Recognizing Hand Gesture in Still Infrared Images by CapsNet

In recent years, gesture recognition has been becoming a hot topic that attracts much attention from the computer vision community because of its great potential in many real-world applications. There is a need to design a robust hand-based gesture recognition algorithm to cope with hand gesture recognition tasks. Infrared image recognition has the characteristic of not being disturbed by illumination variation. As a promising alternative to Convolutional Neural Networks (CNN), Capsule Networks (CapsNet) can represent the orientations of features and capture the spatial relationships between features of an entity, which makes CapsNet possess higher generalization ability. In this paper, we propose IRHGR-CapsNet to investigate hand gesture recognition in still infrared images. To evaluate the testing accuracy, the convergence ability and the generalization ability of IRHGR-CapsNet comprehensively, we split the original dataset into three subsets according to different split proportions and get four different dataset split modes for experiments. We can achieve almost 100.00% testing accuracy on all the four dataset split modes. Meanwhile, we can also demonstrate that our proposed IRHGR-CapsNet has a strong convergence ability and generalization ability.