A Switch State Recognition Method based on Improved VGG19 network

Intelligent recognition of switch state based on computer vision is one of the important research contents of power intelligence. For solving the problem of low recognition rate under poor imaging environment, this paper presents an improved VGG19 network. There are 3 improved aspects: 1) in the first convolutional layer, the convolution features are extracted from the RGB image and the LBP image of the switch, and then the two convolution features are averaged as the output result of the first convolution layer, in order to enhance the adaptability of features to illumination changes and the ability to extract detailed features; 2) in the subsequent 15 convolution layers, the convolution feature output of each layer is averaged with the original input, and then convolved again, so as to extract detail features by secondary convolution; 3) we selects CReLU instead of ReLU as activation function for reducing the redundancy of the convolution kernel. Experiments show that the new method can greatly improve the performance of switch state recognition.