CAFC-Net: A Critical and Align Feature Constructing Network for Oriented Ship Detection in Aerial Images

Area of Science:

• Computer Vision
• Remote Sensing
• Deep Learning

Background:

• Ship detection is crucial for maritime surveillance and management.
• Convolutional Neural Networks (CNNs) have advanced ship detection, but challenges remain.
• Issues include large scale variation, aspect ratios, and dense distribution in aerial imagery.

Purpose of the Study:

• To propose an end-to-end single-stage rotation detector, the Critical and Align Feature Constructing Network (CAFC-Net).
• To enhance ship detection accuracy in aerial images by addressing existing limitations.

Main Methods:

• Developed CAFC-Net comprising Biased Attention Module (BAM), Feature Alignment Module (FAM), and Distinctive Detection Module (DDM).
• BAM extracts critical features for classification and regression.
• FAM generates high-quality anchor boxes using biased regression features and Alignment Convolution.
• DDM produces orientation-sensitive and reconstructs orientation-invariant features.

Main Results:

• CAFC-Net demonstrated state-of-the-art performance on HRS2016 and self-built ship datasets.
• The proposed modules effectively address scale variation, aspect ratio, and dense distribution challenges.
• Improved accuracy in both classification and localization of ships.

Conclusions:

• CAFC-Net significantly improves ship detection accuracy in aerial imagery.
• The novel architecture and modules offer a robust solution for challenging remote sensing scenarios.
• This work advances the field of object detection in computer vision.