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

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Aerial remote sensing images present challenges for small object detection due to scale, texture, and background interference.
• Existing methods struggle with fine-grained feature representation and long-range dependency modeling for small objects.

Purpose of the Study:

• To propose MambaIR-YOLO, a feature-guided lightweight state-space framework for effective aerial small-object detection.
• To improve the representation of small objects by enhancing fine-grained feature modeling, long-range dependency learning, and high-resolution spatial information preservation.

Main Methods:

• Introduced a feature-level MambaIR_SR (feature-level Mamba-based super-resolution guidance) training auxiliary branch for detail-guided information fusion.
• Developed the Object Detail State-Space Block (ODSSBlock) using Lightweight Mamba for unified local detail and long-range context modeling.
• Incorporated a Feature Modulation Block (FMB) for high-frequency structural information enhancement and a Coordinate and Channel Attention (C3CA) module.
• Utilized a Lightweight Decoupled Head (LADH) for single-scale dense prediction on high-resolution features.

Main Results:

• Achieved an average mAP50 of 84.19 ± 0.03% on the VEDAI dataset.
• Maintained a low computational cost with only 4.46 M parameters and 19.97 GFLOPs.
• Outperformed baseline methods in small-object detection accuracy while ensuring real-time performance.

Conclusions:

• MambaIR-YOLO effectively addresses the challenges of aerial small-object detection.
• The proposed framework offers a lightweight and efficient solution with superior performance.
• The MambaIR_SR branch, used only during training, adds no inference overhead.