Gesture recognition based on transfer learning
Xue Fei Wu, Xiaoru Song, Song Gao, Chaobo Chen · 2019
At present, the general methods of gesture recognition are machine learning method and deep learning method. However, machine learning method has the disadvantage of extracting image features manually, while deep learning method requires high hardware requirements and is easy to over-fitting in small samples. To solve above problems, this paper proposes gesture recognition based on transfer learning. Firstly, transfer the pre-trained model and convolutional layer parameters on the Image-Net dataset to the gesture dataset with small samples. And then frozen the convolution layer as the feature extractor and fine-tune the output layer of the network to fit the target data. Finally, initialize the 3 fully connection layer of the network with improved normal distribution initialization method and retrain the net again. The simulation results show that the proposed method has a good recognition rate in the small sample gesture image dataset and avoid the over-fitting, which has a great significance for gesture recognition.