Research on Dynamic Pruning and Distillation Quantization for Expression Recognition Based on MobileNetV3

DOI: https://doi.org/10.62517/jike.202604218

Author(s)

Peihong Li

Affiliation(s)

School of Information Engineering, Beijing Institute of Petrochemical Technology, Beijing, China *Corresponding Author

Abstract

To address the lightweight requirements for real-time facial expression recognition in mobile and edge devices, this paper proposes a collaborative optimization method based on MobileNetV3 incorporating dynamic pruning, knowledge distillation, and quantization-aware training. A lightweight gate network is designed to enable adaptive dynamic pruning of inputs and automatically allocate computing resources according to expression image complexity. A teacher-student architecture is employed for bidirectional knowledge distillation to compensate for accuracy loss caused by model compression. INT8 quantization-aware training is integrated to further reduce model size and inference latency, while Focal Loss and label smoothing loss are utilized to enhance robustness on imbalanced datasets. Experiments on the FER2013 dataset demonstrate that the proposed method achieves over 50% reduction in parameter and computational costs, 65% model size reduction, and 2.3-fold inference speed improvement with only 0.8% accuracy degradation. Outperforming mainstream lightweight networks in accuracy-efficiency balance, this approach provides an efficient lightweight solution for edge-based facial expression recognition.

Keywords

Dynamic Pruning; Knowledge Distillation; Quantitative Perception Training; MobileNetV3; Expression Recognition; Model Compression

References

[1] Li Beibei;Zhu Jiansheng; Li Suwen; Dai Linlin; Yan Zhiyuan; Ma Liangde;. Real-Time Facial Expression Recognition on Res-MobileNetV3[J]. China Communications, 2025(03):60-70. [2] Yajing Li; Xiaoyan Xiong; Wenbin Xin; Jiahai Huang; Huimin Hao;. MobileNetV3-CenterNet: A Target Recognition Method for Avoiding Missed Detection Effectively Based on a Lightweight Network[J]. Journal of Beijing Institute of Technology, 2023(01):85-97. [3] Lu Haoyang. Research and Application of Facial Expression Recognition Based on Lightweight Convolutional Neural Networks[D]. Xi’an University of Technology, 2025. [4] Shen Hao. Facial expression recognition based on lightweight convolutional networks[D]. Tianjin University, 2021. [5] Wei Xinguang. Research on Facial Expression Recognition Method Based on Convolutional Neural Networks[D]. Shandong University, 2023. [6] Ding Qin. Research on Student Behavior Recognition and Facial Expression Classification Methods for Practical Classroom Applications[D]. Anhui University of Science and Technology, 2025. [7] Zhang X. Research on Remote Sensing Image Classification and Detection Method Based on Binary Neural Networks[D]. Northeast Forestry University, 2023. [8] Wang Dingkun. Lightweight Plant Disease Recognition Model and Transplantation Based on Knowledge Distillation and Channel Pruning[D]. East China Jiaotong University, 2023. [9] Hu Zhou. Optimization and Application of Compression Algorithm for Convolutional Neural Network Models[D]. Hefei University of Technology, 2021. [10] Liu Yuanyuan; Wang Dingkun; Wu Lei; Huang Dechang; Zhu Lu;. Lightweight model-based plant disease recognition using knowledge distillation and model pruning[J]. Zhejiang Agricultural Journal, 2023(09):234-248. [11] Zhang Longhao. Research and Application of Rock Identification Based on Deep Learning[D]. Guangxi Minzu University, 2024. [12] Hu Xiaohui; Liu Yahui; Gong Lipeng; Chi Xiaojie; Sun Ranran; Wang Ziwei;. Research on Vegetation Classification for Tailings Dam Restoration Based on MobileNetV2 Lightweight Network[J]. Journal of Chifeng University (Natural Science Edition), 2025(01):100-105. [13] Wang Tingting; Huang Zhixian; Wang Hongtao; Yang Minghao; Zhao Wanchun;. Rock slice lithology recognition based on MobileNetV2[J]. Journal of Jilin University (Earth Science Edition), 2024(04):354-364. [14] Zhou Hanmi; Chen Jiageng; Dai Zhiguang; Niu Xiaoli; Qin Long; Xiang Youzhen; Zhao Long;. Recognition of apple leaf diseases based on lightweight residual networks[J]. Fujian Agricultural Journal, 2024(01):87-96. [15] Zhao Chenggong; Hu Jiyuan; Yang Xingyu; Chen Lei; Jiang Hongxu;. Post-training quantization method for CNN-Transformer hybrid models on UAV platforms[J/OL]. Journal of Beihang University, 1-10[2026-04-11]. [16] Zhang Fukai. Research on Citrus Pest and Disease Recognition Technology Based on Attention Mechanism[D]. Southwest University, 2024.