Rain Detection in Television Images Using Artificial Neural Network

An algorithm for rain detection based on television imagery implemented using an artificial neural network is presented. An overview of existing architectures of artificial neural networks commonly employed for processing TV images, including both fully connected and convolutional neural networks, is provided. The choice of the pre-trained convolutional neural network VGG-16 as the basis for the developed algorithm is justified. A modification of the VGG-16 network structure aimed at adapting it to the task of rain detection has been proposed, which involves replacing the fully connected part of the network with one fully connected layer containing only one artificial neuron. The resulting artificial neural network was trained using a synthetic image database. The accuracy of rain detection by the proposed neural network according to the Accuracy metric on the test sample reached 90.5%.