Scholarlink

<p>The detection of text instances in natural scenes is a critical task for numerous real-world applications, including augmented reality, content-based image retrieval, and autonomous navigation systems. However, this challenge becomes significantly difficult when text regions are partially obscured by other objects or background elements, a phenomenon known as occlusion. To resolve this issue, conventional approaches mostly rely on low-level visual features, which often falter in such complex scenarios, lacking the contextual understanding and semantic reasoning capabilities. This paper introduces a novel framework that harnesses the power of context-aware semantic understanding derived from sketch-level image representations. The proposed method comprises two key components: the sketch extraction and recovery module to remove the occlusion, and text detection from recovered sketches. The sketch extraction and recovery module plays a pivotal role in enabling the model to comprehend the structural and contextual nuances of both occluded and non-occluded regions within the sketch domain. By capturing the essential structural and geometric characteristics of text instances while suppressing irrelevant details and noise, this module generates abstracted representations that are inherently robust to occlusions and cluttered backgrounds. A deep learning-based detection network employs the powerful transformer architecture, to learn rich contextual representations that capture the intricate relationships between text instances, local surroundings, and the global scene context within the sketch domain. To validate the efficacy of our proposed approach, we conducted extensive evaluations on two challenging datasets: a custom-curated occluded scene text detection dataset and the widely-used Total Text benchmark. The results demonstrate the superior performance of our method under the occluded text detection scenarios, outperforming state-of-the-art techniques. Notably, our approach also attained competitive performance on the Total Text benchmark, further underscoring its versatility and effectiveness in both occluded and non-occluded settings.</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

Occluded scene text detection via context-awareness from sketch-level image representations

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Minh-Trieu Tran,
  • Guee-Sang Lee

摘要

The detection of text instances in natural scenes is a critical task for numerous real-world applications, including augmented reality, content-based image retrieval, and autonomous navigation systems. However, this challenge becomes significantly difficult when text regions are partially obscured by other objects or background elements, a phenomenon known as occlusion. To resolve this issue, conventional approaches mostly rely on low-level visual features, which often falter in such complex scenarios, lacking the contextual understanding and semantic reasoning capabilities. This paper introduces a novel framework that harnesses the power of context-aware semantic understanding derived from sketch-level image representations. The proposed method comprises two key components: the sketch extraction and recovery module to remove the occlusion, and text detection from recovered sketches. The sketch extraction and recovery module plays a pivotal role in enabling the model to comprehend the structural and contextual nuances of both occluded and non-occluded regions within the sketch domain. By capturing the essential structural and geometric characteristics of text instances while suppressing irrelevant details and noise, this module generates abstracted representations that are inherently robust to occlusions and cluttered backgrounds. A deep learning-based detection network employs the powerful transformer architecture, to learn rich contextual representations that capture the intricate relationships between text instances, local surroundings, and the global scene context within the sketch domain. To validate the efficacy of our proposed approach, we conducted extensive evaluations on two challenging datasets: a custom-curated occluded scene text detection dataset and the widely-used Total Text benchmark. The results demonstrate the superior performance of our method under the occluded text detection scenarios, outperforming state-of-the-art techniques. Notably, our approach also attained competitive performance on the Total Text benchmark, further underscoring its versatility and effectiveness in both occluded and non-occluded settings.