SFIDM: Few-Shot Object Detection in Remote Sensing Images with Spatial-Frequency Interaction and Distribution Matching DOI Creative Commons
Yong Wang, Jingtao Li, Guo Jiahui

et al.

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(6), P. 972 - 972

Published: March 10, 2025

Few-shot object detection (FSOD) in remote sensing images (RSIs) faces challenges such as data scarcity, difficulty detecting small objects, and underutilization of frequency-domain information. Existing methods often rely on spatial-domain features, neglecting the complementary insights from low- high-frequency characteristics. Additionally, their performance objects is hindered by inadequate feature extraction cluttered backgrounds. To tackle these problems, we propose a novel framework Spatial-Frequency Interaction Distribution Matching (SFIDM), which significantly enhances FSOD RSIs. SFIDM focuses rapid adaptation to target datasets efficient fine-tuning with limited data. First, improve representation, introduce (SFI) module, leverages complementarity between low-frequency By decomposing input into frequency components, SFI module extracts features critical for classification precise localization, enabling capture fine details essential objects. Secondly, resolve limitations traditional label assignment strategies when dealing bounding boxes, construct (DM) models boxes 2D Gaussian distributions. This allows accurate subtle offsets overlapping or non-overlapping Moreover, leverage learned base-class information improved class detection, employ reweighting adaptively fuses extracted backbone network generate representations better suited downstream tasks. We conducted extensive experiments two benchmark demonstrate effectiveness improvements achieved proposed framework.

Language: Английский

SFIDM: Few-Shot Object Detection in Remote Sensing Images with Spatial-Frequency Interaction and Distribution Matching DOI Creative Commons
Yong Wang, Jingtao Li, Guo Jiahui

et al.

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(6), P. 972 - 972

Published: March 10, 2025

Few-shot object detection (FSOD) in remote sensing images (RSIs) faces challenges such as data scarcity, difficulty detecting small objects, and underutilization of frequency-domain information. Existing methods often rely on spatial-domain features, neglecting the complementary insights from low- high-frequency characteristics. Additionally, their performance objects is hindered by inadequate feature extraction cluttered backgrounds. To tackle these problems, we propose a novel framework Spatial-Frequency Interaction Distribution Matching (SFIDM), which significantly enhances FSOD RSIs. SFIDM focuses rapid adaptation to target datasets efficient fine-tuning with limited data. First, improve representation, introduce (SFI) module, leverages complementarity between low-frequency By decomposing input into frequency components, SFI module extracts features critical for classification precise localization, enabling capture fine details essential objects. Secondly, resolve limitations traditional label assignment strategies when dealing bounding boxes, construct (DM) models boxes 2D Gaussian distributions. This allows accurate subtle offsets overlapping or non-overlapping Moreover, leverage learned base-class information improved class detection, employ reweighting adaptively fuses extracted backbone network generate representations better suited downstream tasks. We conducted extensive experiments two benchmark demonstrate effectiveness improvements achieved proposed framework.

Language: Английский

Citations

0