Scholarlink

<p>Chip packaging drawings are an important component of traditional design documents. In the transition from traditional to digital documentation, the automatic detection of annotated text becomes a critical step, which raises higher requirements for detection accuracy and efficiency. To address this, this paper proposes a lightweight text detector, ACIMNet. The method utilizes ResNet18_vd as the backbone network to enhance feature extraction capabilities, effectively capturing complex patterns and fine-grained features. Furthermore, a Dilated Channel Interaction (DCI) module is designed to enable synergistic interaction between channel and spatial dimensions, thereby improving the sensitivity of feature selection. During the feature fusion phase, a Multidimensional Scale Fusion (MDSF) module is introduced to further enhance the network’s scale robustness. A dataset consisting of 1525 manually annotated chip packaging drawings is constructed for evaluation. Experimental results show that ACIMNet achieves 95.1% precision and 90.1% recall, while maintaining high inference speed. Compared to most existing text detectors, ACIMNet demonstrates superior performance, validating the effectiveness of the proposed method.</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

ACIMNet: a text detection method for chip package drawings based on dilated channel interaction and multidimensional scale fusion

  • 期刊论文

  • 发表日期：

    2025年6月3日

查看全文
  • Kaixin Liu,
  • Guofu Feng,
  • Ming Chen

摘要

Chip packaging drawings are an important component of traditional design documents. In the transition from traditional to digital documentation, the automatic detection of annotated text becomes a critical step, which raises higher requirements for detection accuracy and efficiency. To address this, this paper proposes a lightweight text detector, ACIMNet. The method utilizes ResNet18_vd as the backbone network to enhance feature extraction capabilities, effectively capturing complex patterns and fine-grained features. Furthermore, a Dilated Channel Interaction (DCI) module is designed to enable synergistic interaction between channel and spatial dimensions, thereby improving the sensitivity of feature selection. During the feature fusion phase, a Multidimensional Scale Fusion (MDSF) module is introduced to further enhance the network’s scale robustness. A dataset consisting of 1525 manually annotated chip packaging drawings is constructed for evaluation. Experimental results show that ACIMNet achieves 95.1% precision and 90.1% recall, while maintaining high inference speed. Compared to most existing text detectors, ACIMNet demonstrates superior performance, validating the effectiveness of the proposed method.