SMESP-DiffDet: Small-Object MoE Enhanced Spatial Prior-Guided DiffDet for Visible-Infrared Object Detection
Abstract
Small-object detection remains highly challenging due to their small size, weak texture, and severe environmental interference, even when both visible and infrared modalities are used, especially in remote sensing images. To address this issue, this article proposes small-object MoE-enhanced spatial prior-guided DiffDet (SMESP-DiffDet) for visible–infrared object detection, which incorporates two key components: the small-object MoE enhancement (SME) module and the spatial prior-guided box renewal strategy (SP-BRS). Specifically, unlike existing mixture-of-expert (MoE) models that rely on implicit experts, the SME module introduces dedicated experts for small-object detection to enhance localization and classification accuracy by capturing more discriminative feature representations of small objects. This process is further guided by a small-object constraint loss. Based on this, instead of supplementing random proposal boxes during diffusion decoding, the proposed SP-BRS leverages the small-object spatial priors learned in SME to guide the candidate box renewal process, providing potential small-object regions for decoding and thus improving detection accuracy and efficiency. Extensive experiments on three representative public datasets demonstrate that SMESP-DiffDet significantly improves visible–infrared small-object detection performance, achieving up to approximately 2.5% improvement in mAP50 compared with state-of-the-art methods.