Scholarlink

<p>Image cropping plays a crucial role in various computer vision applications. This paper presents a novel approach for high-resolution image cropping, aiming to generate multiple aesthetically pleasing regions. In this paper, we propose an efficient model incorporating an attention module and feature extraction module at different scales. Unlike traditional approaches relying solely on manually annotated crops, our approach combines machine-generated annotation with a limited number of manual annotations, applying the annotation score alignment method to construct the training dataset. Specially, the proposed cropping model utilizes a multi-scale channel attention module to enhance the selection of visually appealing regions. Both subjective and objective experiments conducted on datasets demonstrate the effectiveness of our approach in producing diverse and visually appealing crops. The code is available at:<a href="https://github.com/CunHua-YYT/MRC">https://github.com/CunHua-YYT/MRC</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

Hybrid annotation alignment-based multi-region crop model for high-resolution image

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Yuetao Yuan,
  • Shuchang Xu,
  • Junjie Cheng,
  • Shudong Lin

摘要

Image cropping plays a crucial role in various computer vision applications. This paper presents a novel approach for high-resolution image cropping, aiming to generate multiple aesthetically pleasing regions. In this paper, we propose an efficient model incorporating an attention module and feature extraction module at different scales. Unlike traditional approaches relying solely on manually annotated crops, our approach combines machine-generated annotation with a limited number of manual annotations, applying the annotation score alignment method to construct the training dataset. Specially, the proposed cropping model utilizes a multi-scale channel attention module to enhance the selection of visually appealing regions. Both subjective and objective experiments conducted on datasets demonstrate the effectiveness of our approach in producing diverse and visually appealing crops. The code is available at:https://github.com/CunHua-YYT/MRC.