Infrared and Visible Image Fusion With Language-Driven Loss and Knowledge Distillation

Infrared and visible image fusion (IVIF) has attracted much attention owing to the highly complementary properties of the two image modalities. Due to the lack of ground-truth fused images, the fusion output of current deep-learning-based methods heavily depends on the loss functions defined mathematically. As it is hard to well mathematically define the fused image without ground truth, the performance of existing fusion methods is limited. In this article, we first propose to use natural language to express the objective of IVIF, which can avoid the explicit mathematical modeling of fusion output in current losses, and make full use of the advantage of language expression to improve the fusion performance. For this purpose, we present a comprehensive language-expressed fusion objective and encode relevant texts into the multimodal embedding space using CLIP. A language-driven fusion model is then constructed in the embedding space by establishing the relationship among the embedded vectors representing the fusion objective and input image modalities. Finally, a language-driven loss is derived to make the actual IVIF aligned with the embedded language-driven fusion model via supervised training. In addition, to ensure that the loss works well with high robustness and generalization in practice, we introduce a novel regularization and patch filtering approach in the training. For real-time applications, we further propose a novel knowledge distillation framework, in which the student conducts self-learning under the supervision of the teacher that has been trained with the language-driven loss. This enables the student to maintain comparable performance and, more importantly, achieve real-time inference even on high-resolution images. Experiments show that our method can obtain much better fusion results than existing techniques.