MambaIR-YOLO: A Feature-Guided Lightweight State-Space Framework for Aerial Small-Object Detection

To address the challenges of extremely small object scales, weak texture information, and severe background interference in aerial remote sensing images, we propose MambaIR-YOLO, a feature-guided lightweight state-space framework for aerial small-object detection. Based on the YOLOv5 architecture, this method introduces systematic improvements in three key areas—fine-grained feature modeling, long-range dependency learning, and high-resolution spatial information preservation—while ensuring real-time performance. Specifically, a feature-level MambaIR_SR (feature-level Mamba-based super-resolution guidance) training auxiliary branch is designed to generate high-resolution detail-guided information by fusing shallow-level detail features with deep-level semantic features, improving the representation of small-object edges and textures during the training phase. In the main network, we introduce the Object Detail State-Space Block (ODSSBlock), driven by Lightweight Mamba. Through a channel-compressed state-space modeling mechanism, the ODSSBlock unifies local detail preservation and long-range context modeling with low computational overhead. Concurrently, a Feature Modulation Block (FMB) is constructed at the shallow feature level to enhance the representation of high-frequency structural information, thereby mitigating the irreversible detail degradation caused by multiple downsampling steps. In the feature fusion and detection stages, we introduce a Coordinate and Channel Attention (C3CA) attention enhancement module and construct a lightweight decoupled detection head based on a Lightweight Decoupled Head (LADH). This performs single-scale dense prediction on high-resolution features, improving small-object localization and reducing information loss. Notably, the MambaIR_SR branch is exclusively utilized for optimization during the training phase and is entirely discarded during inference, introducing no additional computational overhead. Experimental results on the VEDAI dataset demonstrate that the proposed method achieves an average mAP50 of 84.19 ± 0.03% over three independent runs, with only 4.46 M parameters and 19.97 GFLOPs, outperforming the baseline methods while maintaining low computational cost.