PGTFusion: Pseudo ground truth guided infrared and visible image fusion

Infrared and visible image fusion aims to produce a unified representation that is not only visually informative but also semantically meaningful for downstream vision tasks. However, existing supervised fusion methods are fundamentally constrained by the absence of reliable ground truth images, while most unsupervised approaches optimize low-level appearance consistency without explicitly preserving semantic targets. This semantic misalignment often leads to fused images that either suppress critical infrared objects or introduce ambiguous structures, limiting their practical utility. To address this problem, we propose PGTFusion, a novel pseudo ground truth generation framework that introduces explicit semantic guidance into the fusion process. Specifically, salient infrared targets are identified through object detection and semantic segmentation, and are adaptively integrated with visible images to construct pseudo ground truths that preserve task-relevant thermal information while maintaining natural visual structures. In addition, a gradient enhancement module based on guided filtering is designed to reinforce structural and textural fidelity by explicitly supervising edge and detail reconstruction. Extensive experiments demonstrate that PGTFusion consistently outperforms nine state-of-the-art methods, producing fused images with improved semantic target visibility, structural consistency and perceptual quality, thereby offering a more task-oriented and semantically grounded solution for infrared and visible image fusion. Code of this work will be released at https://github.com/uzeful/PGTFusion .