EMe-DETR: an efficient and multi-scale enhanced solution for multi-object detection in optical remote sensing images

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Object detection in optical remote sensing faces challenges with small and densely packed objects.
• Vision Transformers (ViTs) show promise but struggle with fine-grained feature extraction for small objects.
• Existing networks often compromise accuracy for efficiency.

Purpose of the Study:

• To develop an efficient and accurate object detection network for optical remote sensing images.
• To address the challenges of small object detection and dense object overlap.
• To balance recognition accuracy with model complexity.

Main Methods:

• Proposes EMe-DETR, an enhanced network based on the real-time detection transformer (RT-DETR).
• Introduces an efficient feature-aware interaction module (EFIM) for improved feature extraction.
• Implements a lightweight multi-scale enhancement pyramid to maintain performance with reduced complexity.

Main Results:

• EMe-DETR achieved 88.43% mAP on the DIOR dataset and 94.90% mAP on the HRRSD dataset.
• The network demonstrates effective feature extraction from complex backgrounds.
• The proposed methods successfully balance accuracy and computational efficiency.

Conclusions:

• EMe-DETR offers a robust and innovative solution for object detection in optical remote sensing.
• The network effectively handles small objects and dense scenes.
• The efficient design makes it suitable for practical applications requiring high accuracy and speed.