DBNDiff: Dual-branch network-based diffusion model for infrared ship image super-resolution

Abstract

Infrared ship image super-resolution (SR) is important for dim and small ship object detection and tracking. However, there are still challenges for large-scale factors SR of infrared ship images, as infrared images require a greater amount of global edge information compared to visible images. To overcome this challenge, we introduce a novel dual-branch network-based diffusion model (DBNDiff) for infrared ship image SR, which incorporates a noise prediction (NP) branch and an edge reconstruction (ER) branch within its conditional noise prediction network (CNPN). In the NP branch, to perform better noise prediction, a hybrid cross-attention (HCA) block is used for the interaction between global and local information. In the ER branch, ER blocks are stacked to extract edge information. Furthermore, an edge loss function is introduced to preserve more edges and details. Extensive experiments on infrared ship image datasets highlight that our DBNDiff outperforms other SR methods, especially showing the best visual quality at large-scale factors SR tasks.