LDA-DETR: A lightweight dynamic attention-enhanced DETR for small object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Small object detection faces challenges like complex backgrounds and dense distributions.
• Current state-of-the-art detectors often compromise accuracy for efficiency or vice-versa.
• Deploying advanced algorithms in resource-constrained environments remains difficult.

Purpose of the Study:

• To propose a lightweight dynamic attention-enhanced DETR (LDA-DETR) for improved small object detection.
• To enhance the balance between accuracy and efficiency in object detection models.
• To address limitations in feature representation and fusion for small objects.

Main Methods:

• Designed a lightweight feature extraction backbone (LFEB) using residual structures and partial convolutions.
• Developed a Dynamic Multi-Scale Fusion Module (DMSFM) for adaptive feature fusion.
• Constructed an Attention-Enhanced Fusion Network (AEFN) to refine features with attention mechanisms and cascading operations.

Main Results:

• LDA-DETR demonstrates superior performance compared to state-of-the-art methods.
• The model achieves improved accuracy and efficiency, particularly for small object detection.
• Experiments on RSOD, NWPU VHR-10, URPC2020, and VisDrone-DET datasets validate the effectiveness.

Conclusions:

• LDA-DETR offers a promising solution for efficient and accurate small object detection.
• The proposed architecture effectively handles complex background interference and dense distributions.
• The method is well-suited for practical applications with limited computational resources.