Multi path attention and scale aware fusion for accurate object detection in remote sensing imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Accurate object detection in remote sensing is vital for intelligent systems.
• Existing methods struggle with feature representation, semantic modeling, and multi-scale fusion, especially for small targets.
• Small objects are prone to omission and misclassification in current detection models.

Purpose of the Study:

• To introduce HyperFusion-DEIM, a novel cascaded detection paradigm.
• To enhance object-level representations, contextual semantic dependencies, and scale-aware feature integration.
• To address the limitations of current detectors in identifying small-scale targets in remote sensing imagery.

Main Methods:

• Developed the Multi-Path Attention Network (MAPNet) with Multi-Path Attention Fusion (MPAF) and Shallow Robust Feature Downsampling (SRFD) for small object recognition.
• Integrated the Scale-Aware Feature Enhancement (SAFE) encoder with Multi-level Feature Concentration (MFC) for cross-layer geometric alignment.
• Incorporated Transformer layers with HyperACE for long-range semantic correlation while preserving spatial fidelity.

Main Results:

• HyperFusion-DEIM achieved 64.5% AP on the SIMD benchmark, outperforming RT-DETR and DEIM.
• On the VEDAI benchmark, it surpassed YOLOv12, YOLOv13, RT-DETRv2, and DEIM.
• The model demonstrated real-time performance, reaching 296.33 FPS on SIMD and 79.7 FPS on VEDAI.

Conclusions:

• HyperFusion-DEIM significantly improves object detection accuracy and robustness in remote sensing.
• The proposed model effectively handles small-scale targets and complex feature fusion.
• HyperFusion-DEIM offers a practical solution for real-time detection in resource-constrained environments.