Highly Adaptive Ship Detection Based on Arbitrary Quadrilateral Bounding Box

With the rapid development of space remote sensing technology, accurate ship detection based on high-resolution optical remote sensing images has steadily attracted considerable research interest. However, most of the current methods adopt a fixed horizontal detection frame to predict the target. Although these methods have good detection accuracy, because the ship's orientation is arbitrary in reality, a large error occurs in the matching degree of the detection effective area, resulting in inaccurate target detection. Therefore, this paper proposes a ship detection algorithm based on an arbitrary quadrilateral prediction frame. We redefine the loss function and directly predict the detection frame's four vertices through the designed eight-parameter regression process. In addition, the convolutional block attention module (CBAM) is introduced to optimize the original network structure, and the clustering method is used to optimize the calculation of the anchor point. To replace the intersection over union (IoU), which cannot distinguish different alignments of objects, we adopt a generalized intersection over union (GIoU). Finally, we conduct experiments based on the DOTA ship dataset and the HRSC2016 dataset. The results show that our method is better than YOLOv3 and other commonly used target detection algorithms in terms of accuracy and visualization. Meanwhile, we compared with SOTA algorithm in real-time and dense ship detection. Experimental results prove that its speed and performance on mobile platform are in the lead, and it has a great effect on dense ship detection.