A Pre-Training Strategy for Convolutional Neural Network Applied to Chinese Digital Gesture Recognition

In this paper, we present an approach to classify Chinese digital gesture based on convolutional neural network (CNN). Principal Component Analysis (PCA) is employed to learn convolution kernels as the pre-training strategy. The learned convolution kernels are used for extracting features instead of the random convolution kernels. The convolutional layers can be directly implemented without any further training, such as Back Propagation (BP). For better understanding, we name the proposed architecture for PCA-based Convolutional Neural Network (PCNN). The dataset is divided into six gesture classes including 14500 gesture images, with 12000 images for training and 2500 images for testing. We examine the robustness of the PCNN against noises and distortions. In addition, the MNIST database of handwritten digits is employed to assess the suitability of the PCNN. Different from the CNN, the PCNN reduces the high computational cost of convolution kernels training. About one-fifth of the training time is shortened. The result shows that our approach classifies six gesture classes with 99.92% accuracy. Multiple experiments manifest the PCNN serving as an efficient approach for image processing and object recognition.