Infrared-visible image fusion with double-attention mechanism and adaptive interaction loss

Infrared and visible light image fusion aims to synthesize a more informative result by extracting and integrating complementary salient features within two heterogeneous modalities. Recent research has shown that capturing explicit self-similarity and implicit cross-correlation with the aid of an attention mechanism has garnered significant interest and presents several advantages. However, exploring the complementary relationships more comprehensively and optimizing the interaction degrees of double attentions quantitatively is still a challenging issue. In this paper, a novel infrared and visible light image fusion method exploring a double-attention mechanism is proposed. Specifically, our approach excavates intra- and inter-attention features of source images through a two-step feature extraction strategy and integrates them with an intra-attention block in the feature fusion stage. Additionally, to regulate the interaction of two kinds of attentions optimally, an adaptive interaction loss term is devised. In these ways, the salient infrared targets and visible texture details can be integrated more effectively. In the experiments, the proposed method was contrasted with seven state-of-the-art methods on the TNO and RoadScene datasets. The comprehensive subjective and objective comparisons demonstrate the superiority of our method. In addition, a thorough experiment and discussion on the interaction of intra- and inter-information is presented to validate and analyze the effectiveness of our work further.