Remote Sensing Image Generation via Object Text Decoupling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Remote sensing images present challenges for generative models due to complex object structures and varied locations.
• Conventional methods struggle with object fidelity and texture in generated remote sensing imagery.
• Existing object-level controls in generation are limited by layout restrictions.

Purpose of the Study:

• To develop a novel method for high object fidelity, layout diversity, and customizable generation of remote sensing images.
• To overcome the limitations of layout restrictions in previous object-level controlled generation approaches.
• To enhance the quality and diversity of generated remote sensing images through improved object control.

Main Methods:

• Proposed Object Text Decoupling Generative Adversarial Network (OTD-GAN) for remote sensing image generation.
• Developed an object text decoupling module to predict semantically consistent textual representations for objects.
• Implemented an object text semantic enhancement module to fuse object-specific text with global captions.

Main Results:

• OTD-GAN achieves object-level controls without imposing layout restrictions.
• The method enhances generation quality by integrating decoupled object text representations.
• Experiments on NWPU-Captions and RSICD datasets show leading performance against state-of-the-art methods.

Conclusions:

• OTD-GAN successfully enables customizable remote sensing image generation with high object fidelity and layout diversity.
• Decoupling object text representations from global captions is effective for improving generation quality.
• The proposed approach offers a flexible and powerful solution for advanced remote sensing image synthesis.