Infrared and Visible Image Fusion Based on Three-branch Adversarial Learning and Compensation Attention Mechanism

The existing deep learning image fusion methods rely on convolution to extract features and do not consider the global features of the source image. Moreover, the fusion results are prone to texture blurring, low contrast, etc. Therefore, this study proposes an infrared and visible image fusion method with adversarial learning and compensated attention. First, the generator network uses dense blocks and the compensated attention mechanism to construct three local-global branches to extract feature information. The compensated attention mechanism is then constructed using channel features and spatial feature variations to extract global information, infrared targets, and visible light detail representations. Subsequently, a focusing dual-adversarial discriminator is designed to determine the similarity distribution between the fusion result and source image. Finally, the public dataset TNO and RoadScene are selected for the experiments and compared with nine representative image fusion methods. The method proposed in this study not only obtains fusion results with clearer texture details and better contrast, but also outperforms other advanced methods in terms of the objective metrics.