AIR-GANet: multi-head attention integrated residual dense block based generative adversarial network for visible and infrared image fusion

AIR-GANet:一种基于多头注意力机制的残差密集块生成对抗网络,用于可见光和红外图像融合

阅读:1

Abstract

Image fusion is a sophisticated improvement technique aimed at integrating data from multiple sensors into a single, detailed and coherent image. This integrated representation supports and simplifies subsequent processing tasks. This article presents a well-crafted fusion method depending on Multi-Head Attention (MA) and Residual Dense Block (RDB) based Generative Adversarial Network (AIR-GANet) for infrared and visible image fusion. The framework utilizes RDB for extracting local features, Multi-Head Attention for capturing global dependencies, ensuring a comprehensive representation of both infrared and visible modalities. A dual discriminator mechanism is incorporated with separate discriminators to refine the fused image through adversarial training, preserving both thermal details and texture information. The fusion module incorporates local and global features into a unified feature map, which is then decoded into the final fused image. By applying AIR-GANet to input infrared ([Formula: see text]and visible ([Formula: see text]) images, the resulting fused image maintains crucial informative data from both sources, thus improving overall image quality and enabling high-performance across multiple quantitative evaluation metrics. Extensive evaluations on the TNO, Road Scene and Multi-Spectral Road Scenarios (MSRS) datasets demonstrate that the model attains average Entropy (EN) of 7.4085 and Mutual Information (MI) of 3.1531 on TNO, corresponding values of 7.7086 and 3.6953 on Road Scene, 7.8260 and 3.8502 on MSRS confirming its superior fusion capability across diverse scenarios.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。