Convolutional Neural Network Architecture for Semaphore Recognition

This paper proposes convolutional neural networks SRNet and Tiny-SRNet for human semaphore action recognition. SRNet is composed of 5 layers of convolution and 3 layers of fully connected layers. In addition to the first convolution layer., a batch normalization layer is added before each convolution layer. In order to enable deep learning algorithms to be applied to both mobile and embedded platforms., Tiny-SRNet removes the full connected layers in SRNet and replaces them with a convolutional layer and a global average pooling layer. The experimental results show that compared with the mainstream classification models AlexNet., GoogleNet and VGGI6., SRNet achieves the highest recognition rate of 98.9% on the semaphore dataset, and Tiny-SRNet compresses its model size to 1/24 of SRNet with a reduction of 1.7% accuracy.