Scholarlink

<p>With the rapid advancement of remote sensing technologies, deep learning–based object detection has been widely adopted for large-scale optical remote sensing imagery. Despite the remarkable progress achieved by convolutional and transformer-based detectors, their performance remains significantly limited by the long-tailed distribution that is inherently present in real-world remote sensing datasets. Existing approaches address this issue primarily from data rebalancing or class-frequency perspectives, while largely overlooking the underlying optimization difficulty that governs the learning behaviour of tail and hard samples. To address this limitation, we propose a novel optimization-driven long-tailed object detection framework for optical remote sensing images, which explicitly regulates training dynamics rather than relying solely on statistical class distributions. The proposed method introduces an Optimization-Driven Feature Emphasis (ODFE) mechanism to adaptively highlight discriminative multi-scale features that are critical for hard and tail samples during optimization. In addition, an Optimization-Balanced Mosaic (OBM) augmentation strategy is designed to increase the effective exposure of optimization-tail samples, alleviating biased gradient updates caused by conventional class-agnostic augmentation. Furthermore, an Optimization-Aware Classification Loss (OACL) is formulated to reweight samples according to their optimization difficulty, ensuring stable convergence and balanced learning across head and tail categories. Extensive experiments conducted on three public optical remote sensing benchmarks, namely DIOR, DOTA, and FAIR1M, demonstrate that the proposed framework achieves competitive or superior performance compared with reproduced baselines evaluated under identical training settings; results from original papers are clearly annotated and interpreted with appropriate caution. In particular, notable performance gains are observed on small objects, background-cluttered categories, and fine-grained long-tailed classes. These results suggest that explicitly modeling optimization difficulty provides a more effective learning strategy for long-tailed object detection in optical remote sensing imagery, within the scope of the experimental settings considered.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimization-driven feature learning for long-tailed object detection in optical remote sensing imagery

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Nishi Madaan,
  • Rahul Malik,
  • Utsav Upadhyay,
  • Alok Kumar,
  • Satyabrata Roy,
  • Umashankar Rawat

摘要

With the rapid advancement of remote sensing technologies, deep learning–based object detection has been widely adopted for large-scale optical remote sensing imagery. Despite the remarkable progress achieved by convolutional and transformer-based detectors, their performance remains significantly limited by the long-tailed distribution that is inherently present in real-world remote sensing datasets. Existing approaches address this issue primarily from data rebalancing or class-frequency perspectives, while largely overlooking the underlying optimization difficulty that governs the learning behaviour of tail and hard samples. To address this limitation, we propose a novel optimization-driven long-tailed object detection framework for optical remote sensing images, which explicitly regulates training dynamics rather than relying solely on statistical class distributions. The proposed method introduces an Optimization-Driven Feature Emphasis (ODFE) mechanism to adaptively highlight discriminative multi-scale features that are critical for hard and tail samples during optimization. In addition, an Optimization-Balanced Mosaic (OBM) augmentation strategy is designed to increase the effective exposure of optimization-tail samples, alleviating biased gradient updates caused by conventional class-agnostic augmentation. Furthermore, an Optimization-Aware Classification Loss (OACL) is formulated to reweight samples according to their optimization difficulty, ensuring stable convergence and balanced learning across head and tail categories. Extensive experiments conducted on three public optical remote sensing benchmarks, namely DIOR, DOTA, and FAIR1M, demonstrate that the proposed framework achieves competitive or superior performance compared with reproduced baselines evaluated under identical training settings; results from original papers are clearly annotated and interpreted with appropriate caution. In particular, notable performance gains are observed on small objects, background-cluttered categories, and fine-grained long-tailed classes. These results suggest that explicitly modeling optimization difficulty provides a more effective learning strategy for long-tailed object detection in optical remote sensing imagery, within the scope of the experimental settings considered.