LG Fusion++: Adaptive Nonlinear Frequency-Aware and Task-Driven Multimodal Image Fusion

Abstract

Multimodal image fusion aims to combine the complementary information from infrared, visible and medical imaging modalities to generate a single enhanced image. Existing methods based on wavelet transform and convolutional neural network are often not able to simultaneously preserve the global structural information and the local texture details well. In this paper, we propose LG Fusion++, an adaptive nonlinear frequency-aware fusion framework that can effectively separate low-frequency structural information and high-frequency texture details. The framework incorporates lightweight attention mechanisms and Mamba based context modeling for efficient global feature learning. A contrast-aware perceptual loss and multi-task optimization strategy are adopted to achieve better edge preservation, texture enhancement and downstream task performance. The experimental results indicate that the proposed framework outperforms existing approaches in terms of entropy, edge preservation, and structural similarity, making it suitable for surveillance, medical imaging, and autonomous systems.