Hybrid Gaussian Deformation for Efficient Remote Sensing Object Detection | JoVE Visualize

Area of Science:

Computer Vision
Remote Sensing
Artificial Intelligence

Background:

Large-scale high-resolution remote sensing images (LSHR) offer detailed object information but incur significant computational costs.
Current object detection methods for LSHR often struggle with balancing accuracy and computational efficiency.
Existing approaches rely on high-resolution inputs, limiting performance gains from computational optimizations.

Purpose of the Study:

To develop an efficient object detection framework for LSHR that reduces computational load without sacrificing accuracy.
To propose a method that intelligently processes image data by preserving object details and compressing background regions.
To introduce novel modules for dynamic sampling, feature extraction, and fusion tailored for LSHR object detection.

Main Methods:

A hybrid Gaussian deformation module was designed for dynamic sampling, adjusting density based on region relevance to enhance object feature preservation.
A bilateral deform-uniform detection framework was introduced, utilizing both deformed low-resolution and original high-resolution images.
Key components include a deformed deep backbone for semantic information, a uniform shallow backbone for spatial details, a deformation-aware feature registration module, and a feature relationship interaction fusion module.

Main Results:

The proposed method significantly reduces computational costs associated with LSHR object detection.
Experimental results on three challenging datasets demonstrate superior performance compared to existing state-of-the-art methods.
The framework effectively balances the trade-off between detection accuracy and computational efficiency.

Conclusions:

The developed framework offers an effective solution for efficient object detection in LSHR.
Dynamic sampling and a dual-stream processing approach are key to achieving high performance with reduced computation.
This work advances the field of remote sensing image analysis by enabling more efficient and accurate object detection.