Scholarlink

<p>Handwritten mathematical expression recognition (HMER) plays a crucial role in intelligent education and document automation. However, existing methods often struggle with symbol ambiguities, stroke irregularities, and visually similar patterns. To overcome these challenges, we introduce WaveMER, a novel framework that leverages discrete wavelet transform (DWT) to explicitly introduce frequency-domain information into the recognition pipeline. WaveMER couples a DenseNet-based spatial encoder with a DWT-derived frequency pathway, in which FERB produces frequency features and WFA highlights salient channels. An adaptive cross-domain fusion block then aligns and fuses the spatial and frequency streams, enabling the model to capture both global structural patterns and fine-grained frequency cues. Comprehensive experiments on the CROHME 2014, 2016, and 2019 benchmarks demonstrate the superiority of WaveMER, achieving expression recognition rates of 60.55%, 60.68%, and 62.38%, respectively. Here, we show that incorporating frequency-domain information significantly enhances HMER performance, offering a robust solution for accurate mathematical expression recognition. The source code is available at <a href="https://github.com/Ethereal-dot/WaveMER">https://github.com/Ethereal-dot/WaveMER</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

Enhanced handwritten mathematical expression recognition via wavelet feature integration

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Pin Wu,
  • Yubo Wang,
  • Yiguo Yang

摘要

Handwritten mathematical expression recognition (HMER) plays a crucial role in intelligent education and document automation. However, existing methods often struggle with symbol ambiguities, stroke irregularities, and visually similar patterns. To overcome these challenges, we introduce WaveMER, a novel framework that leverages discrete wavelet transform (DWT) to explicitly introduce frequency-domain information into the recognition pipeline. WaveMER couples a DenseNet-based spatial encoder with a DWT-derived frequency pathway, in which FERB produces frequency features and WFA highlights salient channels. An adaptive cross-domain fusion block then aligns and fuses the spatial and frequency streams, enabling the model to capture both global structural patterns and fine-grained frequency cues. Comprehensive experiments on the CROHME 2014, 2016, and 2019 benchmarks demonstrate the superiority of WaveMER, achieving expression recognition rates of 60.55%, 60.68%, and 62.38%, respectively. Here, we show that incorporating frequency-domain information significantly enhances HMER performance, offering a robust solution for accurate mathematical expression recognition. The source code is available at https://github.com/Ethereal-dot/WaveMER.