CCDN-DETR: A Detection Transformer Based on Constrained Contrast Denoising for Multi-Class Synthetic Aperture Radar...

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) are effective for SAR object detection, particularly for ship targets.
• Integrating transformer structures into SAR detectors enhances target localization, but existing methods do not fully utilize self-attention's long-range modeling.
• Multi-class SAR target detection remains an area with limited research.

Purpose of the Study:

• To propose a novel SAR detector, CCDN-DETR, based on the Detection Transformer (DETR) framework.
• To address limitations in existing SAR detectors, including fully leveraging self-attention and improving multi-class detection.
• To adapt transformer-based detectors for the multiscale characteristics of SAR data.

Main Methods:

• Developed CCDN-DETR, a SAR detector building upon the Detection Transformer (DETR) framework.
• Introduced cross-scale encoders to model and fuse information across different scales in SAR data.
• Optimized decoder input by employing IOU loss for object query initialization and incorporating constrained contrastive denoising training.

Main Results:

• CCDN-DETR achieved a mean Average Precision (mAP) of 91.9% on a combined SSDD, HRSID, and SAR-AIRcraft dataset.
• Demonstrated strong performance on the multi-class MSAR dataset with an 83.7% mAP, outperforming CNN-based models.
• The proposed methods enhanced model convergence speed and improved detection of diverse SAR target categories.

Conclusions:

• CCDN-DETR effectively leverages transformer architectures for improved SAR object detection, particularly for multi-class scenarios.
• The integration of cross-scale encoders and optimized query selection schemes addresses SAR data's multiscale nature.
• This research advances SAR target recognition by offering a more robust and accurate detection model.