Combining temporal interpolation and DCNN for faster recognition of micro-expressions in video sequences

• Published in: 2016 International Conference on Advances in Computing, Communications and Informatics (ICACCI) pp. 699 - 703
• Main Authors: Mayya, Veena, Pai, Radhika M., Manohara Pai, M.M.
• Format: Conference Proceeding
• Language: English; Japanese
• Published: IEEE 01.09.2016
• Subjects: CASME II; Computational modeling; Computer Vision; Confusion Matrix; Facial features; Feature extraction; Graphics processing units; Informatics; Interpolation; Machine Learning; Micro-expressions recognition; SMIC; Video sequences
• DOI: 10.1109/ICACCI.2016.7732128

Micro-expressions are the hidden human emotions that are short lived and are very hard to detect them in real time conversations. Micro-expressions recognition has proven to be an important behavior source for lie detection during crime interrogation. SMIC and CASME II are the two widely used, spontaneous micro-expressions datasets which are available publicly with baseline results that uses LBP-TOP for feature extraction. Estimation of correct parameters is the key factor for feature extraction using LBP-TOP, which results in long computation time. In this paper, the video sequences are interpolated using temporal interpolation(TIM) and then the facial features are extracted using deep convolutional neural network(DCNN) on CUDA enabled General Purpose Graphics Processing Unit(GPGPU) system. Results show that the proposed combination of DCNN and TIM can achieve better performance than the results published in baseline publications. The feature extraction time is reduced due to the usage of GPU enabled systems.