End-to-end Trainable Ship Detection in SAR Images with Single Level Features

Abstract

Kongsberg Satellite Services (KSAT) use machine learning and manual analysis done by synthetic aperture radar (SAR) specialists on SAR images in real time to provide a ship detection service.

KSATs current machine learning model has a limited ability to distinguish ships close to each other. For this reason, we aim to employ an end-to-end trainable object detection model, as they can better distinguish nearby objects, since they are not limited by heuristic post processing. Since heuristic post processing in object detection limit the models ability to distinguish ships close to each other, we investigate challenges related to employing an end-to-end trainable ship detection model. Since access to ground truth annotations in SAR images is limited, size and rotation labels are not available for all ships, and rotation labels are inaccurate. Since KSATs internal datasets are collected as part of a time critical operational service, position labels are not exact. Since existing evaluation metrics for object detection are too strict, they do not reflect user needs for this service.

To tolerate missing size and rotation annotations, we base loss label assignment on the distance between objects instead of their IoU, and replace DIoU bounding box loss with a novel size regression loss named Size IoU (SIoU) combined with smooth L1 position loss. To tolerate inaccurate rotation labels, we propose angular direction vector (ADV) regression. To tolerate inaccurate position labels, the loss label assignment makes all predictions responsible for large overlapping regions instead of small disjoint regions. To compare models performance according to user needs, we propose an evaluation metric named Distance-AP (dAP), which is based on mAP, but replaces the IoU overlap threshold with an object center point distance threshold. To reduce duplicate ship predictions, we propose multi layer attention.

Using the LS-SSDD SAR ship dataset, we find that replacing IoU based label assignment with position based label assignment increases dAP from 79% to 86%, and that replacing DIoU with SIoU decreases dAP by only 1%. Using a rotation regression benchmark where datasets have different amounts of rotation label noise, we find that ADV outperforms CSL in terms of mean predicted inaccuracy at all noise levels, and median predicted inaccuracy at high noise levels. Using an object detection benchmark where the datasets have varying amount of position label inaccuracy, we find that the proposed loss label assignment tolerates large amounts of noise without reduced performance. Using KSATs dataset of Sentinel 1 images, we measure 83% dAP.

The proposed mechanisms allow effective training of a ship detection model, despite the missing size and rotation annotations, inaccurate position annotations, and inaccurate rotation annotations. We believe this is useful for KSATs ship detection service, as it can better distinguish nearby ships. However, more work is required to compare its performance with their existing solution. Source code is available at https://github.com/matill/Ship-detection