MSRS-DETR: End-to-End Object Detection for Multi-Scale Remote Sensing.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Object detection in remote sensing imagery (RSI) is crucial for various applications.
• Current detectors struggle with spectral bias, limiting high-frequency information learning, which causes performance issues with cluttered backgrounds, distractors, and multi-scale targets, particularly small objects.

Purpose of the Study:

• To address the limitations of existing RSI object detection methods.
• To propose an end-to-end framework that integrates spatial and frequency domain analysis for enhanced object detection.
• To improve the detection of small and multi-scale objects in complex remote sensing scenes.

Main Methods:

• Developed MSRS-DETR, an end-to-end framework integrating spatial and frequency cues.
• Introduced C2f_FATNET, a frequency-attention-enhanced lightweight residual backbone for richer dual-domain features.
• Incorporated an Entanglement Transformer Block (ETB) for cross-domain frequency-spatial interaction and background interference suppression.
• Utilized S2-CCFF, a shallow-feature-extended bidirectional fusion path to enhance small object detail retention.

Main Results:

• MSRS-DETR demonstrated significant improvements over baseline methods on HRSC2016 and ShipRSImageNet datasets.
• Reduced model parameters by 29.1% and increased inference speed by 12.4% and 8.4%.
• Achieved a notable increase in mean Average Precision (mAP_50-95) of 1.69% and 2.16% respectively, indicating enhanced detection accuracy.

Conclusions:

• The proposed spatial-frequency fusion paradigm in MSRS-DETR effectively overcomes the limitations of traditional RSI object detectors.
• MSRS-DETR offers a more efficient and accurate solution for object detection in remote sensing, particularly for challenging scenarios involving small and cluttered objects.
• The framework shows strong effectiveness and generalizability across different remote sensing datasets.