Reference-Guided Global-Local Context Fusion Network for Residual Image Restoration

Abstract

In optical measurement environments for weapons system test and evaluation, imagery is frequently degraded by haze and smoke, impairing downstream analysis. Existing methods rely on physics-based models or single-image deep learning, both struggling under non-uniform haze or recovering occluded structures. This paper proposes a residual-based restoration network for fixed-camera settings where a clean reference image exists before degradation. Restoration is redefined as estimating the change relative to the reference rather than reconstructing all pixels. Three key designs are introduced: a shared-encoder dual-input structure with absolute-difference skip connections focusing on changed regions; a dual global context fusion module injecting a FiLM-conditioned global vector and spatial change map into the bottleneck and final decoder; and a multi-objective loss combining Charbonnier-SSIM, change-weighted reconstruction, and mask-based residual alignment. Experiments in the target fixed-camera reference-guided setting demonstrate substantial gains over single-image baselines.