LFODet: Lightweight Few-Shot Object Detection with Meta-Learning in Remote Sensing Images

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Object detection in remote sensing faces challenges balancing accuracy and model size.
• Convolutional neural networks often require large datasets, hindering few-shot detection of novel classes.
• Adapting models to new classes with limited samples is crucial for practical applications.

Purpose of the Study:

• To propose LFODet, a lightweight few-shot object detection network using meta-learning.
• To enhance feature representation for improved semantic discriminability and spatial detail preservation.
• To reduce computational complexity for resource-constrained remote sensing applications.

Main Methods:

• Developed a lightweight few-shot object detection network (LFODet) based on meta-learning.
• Integrated Semantic Ghost Channel Attention (GCA) and Fine-Grained Ghost Spatial Attention (GSA).
• Utilized Ghost convolutions and a three-stage training process: base-class pre-training, meta-learner optimization, and few-shot fine-tuning.

Main Results:

• LFODet demonstrated stable and balanced performance across various few-shot learning scenarios.
• Achieved efficient feature representation by enhancing semantic and spatial details.
• Validated on benchmark datasets (DIOR and NWPU VHR-10).

Conclusions:

• LFODet offers a practical solution for resource-constrained remote sensing.
• The network enables rapid adaptation to new targets with limited samples.
• Successfully balances detection accuracy with model lightweightness.