Few-Shot Object Detection in Remote Sensing Images via Data Clearing and Stationary Meta-Learning.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Few-shot object detection (FSOD) faces challenges due to limited remote sensing data.
• Remote sensing images (RSIs) present unique difficulties for object detection.
• Existing methods struggle with incomplete annotations and feature vector variance.

Purpose of the Study:

• To develop a meta-learning-based object detector for RSIs with limited data.
• To enhance the accuracy and robustness of few-shot object detection in remote sensing.
• To address data imbalance and feature representation issues in FSOD.

Main Methods:

• A Balanced Few-Shot Object Detector (B-FSDet) built on YOLOv9 (GELAN-C).
• Data clearing strategy for balanced category input.
• Stationary feature extraction and prediction for stable meta-learning.
• Inter-class discrimination support loss for improved class separability.

Main Results:

• The B-FSDet demonstrates promising performance on DIOR and NWPU VHR-10.v2 datasets.
• Comparative analysis shows superior results against state-of-the-art detectors.
• The proposed methods effectively handle incomplete annotations and feature variance.

Conclusions:

• The meta-learning-based B-FSDet significantly advances FSOD for remote sensing.
• The novel strategies improve data balance, feature representation, and inter-class discrimination.
• This work offers a robust solution for object detection with scarce data in RSIs.