A convolutional neural network approach for semaphore flag signaling recognition

This paper proposes a recognition approach for Semaphore flag signaling (SFS). We use the improved convolutional neural network (CNN) to classify the SFS. In the experiment we made Semaphore flag signaling system (SFSS), which based on CNN. The image can be directly input into the SFSS. Each alphabetic character or control signal is indicated by a particular flag pattern. We shoot the SFS videos by a monocular camera. The dataset is divided into five SFS classes. The improved CNN uses the Relu activation function, the max-pooling methods. It's alway use SFS data whitening and grayscale preprocessing methods. The improved CNN provides for partial invariance to different light, angles, scenes, and a group of people. The result shows that our approach classifies five SFS classes with 99.95% accuracy.