Data-Driven Bidirectional Spatial-Adaptive Network for Weakly Supervised Object Detection in Remote Sensing Images

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Weakly-supervised object detection (WSOD) methods in remote sensing images (RSIs) struggle with small-scale instances, rare poses, and crowded scenes.
• Current WSOD approaches often overlook valuable candidate proposals by focusing solely on top-scoring regions.

Purpose of the Study:

• To develop an advanced WSOD network for RSIs that mitigates challenges posed by scale, pose variations, and crowded scenes.
• To improve the excavation of entire instances and enhance feature learning by addressing limitations of existing methods.

Main Methods:

• Proposes a data-driven bidirectional spatial-adaptive network (BSANet) incorporating a forward-reverse spatial dropout (FRSD) module.
• The FRSD module acts as a data-driven hard attention mechanism, adaptively sampling and reconstructing spatial regions to uncover latent features.
• Introduces a soft attention branch to model both pixel-level and region-level attention, leveraging complementary benefits.

Main Results:

• The proposed BSANet effectively reduces instance ambiguity caused by extreme scales, poses, and crowded scenes.
• Experimental results on NWPU VHR-10.v2 and DIOR datasets demonstrate significant improvements in detection performance.
• The method achieves new state-of-the-art results on challenging remote sensing object detection benchmarks.

Conclusions:

• The BSANet offers a robust solution for weakly-supervised object detection in remote sensing images.
• The integration of bidirectional spatial adaptation and combined soft/hard attention mechanisms enhances the ability to detect challenging objects.
• This work advances the field of WSOD in remote sensing by setting a new performance benchmark.