Scholarlink

<p>Ship orientation angle prediction (SOAP) is crucial in optical remote sensing image processing, often relying on deep convolutional neural networks (CNNs) for accuracy. This paper introduces an innovative framework that reduces SOAP’s inference cost without sacrificing precision. We first present Mobile-SOAP, a lightweight, advanced model based on MobileNetV2. Additionally, we introduce four more compact SOAP models, each substituting Mobile-SOAP’s convolutional blocks or fully-connected layers with smaller modules, albeit with a noticeable performance drop. To enhance their performance, we propose SOAP-KD, a novel knowledge distillation (KD) framework, using Mobile-SOAP as the teacher. SOAP-KD includes a feature-based guidance loss and a synthetic sample-based knowledge transfer mechanism. Experiments on the FGSC-23 dataset validate Mobile-SOAP’s superiority and SOAP-KD’s effectiveness in boosting the performance of the four compact models. Our codes are available at <a href="https://github.com/UBCDingXin/SOAP-KD">https://github.com/UBCDingXin/SOAP-KD</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

Regression-oriented knowledge distillation for lightweight ship orientation angle prediction with optical remote sensing images

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Zhan Shi,
  • Xin Ding,
  • Peng Ding,
  • Chun Yang,
  • Ru Huang,
  • Xiaoxuan Song

摘要

Ship orientation angle prediction (SOAP) is crucial in optical remote sensing image processing, often relying on deep convolutional neural networks (CNNs) for accuracy. This paper introduces an innovative framework that reduces SOAP’s inference cost without sacrificing precision. We first present Mobile-SOAP, a lightweight, advanced model based on MobileNetV2. Additionally, we introduce four more compact SOAP models, each substituting Mobile-SOAP’s convolutional blocks or fully-connected layers with smaller modules, albeit with a noticeable performance drop. To enhance their performance, we propose SOAP-KD, a novel knowledge distillation (KD) framework, using Mobile-SOAP as the teacher. SOAP-KD includes a feature-based guidance loss and a synthetic sample-based knowledge transfer mechanism. Experiments on the FGSC-23 dataset validate Mobile-SOAP’s superiority and SOAP-KD’s effectiveness in boosting the performance of the four compact models. Our codes are available at https://github.com/UBCDingXin/SOAP-KD.