图像融合
计算机视觉
人工智能
红外线的
融合
图像处理
计算机科学
图像(数学)
可见光谱
光学
图像分割
传感器融合
材料科学
迭代重建
可视化
光学滤波器
图像配准
特征提取
图像传感器
图像分辨率
辐射测量
图像质量
作者
Liangjie Song,Gang Liu,Hanlin Xu,Bo Liu,Yao Qian,Yiyao Zhao
标识
DOI:10.1109/tcsvt.2026.3727898
摘要
Existing infrared and visible image fusion methods suffer from limited texture perception capabilities, constrained by training on source images with scarce textures and inherent network limitations in extracting and restoring fine details. To address this, we propose TAFuse, a novel Texture-Aware fusion framework designed to explicitly enhance texture perception. TAFuse employs a multi-scale autoencoder architecture featuring a scene-adaptive adjustment module to dynamically perceive and enhance texture representations based on content relevance. During encoding, the network extracts and refines multi-scale spatial features, subsequently transforming them into wavelet and frequency domains for fine-grained texture perception. The decoding stage utilizes a Fourier–Butterworth hybrid reconstruction core to restore structural details, actively leveraging the perceived cross-scale high-frequency components to guide reconstruction. Furthermore, we introduce a gradient enhancement loss and a low frequency masking enhancement loss to explicitly supervise the texture perception process. The former employs a Difference-of-Gaussians filter for controllable texture enhancement via gradient manipulation, while the latter uses a low-frequency mask to protect salient targets from over-enhancement, ensuring balanced perception. Extensive experiments on standard IVIF benchmarks, downstream tasks, and medical fusion scenarios demonstrate TAFuse’s superior texture perception capabilities.
科研通智能强力驱动
Strongly Powered by AbleSci AI