The method uses gesture image as input

passing through four convolutional layers

followed by one pooling layer for every two convolutional layers. The first convolution layer uses 64 3-channel convolution kernels of size 3 × 3 to extract features from 128 × 128 original gesture images

and the second convolution layer uses 32 convolution kernels of size 3 × 3 for feature extraction. The 2 × 2 pooling window is used for feature reduction. The third and fourth convolution layers use 32 convolution kernels of size 3 × 3 and 16 convolution kernels of size 3 ×3 to extract features respectively. Then

use the 2 × 2 pooling window to reduce the dimension. After the network is fully connected

some nodes and their associated input and output connections are removed by using the Dropout technology

then sent to the Soft Max layer for classification. The experimental results show that the accuracy of gesture recognition can reach about 97 %

which verifies the validity of the proposed model.