Convolutional Neural Network-Based Novel Framework for Ship Recognition in Poor Visibility Maritime Surveillance Videos

Ship recognition is a central component of the information perception for smart ships. Traditionally, the ship recognition mission is completed by remote sensing or synthetic aperture radar (SAR) or manual calibration via maritime surveillance video. However, the difficulty arises in implementing the mission when it suffers the weather with a poor visibility. The recognition accuracy may deteriorate dramatically in such a hostile environment, especially in the heavy fog environment at sea when using conventional methods. In this case, we develop a novel framework for ship recognition based on a smart recognition system that relies on the dark channel defogging and the neural networks. First, a dark-channel defogging algorithm is utilized to remove the adverse impact of poor visibility on recognition. After that, three different neural networks are used to train the defogged images (Specifically: Faster-RCNN, SSD, YOLOv3). Finally, we present qualitative and quantitative analyses of three different neural network models. The results demonstrate the effectiveness of the proposed method in a hostile environment, especially in fog weather with poor visibility at sea.