7 November 2024 in Expert Systems with Applications
17 September 2024 in Displays
28 January 2026 in Expert Systems with Applications
30 October 2025 in Knowledge-Based Systems
17 August 2026 in SSRN Electronic Journal
Infrared and visible image fusion aims to collaboratively preserve the salient thermal targets in infrared images and the rich texture details in visible images. However, existing deep learning methods adopt a pixel-level feature learning paradigm based on Euclidean space, relying on local receptive fields or regular grid modeling, which makes it difficult to explicitly characterize the topological structural relationships between functional regions of the image. This results in the loss of structural detail continuity in visible scenes and fails to fully capture the semantic integrity and regional correlation features of infrared target areas. To address these issues, this paper proposes an infrared and visible image fusion method based on regional topological representation learning and cross-representation interaction. First, the image feature learning is extended from pixel-level representation to region-level topological relationship modeling by constructing a regional structure graph to explicitly depict the structural associations between functional regions of the image. Second, a collaborative learning framework for region-level semantics and pixel-level details is established to achieve complementary modeling of these two heterogeneous representations. Finally, a cross-representation interaction mechanism is designed to realize the collaborative optimization and consistency enhancement of region-level semantic and local texture features. Experimental results on multiple public datasets demonstrate that the proposed method outperforms many existing approaches in maintaining structural continuity, target saliency, and visual fidelity.