Fully Funded PhD: Vision Language Models for Micro-Expression Analysis

Micro-expressions are tiny, involuntary facial movements that can occur when a person tries to hide an emotion, but are difficult to analyse because they are subtle, short-lived, and easily confused with ordinary facial movement. While many methods have been developed to recognise and localise micro-expressions, these results remain difficult to understand by people and lack interpretable evidence.

We are looking for candidates to research computer vision, machine learning, and modern vision-language models to create new methods of analysing micro-expressions. Rather than only predicting a label, the proposed models will be designed to identify where the relevant facial evidence occurs, describe what subtle change has been observed, and explain why a prediction has been made.

The project will deliver new methods, benchmark protocols, and evaluation frameworks for explainable and trustworthy micro-expression analysis, with relevance to human-centred AI, and healthcare-related affect analysis. The core novelty of the project is to shift micro-expression research from predominantly vision-only pipelines towards multimodal foundation-model approaches that combine subtle visual motion with structured language-level representations such as action units, temporal descriptions, prompt templates, and evidence-based explanations.

The successful candidate will benefit from training facilities at Manchester Met, and the potential to work with external industry stakeholders.

Aim: We will investigate whether vision-language models (VLMs) can improve micro-expression analysis by combining video understanding with language-based reasoning.

• Process micro-facial expression data more efficiently in computer vision and vision language models.
• Create a language-guided representation for subtle facial motion, exploring multimodal data
• Create a spot-then-recognise approach to detect micro-expression intervals and interpret language-aware recognition.
• Generalise and efficiently adapt across different datasets and in-the-wild data.
• Evaluate novel methods and validate against the state-of-the-art in a real-world setting.