TSTBFuse: a two-stage three-branch feature extraction method for infrared and visible image fusion

The purpose of image fusion is to combine information from different source images to produce a comprehensively representative image. Traditional autoencoder architectures often struggle to effectively extract both unique and shared features from these image types. A novel two-stage three-branch feature extraction method (TSTBFuse) was proposed in the study, specialized for the fusion of infrared and visible images. The proposed architecture employed a three-branch encoder that separately captured infrared-specific thermal radiation features, visible-specific texture details, and shared structural information. A two-stage end-to-end training strategy was introduced: the first stage focused on reconstructing the original input images to preserve modality-specific information, while the second stage leveraged the learned representations to generate high-quality fused images. we designed a comprehensive loss function combining mean squared error (MSE), structural similarity index (SSIM), and gradient loss, ensuring both pixel-level accuracy and structural integrity. Extensive experiments on public datasets (TNO, MSRS and RoadScene) demonstrated that TSTBFuse consistently outperformed seven state-of-the-art methods in both subjective and objective evaluations. Furthermore, the method exhibited strong generalization capabilities, successfully extending to challenging tasks such as magnetic resonance imaging-computed tomography (MRI-CT) medical image fusion and red-green-blue (RGB)-infrared image fusion without retraining. The code is publicly available at: https://github.com/QXinYue/TSTBFuse.