BMS-MFZS: A Background Modeling and Suppression Method Based on Multifeature Generalized Zero-Shot in Remote Sensing

Reliable background suppression remains a key challenge in infrared imaging for space and aerial scientific visual learning, as image data are often affected by multiple complex and intertwined factors, such as environmental variability, radiometric inconsistencies, and sensor noise, which make it difficult for models to effectively learn background characteristics for robust suppression. To address this, we propose a background modeling and suppression method based on multifeature generalized zero-shot learning (BMS-MFZS), which innovatively introduces generalized zero-shot learning (GZSL) into background modeling. By leveraging both target features and seen background features, the model reversely infers background representations to suppress the impact of complex backgrounds on recognition. Moreover, this work is the first to incorporate the physical characteristics of both targets and backgrounds as auxiliary features within the GZSL framework, thereby enhancing feature effectiveness and improving the model’s target–background discrimination capability. Finally, a joint semantic–pixel background reconstruction strategy is introduced to enable effective suppression of complex infrared backgrounds by combining global structural modeling with fine-detail refinement. The effectiveness of the proposed BMS-MFZS model is validated through extensive experiments on both real, semiphysical simulation and simulated datasets, demonstrating strong adaptability and suppression performance across multiple complex scenarios.