Source-linked AI summary
Automatic Ship Detection of Remote Sensing Images from Google Earth in Complex Scenes Based on Multi-Scale Rotation Dense Feature Pyramid Networks
Xue Yang, Hao Sun, Kun Fu, Jirui Yang, Xian Sun, Menglong Yan, Zhi Guo
TL;DR
Ship detection in remote-sensing images must handle complex scenes, densely arranged ships, and redundant regions from horizontal detection. The paper proposes R-DFPN with dense feature pyramids, rotation anchors, and multi-scale ROI Align, and reports state-of-the-art performance, including 88.2% Recall, 91.0% Precision, and 89.6% F-measure. Its conclusion is strongest for complex scenes, especially densely arranged ships, but false alarms leave Precision lower than Faster-RCNN and FPN.
Problem
Ship detection is limited by complex application scenes, intensive objects, and redundant detection regions, with narrow ship widths and ship-like port interference further challenging feature discrimination.
Method
R-DFPN combines dense feature pyramids, rotation anchors, and multi-scale ROI Align in an end-to-end multi-scale rotation-region detection framework.
Results
88.2% Recall, 91.0% Precision, and 89.6% F-measure were reported for remote-sensing ship detection, with state-of-the-art performance.
Takeaways & Limitations
R-DFPN performs especially strongly for ship detection in complex scenes and densely arranged objects while reducing redundant detection regions.
Takeaways & Limitations
More false alarms result in much lower Precision for R-DFPN than Faster-RCNN and FPN, motivating future work to reduce false alarms.
Abstract
from arXiv · showhide
Ship detection has been playing a significant role in the field of remote sensing for a long time but it is still full of challenges. The main limitations of traditional ship detection methods usually lie in the complexity of application scenarios, the difficulty of intensive object detection and the redundancy of detection region. In order to solve such problems above, we propose a framework called Rotation Dense Feature Pyramid Networks (R-DFPN) which can effectively detect ship in different scenes including ocean and port. Specifically, we put forward the Dense Feature Pyramid Network (DFPN), which is aimed at solving the problem resulted from the narrow width of the ship. Compared with previous multi-scale detectors such as Feature Pyramid Network (FPN), DFPN builds the high-level semantic feature-maps for all scales by means of dense connections, through which enhances the feature propagation and encourages the feature reuse. Additionally, in the case of ship rotation and dense arrangement, we design a rotation anchor strategy to predict the minimum circumscribed rectangle of the object so as to reduce the redundant detection region and improve the recall. Furthermore, we also propose multi-scale ROI Align for the purpose of maintaining the completeness of semantic and spatial information. Experiments based on remote sensing images from Google Earth for ship detection show that our detection method based on R-DFPN representation has a state-of-the-art performance.
1. Introduction
Ship detection remains challenging in complex remote-sensing scenes because of uncertain imaging conditions, dense objects, and redundant horizontal detection regions. R-DFPN addresses these issues with dense feature pyramids, rotation anchors, and multi-scale ROI Align.
- Ship detection is challenged by lighting, disruptors, ship density, and other uncertainties in remote-sensing images.
- Traditional methods use sea-land segmentation, candidate-region extraction, and post-processing, but have poor practicability in complex scenarios.
- Horizontal detectors can produce large redundant regions for inclined, high-aspect-ratio ships, causing non-maximum suppression difficulties and missed detections.
- R-DFPN is an end-to-end framework combining dense feature pyramids, rotation anchors, and multi-scale ROI Align for ship detection.The framework targets complex scenes, densely arranged objects, and redundant detection regions.
- Dense feature-pyramid connections enhance feature propagation and reuse, while rotation-based detection reduces redundant regions and supports densely arranged ships.
2. Proposed Method
R-DFPN combines dense multi-scale feature fusion with rotation-aware proposal detection and multi-scale ROI Align for ship detection in complex scenes. Its components target narrow ships, dense arrangements, redundant regions, and preservation of semantic and spatial information.
- Overall Framework: R-DFPN consists of DFPN for feature fusion and RDN for rotational proposal generation, regression, and classification.DFPN produces fused multi-scale feature maps; RDN uses rotational proposals before Fast-RCNN prediction.
- Dense Feature Pyramid Network: DFPN densely connects multi-level feature maps, concatenating upsampled and lateral features to enhance propagation and reuse.The network uses top-down and lateral pathways, nearest-neighbor upsampling, concatenation, and convolutional refinement.
- Dense Feature Pyramid Network: DFPN significantly improves detection performance through smoother feature propagation and feature reuse.The authors report this finding from extensive experimental comparison.
