DPFusion: Detail-preserving dual-path infrared and visible image fusion in high-radiance thermal scenes

Infrared and visible image fusion integrates thermal radiation responses with visible textures, providing complementary information for observing high-radiance thermal targets. Most existing deep fusion methods enhance single-modality features before cross-modal fusion; the enhancement supplements context but tends to weaken original edges and weak textures and causes modality bias under strong thermal responses. This paper therefore proposes DPFusion, a detail-preserving dual-path fusion network with a three-scale dual-stream encoder, a parallel residual fusion (PRF) module, and a shared decoder. At each scale, PRF places an original path and an enhanced path in parallel: the former directly fuses the infrared and visible encoded features, while the latter supplements statistical and directional context through triple-statistic enhanced attention (TSEA) and is weighted by a shared cross-modal adaptive fusion (CAF) module; the two outputs are combined residually with scale-wise learnable coefficients. TSEA combines average, maximum, and median statistics to stabilize feature description under unbalanced infrared responses; CAF normalizes the two modality weights along the modality dimension to explicitly model their relative contributions. Training uses self-reconstruction and cross-reconstruction constraints on the recoverability of source information from the fused features, without manual fusion ground truth. On MFNet, M3FD, and LLVIP, DPFusion achieves the best SF and AG, improving by 18.7% and 25.9% over the second-best method on M3FD. Wildfire and field explosion fireball experiments further show that it preserves flame boundaries, fragment contours, and weak dark-region structures under strong thermal radiation, local saturation, and smoke interference, providing image support for structural interpretation of high-radiance transient thermal targets.