# Deep Learning based Micro-expression Recognition: A Survey

**Authors:** Yante Li, Jinsheng Wei, Seyednavid Mohammadifoumani, Yang Liu, Guoying Zhao  
**arXiv:** [2107.02823](https://arxiv.org/abs/2107.02823)  
**Date:** 2021-07-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.02823v1/

## Key points

- Micro-expressions (MEs) are involuntary facial movements that reveal hidden emotions in high-stake situations
- Deep learning (DL) techniques have been applied to micro-expression recognition (MER)
- Early methods for MER relied on traditional appearance and geometry features
- DL approaches have been proposed to improve MER performance and overcome limitations of small-scale datasets
- The paper provides a comprehensive review of deep MER, including datasets, pipeline, benchmarking methods, and a new taxonomy based on DL
- Remaining challenges in deep MER are highlighted and potential directions for designing robust systems are proposed
- The paper aims to bridge the gap between traditional methods and DL-based approaches in MER research

## Summary

This paper presents a comprehensive survey on deep micro-expression recognition (MER), focusing on the application of deep learning (DL) techniques. Micro-expressions (MEs) are involuntary facial movements that reveal hidden emotions in high-stake situations, making them important in various fields such as medical treatment, national security, interrogations, and human-computer interaction systems. Early methods for MER primarily relied on traditional appearance and geometry features. However, with the success of DL in various domains, there has been increasing interest in using neural networks for MER. Unlike macro-expressions, MEs are spontaneous, subtle, and rapid facial movements which pose challenges for data collection due to their difficult-to-capture nature. To address these challenges, various DL approaches have been proposed to improve MER performance and overcome the limitations of small-scale datasets. In this survey, the authors provide a comprehensive review of deep MER by discussing datasets used in research, the deep MER pipeline and benchmarking influential methods. They introduce a new taxonomy for the field that encompasses all aspects of MER based on DL. For each aspect they summarize and discuss both basic approaches and advanced developments. Additionally they highlight remaining challenges in deep MER and propose potential directions for designing robust systems. This survey serves as a reference point for future research in deep MER and aims to bridge the gap between traditional methods and DL-based approaches. Overall this paper is the first systematic and detailed survey specifically focusing on DL-based MER. It evaluates dynamic network inputs blocks structures losses and special tasks in detail while identifying further study directions with the goal of providing researchers with a comprehensive understanding of current advancements in DL-based MER techniques and inspiring future innovations in this field.