SPAFNet: Semantic Prior-Guided Progressive Alignment and Gated Bidirectional Fusion for Unaligned RGB-T Salient Object Detection

Unaligned RGB and thermal images require spatially tolerant fusion when parallax, sensor offset and modality-specific artifacts coexist. SPAFNet combines four-scale dense feature sampling with a shared semantic prior and gated bidirectional convolutional fusion. Sampling displacement depends on inherited coarse flow, local matching expectations and learned residuals. A saliency-supervised gate controls the thermal contribution without direct correspondence calibration. This manuscript develops the architecture’s component interfaces, coordinate conventions and evaluation requirements, supported by schematic figures and an explicit evidence record. The supplied evaluation summary describes two naturally unaligned, three synthetically perturbed and three aligned datasets; complete results for this coverage are unavailable. The supplied UVT20K comparison reports F_m values of 0.887 for SPAFNet and 0.822 for PCNet. The SPAFNet estimate has not been independently reproduced, and common evaluation conditions have not been established. These point estimates do not establish physically correct registration, statistical significance or generalization to unseen imaging systems. Expanded preprint draft: 10 pages, 3 explanatory schematic figures, 4 tables and 12 references. No new experiments were conducted. The editable LaTeX source package is included.