- Rotation Region Detection: Rotation anchors regress oriented proposals to reduce non-maximum-suppression effects and improve recall for densely arranged, high-aspect-ratio ships.Anchors vary by scale, aspect ratio, and angle, with ship-oriented ratios and six rotation angles.
- Rotation Bounding Box: The rotation bounding box uses center coordinates, width, height, and angle, with the angle defined over [−90,0).This representation uniquely determines an arbitrary oriented rectangle.
- Multi-Scale ROI Align: Multi-scale ROI Align processes the proposal’s horizontal circumscribed rectangle with pool sizes 3:16 and 16:3 to reduce feature distortion and misalignment.The method addresses narrow sides of large-aspect-ratio ships and aims to preserve useful semantic and spatial information.
3. Experiments and Results
Experiments use Google Earth remote-sensing scenes and compare R-DFPN configurations, detection methods, efficiency, and precision–recall behavior. R-DFPN achieves the reported state-of-the-art detection performance while maintaining relatively fast training and testing.
- 3.1. Dataset: The dataset contains Google Earth scenes spanning civilian ports, military bases, offshore areas, and far seas, divided into overlapping sub-images.It comprises 1000 large images covering 400 square kilometers and uses red, green, and blue bands after geometric correction.
- 3.1. Experimental Setup: The experiments use ResNet-101 initialization and adopt end-to-end training because it has the same effect as alternating training while being more convenient.Training runs for 80k iterations with staged learning rates.
- 3.2. Accelerating Experiment: Selecting 12,000 highest-confidence anchors and generating 1,200 proposals through non-maximum suppression balances stable assessment indicators against sharply changing computation time.Increasing the anchors/proposals pair can eventually slightly decrease model results.
- 3.3. Comparative Experiment: 88.2% Recall, 91.0% Precision, and 89.6% F-measure are reported as state-of-the-art performance for the proposed method.Table 1 compares detection methods at a confidence threshold of 0.5, with R, P, and F denoting Recall, Precision, and F-measure.
- 3.3. Comparative Experiment: R-DFPN-1 exceeds R2CNN by 1.8% in Recall, indicating that rotation anchors substantially affect Recall in this comparison.The cited comparison concerns methods using multi-scale pool size.
- 3.3. Comparative Experiment: R-DFPN-4 improves Recall and Precision over R-DFPN-3, reaches the highest 89.6% F-measure, and has the best equilibrium point on the precision–recall curves.The method also maintains relatively fast training and testing time according to the reported comparison.
4. Discussion
R-DFPN performs strongly on multi-scale and densely arranged ships, but false alarms and angle-sensitive IoU misjudgments limit Precision. The discussion identifies these errors and evaluates rotation-region methods under revised criteria.
- R-DFPN offers superior performance for multi-scale and high-density objects, although its Precision is lower than a traditional method.The authors attribute the lower Precision to false alarms and misjudgment.
- 4.1. False Alarm: Roofs, container piles, and docked docks have aspect ratios similar to ships and cause detector false alarms in complex port scenes.Figure 10 illustrates false alarms from roofs, container piles, docked docks, and floating objects.
- 4.1. False Alarm: DFPN reuses feature information but does not fully eliminate false alarms; sea-land segmentation and generative adversarial networks are suggested as possible directions.The proposed alternatives are constrained by segmentation accuracy or described as future ideas.
- 4.2. Misjudgment: For a ship with aspect ratio 1:7, a 15-degree angle difference reduces IoU to 0.38, causing missed detections at a confidence threshold of 0.55.The discussion describes this angle-sensitive IoU relationship as a source of misjudgment and compares it with circumscribed-rectangle evaluation.
- 4.2. Misjudgment: Under the new evaluation criteria, rotation-region methods show obvious improvements in Recall and Precision.The revised indicators are reported in Table 3.
5. Conclusions
The paper concludes that R-DFPN combines dense feature fusion, rotation proposals, and multi-scale ROI Align for complex ship-detection scenes. Experiments report state-of-the-art performance, especially for densely arranged ships, while false alarms remain a Precision limitation.
- R-DFPN is a multi-scale rotation-region detector designed for complex scenes, intensive objects, and reduced redundant detection regions.The framework includes DFPN, rotation anchors, and multi-scale ROI Align.
- DFPN enhances feature propagation and reuse, rotation anchors improve proposal quality, and multi-scale ROI Align preserves semantic and spatial information.The second-stage regression box and ROI are horizontal circumscribed rectangles of the proposals.
- More false alarms produce substantially lower R-DFPN Precision than Faster-RCNN and FPN, motivating future work to reduce false alarms.