Ship Detection for Optical Remote Sensing Images Based on Visual Attention Enhanced Network.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Ship detection is crucial for military and civil applications.
• Current convolutional neural network (CNN) methods struggle with large images, complex scenes, false alarms, and inshore vessels.
• Existing methods require improvement in accuracy and efficiency for diverse remote sensing data.

Purpose of the Study:

• To propose an advanced ship detection method for optical remote sensing images.
• To enhance detection performance and positioning accuracy, especially for inshore ships.
• To reduce false alarms and improve efficiency in complex remote sensing scenarios.

Main Methods:

• Developed a light-weight local candidate scene network (L2CSN) for efficient candidate region extraction.
• Proposed a visual attention enhanced Deeply Supervised Object Detection (VA-DSOD) method.
• Integrated semantic feature extraction with a visual attention mechanism within DSOD.

Main Results:

• The proposed VA-DSOD method achieved an average precision (AP) of 89.86%, a 7.53% improvement over the baseline SW+DSOD (82.33% AP).
• Demonstrated superior detection and localization performance compared to the baseline in complex remote sensing scenes.
• Effectively reduced false alarms and improved detection efficiency on Google Earth and GaoFen-2 datasets.

Conclusions:

• The visual attention enhanced network significantly boosts ship detection accuracy and positioning in remote sensing.
• The proposed L2CSN and VA-DSOD methods offer a robust solution for challenging ship detection tasks.
• This approach provides a valuable advancement for both military and civil applications requiring precise maritime surveillance.