Infrared and visible image fusion based on negative-value calibration and adaptive hierarchical activation

Abstract

To address modality contribution imbalance and computational redundancy in cross-modal image fusion, this letter proposes a fusion method based on negative-value calibration and adaptive hierarchical activation. The proposed method dynamically partitions the computation into shallow and deep paths, enabling information-density-aware resource allocation to reduce unnecessary computation while preserving fusion quality in critical regions. In addition, negativevalue calibration is introduced to constrain the attention region, thereby suppressing artifacts and distortions caused by conflicting activations. Using infrared–visible image fusion as an example, experiments on public datasets demonstrate that the proposed method outperforms existing approaches in both visual quality and multiple objective metrics. Specifically, MI, EN, CC, SSIM, SD, and 𝑄𝐴𝐵/𝐹 are improved by 7.05%, 7.83%, 13.14%, 4.23%, 6.71%, and 14.96%, respectively, while the average fusion time is 0.0065 s per image pair. These results indicate that the proposed method improves fusion quality in low-visibility scenes while achieving superior efficiency.