Super Mamba feature enhancement framework for small object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate and timely detection of small objects (dozens of pixels) in infrared images, especially from low-altitude drones with complex backgrounds, remains a significant challenge.
• Existing methods often incur substantial computational costs when learning strong feature representations to separate small objects from intricate backgrounds.

Purpose of the Study:

• To propose the Super Mamba (SMamba) framework for enhanced unmanned aerial vehicle (UAV) infrared small object detection.
• To achieve high-resolution detection of multi-scale objects while balancing accuracy and computational efficiency.

Main Methods:

• Incorporated Receptive Field Attention Convolution (RFAConv) into the backbone network to optimize computing efficiency via dynamic receptive field adjustment.
• Integrated Spatial Attention Mechanism (SAM) and Squeeze-Excitation (SE) into the State Space Model (SSM) for multi-scale and multi-feature extraction.
• Introduced a Feature Enhancement Module (FEM) into the Bidirectional Feature Pyramid Network (BiFPN) neck to improve local context information and detection efficiency for small objects.

Main Results:

• The Super Mamba framework achieved over 92% accuracy (mAP@0.5) on the VEDAI dataset.
• Demonstrated a performance improvement exceeding 20% compared to state-of-the-art models like Yolov5, Yolov8, and Yolov11.
• The framework effectively handles multi-scale objects and complex backgrounds in infrared imagery.

Conclusions:

• The Super Mamba framework offers a significant advancement in UAV infrared small object detection.
• The proposed methods enhance feature representation and contextual understanding, leading to superior detection performance.
• The framework provides an efficient and accurate solution for challenging infrared small object detection tasks.