SPGDet: Vision–Language Priors and Deformable Alignment for Multispectral Aerial Object Detection

Small objects in multispectral aerial images are difficult to distinguish from background clutter, particularly when visible and infrared features are spatially misaligned. We propose Semantic Prior Guided Detection (SPGDet) to combine scene context with adaptive cross-modal sampling. Prompt-guided frequency gated fusion conditions visible features on an RGB-derived scene description and combines them with gated sensor features. Deformable cross-modal feature alignment samples visible features using infrared-conditioned offsets. Spatial–channel prior attention combines these paths into shared memory for modality-specific query decoders, which predict boxes in each sensor’s coordinate frame. Training combines reliability-weighted semantic supervision and cross-modal class agreement with scale-aware localization and edge-aware offset regularization. On DroneVehicle, VTUAV-det, VEDAI512, and RGBTDronePerson, SPGDet achieves AP50 values of 82.6, 87.9, 85.7, and 75.8%, respectively. These scores exceed the strongest compared baseline by 1.4, 1.1, 1.1, and 1.3 percentage points. Component ablations and prompt controls identify contributions from both image-specific context and deformable sampling. Translation experiments further show that the alignment path improves detection under sensor displacement. These results demonstrate the value of combining scene descriptions with local sensor evidence for multispectral aerial detection.