models were selected for face recognition pre-training,and the model with the highest accuracy was chosen for transfer learning to simplify the subsequent training process and improve model accuracy.Then,a gated global filter pruning algorithm was employed to compress the model after transfer learning,reducing computational complexity and memory footprint.To achieve this,we propose a Punish-Reward-Judge three-stage iterative pruning framework to gradually recover model accuracy during the pruning process.Finally,the model was finetuned to further enhance recognition accuracy.The model was evaluated on the mainstream facial expression recognition datasets AffectNet and RAF-DB.Experimental results show that,with a 23%reduction in memory usage for the MobileNetV2 model,a classification accuracy of 63.92%was achieved on the AffectNet dataset,outperforming many larger models.

关 键 词：面部表情识别 模型轻量化 AffectNet 迁移学习 门控剪枝 

分 类 号：TP391.2[自动化与计算机技术—计算机应用技术]