A Prediction and Prior Information Guided SAR Ship Detection Method

Abstract

Data-driven ship detection methods via deep learning algorithms are the recent research hotspot. In this family of models, it is essential to divide the positive and the negative samples. The commonly used strategy is called label assignment. In the previous detectors, label assignment resulted from the handcrafted heuristic techniques. Therefore, it needs to tune the hyperparameters, and improper settings will deteriorate model performance. Moreover, a significant inconsistency between the training and the testing objective is available. To address these issues, a prediction and prior information guided label assignment technique is proposed. A network specific to ship detection is then presented to improve multiscale detection performance. First, the IoU prediction is specified as the estimation of localization precision. The quality of the candidate anchor is evaluated by a combination of classification and positioning. It reduces the inconsistency between training and testing. Besides, the learning status of the current model and the anchors' prior information are exploited simultaneously. Optimum positive samples are selected in an adaptive manner. Finally, a multiscale ship detection network is designed, concentrating on small ships' rich contextual information. After feature fusion and feature enhancement on different scales, shallow texture and deep semantic information are combined to detect multiscale ships. Multiple experiments are conducted on SSDD and HRSID datasets, and the results demonstrate the advantage of the proposed method compared with advanced detectors.

Keywords

• HRSID
• label assignment
• SSDD
• ship detection