AutoMER: Spatiotemporal Neural Architecture Search for Microexpression Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial microexpressions reveal genuine emotions but are challenging to model due to subtle temporal changes.
• Accurate microexpression recognition (MER) is crucial for understanding human emotions.

Purpose of the Study:

• To propose AutoMER, a novel spatiotemporal architecture search algorithm for microexpression recognition.
• To develop efficient 3D convolutional neural network (CNN) architectures for MER.

Main Methods:

• Introduced a parallelogram design-based search space for efficient architecture search.
• Developed a 3-D singleton convolution module for spatiotemporal feature extraction.
• Utilized four candidate operators and two 3-D dilated convolution operators for end-to-end video encoding.
• Conducted network-level search for 3D CNN architectures, a first for MER.

Main Results:

• Searched models evaluated on five benchmark datasets (CASME-I, SMIC, CASME-II, CAS(ME) ∧2, SAMM).
• Generated models quantitatively outperformed existing state-of-the-art approaches in microexpression recognition.
• Validated AutoMER with various configurations, including downsampling, multiscale convolutions, and kernels.

Conclusions:

• AutoMER represents a significant advancement in automated microexpression recognition.
• The proposed approach enables the discovery of highly effective 3D CNN architectures for MER.
• Ablation studies provided insights into the operational effectiveness of AutoMER.