Learning to Holistically Detect Bridges From Large-Size VHR Remote Sensing Imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Bridge detection in remote sensing images (RSIs) is challenging due to variations in scale and aspect ratio.
• Existing deep learning methods struggle with large-size VHR RSIs due to GPU memory limitations and cropping strategies causing fragmented predictions.
• A scarcity of large-scale datasets hinders the performance of deep learning algorithms for bridge detection in RSIs.

Purpose of the Study:

• To address the limitations in bridge detection within large-size very-high-resolution (VHR) remote sensing images (RSIs).
• To introduce a comprehensive dataset and an efficient network for holistic bridge detection.
• To establish a benchmark for evaluating bridge detection algorithms on large-scale RSIs.

Main Methods:

• A large-scale dataset, GLH-Bridge, comprising 6,000 VHR RSIs (2,048x2,048 to 16,384x16,384 pixels) with 59,737 manually annotated bridges (oriented and horizontal bounding boxes).
• An efficient network for holistic bridge detection (HBD-Net) featuring a separate detector-based feature fusion (SDFF) architecture and a shape-sensitive sample re-weighting (SSRW) strategy.
• SDFF utilizes inter-layer feature fusion (IFF) within a dynamic image pyramid (DIP) for multi-scale context, while SSRW balances regression weights for diverse aspect ratios.

Main Results:

• The proposed GLH-Bridge dataset facilitates robust bridge detection in large-size VHR RSIs.
• HBD-Net demonstrates effectiveness in holistic bridge detection, outperforming existing methods on the new benchmark.
• Cross-dataset generalization experiments confirm the strong applicability of the GLH-Bridge dataset.

Conclusions:

• The GLH-Bridge dataset and HBD-Net significantly advance the field of bridge detection in large-scale remote sensing imagery.
• The developed methods overcome challenges related to image size, aspect ratio variations, and data scarcity.
• The findings provide a valuable resource and benchmark for future research in remote sensing object detection.