AutoGesNet: Auto Gesture Recognition Network Based on Neural Architecture Search

The deep-learning-based gesture recognition technologies have developed rapidly in recent years, and more and more convolutional neural network models are proposed. In this paper, we propose a method to automatically generate a convolutional neural network for the gesture recognition task and name the network AutoGesNet, to solve the problem that it is difficult to design a good neural network architecture. To be specific, we firstly fuse and preprocess three gesture recognition data sets. Then we design the overall architecture and the search space of AutoGesNet. And we use reinforcement learning and transfer learning method to automatically generate the detailed architecture of AutoGesNet. Finally, we fine-tune and retrain the searched neural network for two different input sizes. Experiments show that the retrained model achieves above 99% accuracy on NUS Hand Posture Dataset II and our data set, and its parameters and FLOPs are reduced by more than 40% compared with lightweight MobileNet.