Resolution Preserving and Utilization Network for Tiny Object Detection in Large-Size Remote Sensing Imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Tiny object detection (TOD) in large-size remote sensing imagery (LSRSI) is difficult due to foreground signal identification issues with increasing scene size.
• Existing methods struggle to maintain performance as input image dimensions grow.

Purpose of the Study:

• To develop a novel backbone network for efficient TOD in LSRSI.
• To address the challenge of preserving and utilizing low-level spatial features for tiny object extraction in large remote sensing scenes.

Main Methods:

• Introduction of the Resolution Preserving and Utilization Network (RPUN), a backbone architecture designed for low-level spatial feature preservation and utilization.
• Development of GF2UBSv2, a large-scale dataset for tiny urban bridge detection in panchromatic satellite imagery.

Main Results:

• RPUN shows excellent ability in identifying foreground tiny object features, even with increased input sizes.
• The proposed network maintains detection performance comparable to that achieved with smaller input slices.
• Experiments on GF2UBSv2, DIOR, SODA-A, and DOTAv2.0 datasets demonstrate RPUN's superior performance over state-of-the-art methods.

Conclusions:

• RPUN effectively enhances tiny object detection in large-size remote sensing imagery by preserving and utilizing spatial features.
• The proposed network offers a robust solution for TOD in challenging remote sensing scenarios.
• The GF2UBSv2 dataset and RPUN architecture contribute significantly to advancing research in this field.