Visible and thermal image fusion network with diffusion models for high-level visual tasks

Fusion technology enhances the performance of applications such as security, autonomous driving, military surveillance, medical imaging, and environmental monitoring by combining complementary information. The fusion of visible and thermal (RGB-T) images is critical for improving human observation and visual tasks. However, the training of most semantics-driven fusion algorithms combines segmentation and fusion tasks, thereby increasing the computational cost and underutilizing semantic information. Designing a cleaner fusion architecture to mine rich deep semantic features is the key to addressing this issue. A two-stage RGB-T image fusion network with diffusion models is proposed in this paper. In the first stage, the diffusion model is employed to extract multiscale features. This provided rich semantic features and texture edges for the fusion network. In the next stage, semantic feature enhancement module (SFEM) and detail feature enhancement module (DFEM) are proposed to improve the network’s ability to describe small details. An adaptive global-local attention mechanism (AGAM) is used to enhance the weights of key features related to visual tasks. Specifically, we benchmarked the proposed algorithm by creating a new tri-modal sensor driving scene dataset (TSDS), which includes 15234 sets of labeled images (visible, thermal, and polarization degree images). The semantic segmentation model trained on our fusion images achieved 78.41% accuracy, and the object detection model achieved 87.21% MAP. The experimental results indicate that our algorithm outperforms the state-of-the-art image fusion algorithms.