Reference-guided texture transfer with deformable convolutions for indoor image dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Indoor image dehazing is challenging, requiring clear scene recovery without sacrificing structural integrity or introducing artifacts.
• Existing methods often struggle with preserving fine details and natural textures in indoor environments.

Purpose of the Study:

• To develop a reference-guided indoor image dehazing framework that enhances structural fidelity and texture preservation.
• To improve upon previous texture-transfer models by incorporating deformable feature alignment.

Main Methods:

• The DTTN framework extracts multi-scale features from hazy and reference images.
• It employs Top-K patch matching for transferable texture retrieval and a multi-scale deformable feature integration module.
• A gradient density enhancement module reinforces edge and structural consistency.

Main Results:

• DTTN achieved the highest SSIM (0.992) and competitive PSNR (36.59 dB) on the RESIDE-indoor benchmark.
• The method demonstrated a favorable quality-complexity trade-off.
• Ablation studies confirmed the benefits of reference-guided transfer and deformable alignment for structural fidelity.

Conclusions:

• Reference-guided texture transfer combined with deformable alignment is a highly effective strategy for indoor image dehazing.
• The DTTN framework offers a robust solution for recovering clear indoor images while maintaining structural integrity.