Relational Part-Aware Learning for Complex Composite Object Detection in High-Resolution Remote Sensing Images.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Complex composite object detection in high-resolution remote sensing images (RSIs) is challenging due to discrete parts, variable layouts, and blurred boundaries.
• Existing methods struggle with weak inter-relationships and subtle features of composite objects like power plants and harbors.

Purpose of the Study:

• To propose an end-to-end framework, the relational part-aware network (REPAN), for improved detection of complex composite objects in RSIs.
• To explore semantic correlations and extract discriminative features among multiple object parts.

Main Methods:

• A part region proposal network (P-RPN) with butterfly units (BFUs) to locate discriminative regions and alleviate feature-scale confusion.
• A feature relation Transformer (FRT) for joint part-and-global learning to capture spatial relationships and enhance part representation.
• A contextual detector (CD) utilizing multirelation-aware features for classifying and detecting both parts and the whole composite object.

Main Results:

• The proposed REPAN framework consistently surpassed state-of-the-art methods in extensive experiments.
• Evaluation on three remote sensing object detection datasets with four categories demonstrated the method's effectiveness.
• The approach successfully addresses challenges posed by variable layouts and blurred boundaries in composite object detection.

Conclusions:

• The relational part-aware network (REPAN) is an effective and superior method for complex composite object detection in high-resolution remote sensing images.
• The framework's ability to explore semantic correlations and spatial relationships among parts significantly enhances detection accuracy.
• REPAN provides a robust solution for challenging remote sensing object detection tasks, particularly for objects with intricate structures.