MGSC-Net: Multisemantic Guided Spatial–Frequency Collaborative Network for UAV-Based Infrared and Visible Image Fusion

Abstract

Infrared and visible image fusion (IVIF) plays a critical role in unmanned aerial vehicle (UAV)-based all-weather aerial perception. However, in aerial scenarios, thermal targets are typically small scaled and easily submerged in complex backgrounds. Existing methods often face a severe feature compromise dilemma: enhancing the saliency of tiny infrared targets tends to oversmooth the vast visible background textures, whereas improving background clarity inevitably attenuates the thermal energy. This phenomenon primarily stems from the incompatibility between isomorphic networks and dual-modal heterogeneous data, where fragile small target signals are easily diluted in the spatial domain. To fundamentally alleviate these issues and achieve a highly robust cross-modal feature expression, this article proposes a multisemantic guided spatial–frequency collaborative network (MGSC-Net). In particular, the network adopts an asymmetric architecture at the encoder, designing a dual-branch extraction mechanism to independently decouple the thermal features of small targets and high-fidelity visible textures. In the fusion stage, a spatial–frequency collaborative strategy is introduced, utilizing frequency-domain priors to accurately anchor tiny targets and adaptively guide the fusion of spatial semantics. In the decoding stage, a progressive cross-scale reconstruction framework and a composite loss function are formulated. Extensive experiments on UAV and common datasets demonstrate that the proposed MGSC-Net achieves superior performance in both subjective and objective evaluations. It establishes an excellent visual balance between tiny targets and complex backgrounds, effectively improving the detection accuracy of downstream UAV vision tasks. The GitHub link is https://github.com/nchuhxn/MGSC-Net