Search

Facial micro-expression (ME) recognition requires face movement interval as input, but computer methods in spotting ME are still underperformed. This is due to lacking large-scale long video dataset and ME generation methods are in their infancy. This thesis presents methods to address data deficiency issues and introduces a new method for spotting macro- and micro-expressions simultaneously. This thesis introduces SAMM Long Videos (SAMM-LV), which contains 147 annotated long videos, and develops a baseline method to facilitate ME Grand Challenge 2020. Further, a reference-guided style transfer of StarGANv2 is experimented on SAMM-LV to generate a synthetic dataset, namely SAMM-SYNTH. The quality of SAMM-SYNTH is evaluated by using facial action units detected by OpenFace. Quantitative measurement shows high correlations on two Action Units (AU12 and AU6) of the original and synthetic data. In facial expression spotting, a two-stream 3D-Convolutional Neural Network with temporal oriented frame skips that can spot micro- and macro-expression simultaneously is proposed. This method achieves state-of-the-art performance in SAMM-LV and is competitive in CAS(ME)2, it was used as the baseline result of ME Grand Challenge 2021. The F1-score improves to 0.1036 when trained with composite data consisting of SAMM-LV and SAMMSYNTH. On the unseen ME Grand Challenge 2022 evaluation dataset, it achieves F1-score of 0.1531. Finally, a new sequence generation method to explore the capability of deep learning network is proposed. It generates spontaneous facial expressions by using only two input sequences without any labels. SSIM and NIQE were used for image quality analysis and the generated data achieved 0.87 and 23.14. By visualising the movements using optical flow value and absolute frame differences, this method demonstrates its potential in generating subtle ME. For realism evaluation, the generated videos were rated by using two facial expression recognition networks.