Abstract
Ship detection in satellite images is a challenging task. In this paper, we introduce a transfer learned Single Shot MultiBox Detector (SSD) for ship detection. To this end, a state-of-the-art object detection model pre-trained from a large number of natural images was transfer learned for ship detection with limited labeled satellite images. To the best of our knowledge, this could be one of the first studies which introduce SSD into ship detection on satellite images. Experiments demonstrated that our method could achieve 87.9% AP at 47 FPS using NVIDIA TITAN X. In comparison with Faster R-CNN, 6.7% AP improvement could be achieved. Effects of the observation resolution has also been studied with the changing input sizes among 300 × 300, 600 × 600 and 900 × 900. It has been noted that the detection accuracy declined sharply with the decreasing resolution that is mainly caused by the missing small ships.
Highlights
Ship detection from satellite imagery plays an important role in the maritime surveillance, e.g. traffic monitoring, fishing management, oil pollution control etc
There is a fact that the convolutional neural networks provide means to learn rich low and middle level features transferrable to a variety of visual recognition tasks [19]
This paper introduces a transfer learned Single Shot MultiBox Detector (SSD) that transfers visual knowledge between natural image and satellite image
Summary
Ship detection from satellite imagery plays an important role in the maritime surveillance, e.g. traffic monitoring, fishing management, oil pollution control etc. According to the feature extraction method, the previous work on ship detection can be classified into two groups, handcraft feature and machine learning feature. Researchers proposed some adaptive feature extraction methods according to the expert knowledge about the ship in satellite imagery. Due to the variety of ship size, the handcraft feature approaches may lead to poor performance for ship detection. Machine learning feature makes a great profit for object detection These approaches take advantage of some prior knowledge, automatically learn image features, and find potential object characters and the distribution rules over the object which cannot be described by human cognition. Current state-of-theart object detection systems using deep learning are variants of the following approach: hypothesize bounding boxes, resample pixels or features for each box, and apply a high quality classifier. Figure 4. statistic of the ship length in satellite images
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
