Heat-Map Based Emotion and Face Recognition from Thermal Images

Nowadays emotion recognition becomes feasible in the Computer Vision domain with the help of Convolutional Neural Networks. However, the credibility of emotion recognition from daily images or videos is evidently insufficient. As people can easily mimic emotions one after another by fooling the computational models, different defensive approaches should be taken into consideration. Particularly, thermal images taken by thermal cameras visualize the facial and body's heat status, revealing where humans actually feel emotions; therefore, models trained with thermal heat-maps are less subject to fake expressions. Accordingly, heat-maps provide suitable resources for developing more credible emotion recognition models. In this paper, a fast detection algorithm, YOLO, is adapted and trained to detect emotions in thermal images. The detection performance, in terms of average precision and intersection over union, from three detection algorithms, YOLO, ResNet, and DenseNet, is compared and their respective characteristics are discussed.