Evaluation of Ship Detection Capability in SAR Images: A Study Based on Convolutional Neural Networks and Objective Assessment Metrics

SAR images of different qualities directly impact various application tasks, such as target detection, recognition, and classification at different levels. However, the task of analyzing the level of image quality quantitatively is challenging and has dramatic effects on the development and application of synthetic aperture radar (SAR) images. To meet the increasing demand for large-scale SAR data and enhance the level of application of SAR ship detection, this paper proposes a method that can be used to evaluate the ship detection capability of SAR images on the basis of convolutional neural networks and objective assessment metrics (CNN-OAM). This method integrates the powerful feature learning capability of deep learning with the quantitative approach associated with objective assessment metrics. By predicting the local target confidence and consulting global SAR image quality assessment metrics, the proposed method can comprehensively evaluate the suitability of different SAR images with respect to ship detection tasks and facilitate image quality grading. The CNN-OAM method employs the analytic hierarchy process (AHP) for indicator weight assignment, thus ensuring a reasonable balance between local target confidence and global image quality. The experimental results of this research reveal that the CNN-OAM method performs well across multiple datasets and scenarios. Compared to the application suitability method, the evaluation results obtained via the proposed method are closer to expert assessments. Furthermore, the results obtained via the proposed method are highly consistent with actual ship detection accuracy, thereby accurately reflecting the detectability of ship targets in the images.