Multi-Scale Feature Integrated Attention-Based Rotation Network for Object Detection in VHR Aerial Images

Area of Science:

• Computer Vision and Image Analysis
• Remote Sensing and Geospatial Intelligence
• Machine Learning for Object Detection

Background:

• Object detection in very high resolution (VHR) aerial imagery is crucial for numerous applications.
• Traditional horizontal bounding box (HBB) methods struggle with densely distributed, strip-like objects, scale variations, and diverse backgrounds, leading to missed or redundant detections.
• Limitations of HBBs necessitate advanced approaches for robust aerial object detection.

Purpose of the Study:

• To develop an effective object detection framework for VHR aerial images that overcomes the limitations of traditional HBB methods.
• To improve the accuracy and robustness of multi-class object detection, particularly for challenging object types and scenarios.
• To enhance the performance of oriented bounding box (OBB) detection for aerial imagery.

Main Methods:

• Proposed the Multi-scale Feature Integration Attention Rotation Network (MFIAR-Net), a region-based object detection framework.
• Integrated inherent multi-scale pyramid features to create a discriminative feature map.
• Introduced a double-path feature attention network, guided by ground truth mask information, to focus on object regions and reduce noise, coupled with a robust Rotation Detection Network for efficient OBB representation.

Main Results:

• The MFIAR-Net framework demonstrates effectiveness in detecting multi-class objects in VHR aerial images.
• The proposed attention mechanism successfully guides the network to focus on relevant object regions and suppress background noise.
• Experiments on public datasets validate the superior performance of the MFIAR-Net compared to existing methods, especially for challenging object distributions and scales.

Conclusions:

• The MFIAR-Net provides a significant advancement in aerial object detection, particularly for complex scenes with dense and strip-like objects.
• The integration of multi-scale features and attention mechanisms enhances detection accuracy and robustness.
• The framework's ability to generate efficient OBB representations contributes to improved regression and classification performance.