ELFT: Efficient local-global fusion transformer for small object detection.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Transformer models excel in computer vision but struggle with small object detection due to insufficient feature extraction.
• Existing methods face deployment challenges on resource-constrained platforms due to high computational demands.

Purpose of the Study:

• To propose an Efficient Local-Global Fusion Transformer (ELFT) for effective and efficient small object detection.
• To address the limitations of current transformer-based approaches in feature representation and computational complexity.

Main Methods:

• Developed an Efficient Local-Global Fusion Attention (ELGFA) mechanism for enhanced location feature extraction and integration of detailed feature map information.
• Introduced a Grouped Feature Update Module (GFUM) to decrease computational complexity by alternating updates of high-level and low-level features within groups.
• Incorporated a Broadcast Context module (CB) to enrich contextual information for improved small object identification.

Main Results:

• Achieved high mean average precision (mAP) scores: 95.8% on Remote Sensing Object Detection (RSOD), 94.3% on NWPU VHR-10, and 85.2% on PASCAL VOC2007.
• Demonstrated significant reductions in computational burden compared to DINO: 10.4% fewer parameters and 22.7% fewer Floating Point Operations (FLOPs).

Conclusions:

• The proposed ELFT effectively enhances small object detection accuracy.
• ELFT offers a computationally efficient solution suitable for resource-constrained environments, outperforming existing methods.