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

Indoor image dehazing requires recovering clear scene content while preserving structural fidelity and avoiding artificial textures. We propose DTTN, a reference-guided indoor image dehazing framework that extends our previous texture-transfer model with deformable feature alignment. The method extracts multi-scale features from the hazy input and a clean reference image, retrieves transferable textures using Top-K patch matching, and integrates them through a multi-scale deformable feature integration module. A gradient density enhancement module is further introduced to reinforce edge and structural consistency. On the RESIDE-indoor benchmark, DTTN achieves the best SSIM (0.992) among the compared methods and competitive PSNR (36.59 dB), while maintaining a favorable quality-complexity trade-off. Additional internal analysis and ablation studies indicate that reference-guided transfer and deformable alignment improve structural fidelity and remain stable under imperfect reference pairing. These results show that reference-guided texture transfer with deformable alignment is an effective strategy for indoor image dehazing under synthetic benchmark conditions.