EvoASF: Text guided evolutionary adaptive fusion for human preference aligned infrared visible image fusion

Significant progress has been achieved by existing infrared and visible image fusion methods in preserving structural details and enhancing salient targets. However, their performance remains limited under complex imaging conditions due to two major challenges. Firstly, most existing methods mainly optimize objective quality metrics, and the discrepancy between mathematical optimization objectives and human perceptual preferences remains insufficiently investigated. Secondly, the interaction between infrared salient information and visible-light texture structures is often insufficient, resulting in inconsistent detail preservation and perceptual inconsistency. To address these issues, we propose a framework for infrared and visible image fusion that is perceptual-aware and named Evolutionary Adaptive Strategy Fusion (EvoASF). The proposed EvoASF incorporates two key designs. First, it introduces a channel-adaptive feature exchange encoder that estimates the channel-wise importance of channels for features across different modalities. This encoder selectively models the complementary relationship between infrared and visible modalities in a hierarchical feature space. This enhances cross-modal feature interaction and adaptive fusion between infrared saliency responses and visible light texture structure representations. Secondly, it introduces a preference-guided optimization mechanism that combines objective quality measurements, such as structure, gradient and intensity, with perceptual preference signals derived from expert evaluations and large language models. This achieves a better balance between information preservation and subjective visual quality. Extensive experiments conducted on public datasets demonstrate the superior performance of EvoASF in both objective evaluations and subjective visual perception.