Oriented Ship Target Detection Based on Improved CenterNet in Synthetic Aperture Radar Images

Abstract

Deep learning has a wide application prospects in the field of the ship target detection in synthetic aperture radar (SAR) images. The existing researches mainly use anchor-based target detection method, but this kind of methods is not suitable for the ship SAR image with the sparse ship targets. It requires additional computational resources to filter out a large number of overlapped candidate prediction boxes, which tends to result in the inaccurate target position and low detection efficiency. At the same time, most of existing methods use the horizontal box to detect the ship targets, which are not suitable for detecting the large aspect ratio ship targets and densely arranged ship targets. Aiming at the problem of the insufficient data, a data augmentation strategy for the ship SAR images is proposed, which effectively alleviates the problem of over-fitting during training, improves the detection performance and stability of the model. Then, an oriented ship target detection method is presented for the SAR images, which extends the CenterNet to the field of the rotating target detection. Compared with the existing ship detection method of the SAR images based on the deep learning, this method is very easy to implement, has the high detection precision and positioning accuracy, which has been conducive to the ship detection in practice.