Abstract
Expression recognition in the wild is a challenging task because of the interference of various environment. In this paper, we propose a transfer-learning method that utilize two representative transformations from grayscale images as input and fuse their results in decision level to enhance the overall performance, which also address the dimensional mismatch issue when applying pre-trained deep neural network on grayscale images. Each network consists of the body of pre-trained deep network which is freezed during training process, and two other layers that trained for our expression classification task. Our method achieve an accuracy of 63.72%, a 11.18% improvement over baseline results on FER-2013 with fewer computation compared with other deep-learning methods, which demonstrate the complementary effect between two different inputs, as well as the good generalization ability of the features extracted by pre-trained network.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call