Scholarlink

<p>With advancements in horizontal object detection models, oriented object detection models have also made significant strides. Yet existing rotated object detectors often struggle to maintain high accuracy while processing high-resolution remote sensing images in real-time. To address these challenges, we propose a new lightweight model specifically for oriented object detection, named the efficient cross-layer attention detector (ECDet). ECDet integrates several efficient modules, including an efficient reparameterized Transformer-like backbone (ERepViT) to reduce computational costs, and the efficient cross-layer fusion neck (CLF-Neck), a lightweight alternative to traditional pyramid networks for feature fusion with attention mechanism. Additionally, we introduce the lightweight task interaction decoupled (LTID) head, which enhances task-specific performance by providing more detailed, task-aligned information for classification and regression with minimal computational cost. Furthermore, an ensemble loss combined with the phase shifting coder (PSC) mitigates the angle discontinuity issue in regression-based methods. Evaluations on the DOTAv1 and HRSC datasets show that ECDet runs 32% faster than RTMDet-S with higher accuracy, demonstrating its strong potential for practical application. The source code will be release at <a href="https://github.com/tianlianghai/ECDet">https://github.com/tianlianghai/ECDet</a>.</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

ECDet: efficient oriented object detection on the aerial image with cross-layer attention

  • 期刊论文

  • 发表日期：

    2025年1月11日

查看全文
  • Xueqiang Lyu,
  • Lianghai Tian,
  • Shangzhi Teng

摘要

With advancements in horizontal object detection models, oriented object detection models have also made significant strides. Yet existing rotated object detectors often struggle to maintain high accuracy while processing high-resolution remote sensing images in real-time. To address these challenges, we propose a new lightweight model specifically for oriented object detection, named the efficient cross-layer attention detector (ECDet). ECDet integrates several efficient modules, including an efficient reparameterized Transformer-like backbone (ERepViT) to reduce computational costs, and the efficient cross-layer fusion neck (CLF-Neck), a lightweight alternative to traditional pyramid networks for feature fusion with attention mechanism. Additionally, we introduce the lightweight task interaction decoupled (LTID) head, which enhances task-specific performance by providing more detailed, task-aligned information for classification and regression with minimal computational cost. Furthermore, an ensemble loss combined with the phase shifting coder (PSC) mitigates the angle discontinuity issue in regression-based methods. Evaluations on the DOTAv1 and HRSC datasets show that ECDet runs 32% faster than RTMDet-S with higher accuracy, demonstrating its strong potential for practical application. The source code will be release at https://github.com/tianlianghai/ECDet.