Infrared and visible image fusion network based on dense connections and dual-branch attention mechanism

Infrared and visible images can be combined, bringing together the prominent target information from the infrared field and the rich background texture details from the visible field, which is beneficial for high-level vision tasks. Nevertheless, the feature extraction modules in current deep learning-based image fusion algorithms have difficulties in preserving the effective information from both infrared and visible images. To address this problem, this study proposed a fusion network—namely, DCDAM-FusionNet-Relyed on dense connections and a dual-branch attention mechanism, it is capable of adaptively acquiring features from diverse modalities to accomplish efficient information integration. An efficient attention mechanism was combined with dense connections in this method to enhance the network's ability to represent features. A channel corner attention module and a dual-frequency interactive attention mechanism were adopted as well. They were used to extract detailed information from visible light images and prominent information from infrared images respectively. The DCDAM-FusionNet model was trained using pixel and texture losses as loss functions. The TNO and Multi-Spectral Road Scenarios public datasets results demonstrated that the DCDAM-FusionNet model outperformed the present state-of-the-art fusion algorithms in subjective and objective evaluations.