AureNet: A Real-Time Arbitrary-oriented and Ship-based Object Detection
Mingyue Liu, Yingyang Chen, Daorui Ding · 2023
In recent years, due to the bursting computing power, neural-network-based object detection algorithms are emerging in large numbers, which is very promising in monitoring the transportation, especially in remoting sensing of the sea where it is relatively easier to distinguish the background and foreground. But because of the arbitrary angles of the ships, the impacts of illuminations and the too large aspect ratio of the length with respect to width for the ships, ship detection has faced challenging. To improve the accuracy of ship detection, in this paper, we present a new network based on RetinaNet named AureNet where Yolox-HSV algorithm is adopted to simulate the different light intensities for data augmentation. We used RegNet as the backbone of AureNet to extract the features more efficiently. What’s more, we developed a client using C# and Python to lower the threshold to do ship detection for non-professionals. On ship detection dataset HRSC2016, AureNet achieves mAP of 0.5286, which is 1.2% higher than that of the original network RetinaNet, demonstrating the effectiveness of the proposed network.