Semantic-Guided Diffusion for UAV Maritime Target Detection: Fusing RGB -Thermal Registration with Contextual Feature Modulation
Adnan Hashim Abdulkadhim, Alaa Ali Hussein · Kirkuk journal of science · 2026
A major fence to UAV based maritime surveillance is the significant strangeness between RGB and thermal infrared modalities which manifests in divergent radiometric signatures texture patterns and feature distributions. Two main challenges confuse maritime target detection the ever-changing sea background and the minimal pixel footprint of small objects. Yet existing approaches fall small they either exhibit insufficient robustness in multimedia recording or overlook the use of high-level semantic cues when integrating information. In this study we put forward the Semantic Guided Diffusion Framework (SGDF) an integrated architecture that jointly optimizes the registration and fusion of RGB and thermal infrared imagery, thereby boosting detection accuracy for maritime targets. Whereas existing methods decouple registration and fusion into separate stages SGDF adopts a unified diffusion-based architecture. Its three main components are: (1) target-conscious preprocessing that isolates maritime objects using YOLOv11-generated bounding boxes and SAM2 segmentation (2) a textual semantic interaction branch that extracts contextual descriptions with a vision-language model to guide feature modulation, and (3) conditional deployment that concurrently estimates homography and merges multi-modal data through iterative refinement.Extensive validation on the OUCR maritime dataset and public benchmarks reveals that SGDF achieves higher MIOU and VIF than the evaluated baseline methods, with improvements of 21.3% in MIOU over Li et al. (2026) and 28.7% in VIF over Text-DiFuse (2024).