Real-Time Video Processing for Ship Detection Using Transfer Learning

Abstract

AbstractAutomatic ship classification and detection is an interesting research field concerning maritime security. Automatic ship detection systems are important for maritime security and surveillance. These systems can be used to monitor marine traffic, illegal fishing and illegal activities which deal with the prospects of maritime security. This research has gained interest because many ships that are sailing on the ocean or sea do not install transponders which are used for tracking the ships. It will be a serious threat to the nation, mankind, and sea-life if we do not keep an eye on these kinds of ships. Therefore, in this we presented a novel Deep Learning method that will be used to detect ships by using satellite Images. This approach uses TensorFlow object detection API to detect objects in the images as we are concerned with object detection. The dataset used for this purpose is Maritime Satellite Imagery (MASATI)-v2 dataset which consists of various satellite images that are captured under different weather and dynamic conditions. As the real-time satellite monitoring is a kind of video thing, we proposed an approach to perform video processing to detect ships by using the model that is trained using MASATI-v2 dataset. For training the model we are using a Transfer Learning technique by utilizing (Single Shot Detector) SSD MobileNetV2. This algorithm is different from normal convolution neural networks, and it uses depth wise and pointwise separable filters to perform the task. KeywordsShip detectionTransfer learningObject detectionSingle shot detectorMobileNetV2