Dynamic small object feature enhancement and detection for remote sensing images.

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Remote sensing images present challenges like small objects and cluttered backgrounds, hindering accurate and efficient object detection on edge devices.
• Existing models struggle to balance recognition accuracy with the computational efficiency required for embedded systems.
• Real-time detection is crucial for applications like disaster response and traffic monitoring.

Purpose of the Study:

• To develop a lightweight, dynamic feature-enhanced detector (DFE-DETR) optimized for remote sensing applications.
• To improve the balance between accuracy and efficiency for object detection in resource-constrained environments.
• To enhance the detection of small, slender, and irregular objects in complex remote sensing scenes.

Main Methods:

• Proposed DFE-DETR, a refined RT-DETR model incorporating three novel modules: Sparse Attention Enhancement Module (SAEM), Multi-scale Convolutional Attention Enhancement Module (MSCAEM), and Deformable Large Kernel Convolution Module (DLKCM).
• SAEM prunes irrelevant tokens using Top-k scoring to reduce computation.
• MSCAEM utilizes multi-branch depth-wise strip convolutions to enhance detection of slender targets.
• DLKCM employs adaptive sampling for locating irregular objects.

Main Results:

• DFE-DETR achieved 47.34% mAP@0.5 and 65.8 FPS on VisDrone2019 with only 23.6M parameters.
• Attained 82.79% mAP@0.5 and 74.6 FPS on the SIMD dataset.
• Achieved high mAP@0.5 scores of 96.58% on RSOD and 93.28% on NWPU VHR-10.
• Outperformed mainstream YOLO series and specialized UAV detection models in comparative tests.

Conclusions:

• DFE-DETR offers a significant performance breakthrough for remote sensing object detection, especially on embedded devices.
• The model's synergistic modules effectively enhance target perception and contour delineation.
• DFE-DETR provides a novel high-precision, real-time detection solution for critical applications like disaster response and traffic monitoring.