Scholarlink

With the proliferation of portable devices in document digitization, complex deformations in scanned images challenge optical character recognition (OCR). Traditional model-driven methods suffer from limited generalization and high costs, while data-driven approaches require massive data and face edge-deployment hurdles due to bulky architectures. This paper introduces DocFEAt, a lightweight framework fusing hybrid feature engineering and dynamic attention for efficient document dewarping. The Multi-Form Feature Extractor (MFFE) employs hardware-accelerated parallel extraction of six geometric features to construct rich representations. The Dynamic Adaptive Perception (DAP) network uses a hierarchical attention mechanism—including spatial-channel recalibration and cross-scale interactions via a Document Multi-Scale Attention (DMA) module—to refine features with minimal parameters. Experiments on DocUNet and UVDoc benchmarks show DocFEAt achieves outstanding OCR accuracy: a 4.6% Character Error Rate (CER) on UVDoc (36.1% reduction vs. prior arts) and 17.9% CER on DocUNet, while maintaining strong geometric preservation (e.g., MS-SSIM=0.737). With only 6 million parameters, DocFEAt strikes a unique balance between performance and lightweight design.

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

DocFEAt: A Lightweight Model with Hybrid Feature Engineering and Dynamic Attention for Document Dewarping

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Weizhong Zhang,
  • Yingshan Shen,
  • Sizhu Wang,
  • Qi Chen,
  • Junhui Deng,
  • Jiahao Feng

摘要

With the proliferation of portable devices in document digitization, complex deformations in scanned images challenge optical character recognition (OCR). Traditional model-driven methods suffer from limited generalization and high costs, while data-driven approaches require massive data and face edge-deployment hurdles due to bulky architectures. This paper introduces DocFEAt, a lightweight framework fusing hybrid feature engineering and dynamic attention for efficient document dewarping. The Multi-Form Feature Extractor (MFFE) employs hardware-accelerated parallel extraction of six geometric features to construct rich representations. The Dynamic Adaptive Perception (DAP) network uses a hierarchical attention mechanism—including spatial-channel recalibration and cross-scale interactions via a Document Multi-Scale Attention (DMA) module—to refine features with minimal parameters. Experiments on DocUNet and UVDoc benchmarks show DocFEAt achieves outstanding OCR accuracy: a 4.6% Character Error Rate (CER) on UVDoc (36.1% reduction vs. prior arts) and 17.9% CER on DocUNet, while maintaining strong geometric preservation (e.g., MS-SSIM=0.737). With only 6 million parameters, DocFEAt strikes a unique balance between performance and lightweight design.