A Multilayer Fusion Light-Head Detector for SAR Ship Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Synthetic aperture radar (SAR) ship detection faces challenges with traditional methods and existing deep learning models like Faster R-CNN (FRCN).
• FRCN struggles with multiscale SAR ship objects, slow detection speeds, and high false detection rates in complex backgrounds due to fixed receptive fields and imbalanced datasets.

Purpose of the Study:

• To address the limitations of current methods in SAR ship detection.
• To develop a more accurate and faster detection system for multiscale SAR ship objects.

Main Methods:

• Proposed a multilayer fusion light-head detector (MFLHD) integrating shallow high-resolution and deep semantic features for region proposal.
• Introduced a light-head detector with large-kernel separable convolution and position-sensitive pooling for improved detection speed.
• Adapted focal loss to train on hard examples, reducing false alarms in complex backgrounds.

Main Results:

• The MFLHD achieved superior performance on the SAR Ship Detection Dataset (SSDD).
• Demonstrated significant improvements in both accuracy and detection speed compared to existing methods.
• Effectively handled scale variability and reduced false detections in complex SAR imagery.

Conclusions:

• The MFLHD offers a promising solution for efficient and accurate SAR ship detection.
• The proposed fusion strategy and light-head detection network contribute to overcoming key challenges in the field.
• This work advances the application of deep learning for maritime surveillance using SAR imagery.