SPMF: a saliency-based pseudo-multimodality fusion model for data-scarce maritime targets detection

• A seawater-tailored method generates pseudo-infrared imagery from RGB. • A spectral smoothing method prevents aliasing for sparse maritime targets. • Saliency maps guide a dual-backbone network for feature fusion. • Small target detection is enhanced in data-scarce scenarios. Despite the rapid advancement of visual detection, challenges remain in detecting maritime targets due to insufficient foreground-background separation and scarce data; this paper addresses these issues by proposing a Saliency-based Pseudo-Multimodality Fusion detection model (SPMF). Firstly, a spectral enhancement is employed based on ocean spectral absorption preferences, achieving excellent target retention performance. Subsequently, spectrum smooth reconstruction is integrated, yielding high-contrast saliency maps with good anti-aliasing properties. Finally, the salient map, simulating infrared-like features derived from RGB inputs, is incorporated as a pseudo-infrared modality into a cross-modality, cross-level dual backbone network. The experiments were conducted on the SeaDronesSee dataset. Compared to the baseline model, the mAP of SPMF has increased to 53.9 % (+2.0 %), with the PR-AUC rising to 0.879 (+0.02). Additionally, for target categories with small-sample and small-size features, such as Life Saving Appliances (LSA) and Buoys, AP has increased respectively to 38.6 % (+3.0 %) and 56.8 % (+4.6 %); while for Swimmers, Boats, and LSA, APs were the highest among all compared models. Furthermore, experiments on the AFO dataset demonstrate SPMF’s consistent superiority over the baseline. These demonstrate SPMF’s capability to maximally mine target features from single-modality by simulating infrared characteristics, particularly excelling in small-sample, small-size scenarios while reducing reliance on specialized sensors.