Task-Decoupled Learning Strategies for Optimized Multiclass Object Detection From VHR Optical Remote Sensing Imagery

Object detection in remote sensing imagery poses significant challenges due to the vast scale variations and long-tail class distributions inherent in these scenarios. Traditional object detectors often struggle to balance the conflicting requirements of translation invariance for classification and...

Full description

Saved in:
Bibliographic Details
Main Authors: Guangyao Zhou, Guanqun Wang, Yu Huang, Wenzhi Wang, Zhi Qu, Chenghan Yin
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10829776/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Object detection in remote sensing imagery poses significant challenges due to the vast scale variations and long-tail class distributions inherent in these scenarios. Traditional object detectors often struggle to balance the conflicting requirements of translation invariance for classification and translation equivariance for localization. In this article, we propose a decoupled multiclass object detection (DMOD) framework designed to address these challenges. The DMOD framework introduces a novel decoupling strategy, which separates the optimization of the localization and classification tasks, thereby avoiding the suboptimal solutions that arise from their conflicting demands. Specifically, the decoupled localization branch independently optimizes class-agnostic localization using a dedicated regression head, while the decoupled classification branch focuses on category prediction using translation-invariant features. In addition, a training-independent module (TIM) is incorporated to mitigate the impact of long-tail distributions by leveraging information theory to prioritize informative samples during training. Extensive experiments conducted on the DIOR and NWPU VHR-10 datasets demonstrate that our DMOD framework significantly outperforms existing state-of-the-art methods, achieving superior mean average precision scores of 74.79% and 95.23%, respectively. These results validate the effectiveness of the proposed decoupling strategy in enhancing object detection performance in complex remote sensing scenarios.
ISSN:1939-1404
2151-1535