Scholarlink

<p>The problem of identifying the correct authors with the help of samples in handwriting is a bit compromising, considering the way variations change and image distortions come into being. Therefore, in this paper, we intend to improve the offline handwriting recognition by proposing a hybrid architecture that integrates deep learning and classical feature extraction methods. In particular, we extend the UNet model in a dual-path structure: one path is used for local features extracted by the neural network of UNet, and the second path is utilized to include global classical features that can capture invariant features of handwriting against usual deformations like rotation and scaling. This architecture has been tested on three major datasets of handwriting: QUWI, IFN/ENIT, and IAM for the recognition performance of writers. Our results indicate that the proposed hybrid method outperforms the basic UNet model and similar techniques, which achieved the highest recognition rates of 98.96% on the IFN/ENIT database, 96.25% on the IAM database, and 90.56% on the QUWI database. We conclude from this work that the integration of deep learning with traditional feature extraction raises the accuracy of handwriting recognition, provides robustness to normal variations in handwriting, and sets new benchmark standards for author identification.</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 handwriting recognition through hybrid UNet-based architecture with global classical features

  • 期刊论文

  • 发表日期：

    2025年1月18日

查看全文
  • Xiaofei Liu

摘要

The problem of identifying the correct authors with the help of samples in handwriting is a bit compromising, considering the way variations change and image distortions come into being. Therefore, in this paper, we intend to improve the offline handwriting recognition by proposing a hybrid architecture that integrates deep learning and classical feature extraction methods. In particular, we extend the UNet model in a dual-path structure: one path is used for local features extracted by the neural network of UNet, and the second path is utilized to include global classical features that can capture invariant features of handwriting against usual deformations like rotation and scaling. This architecture has been tested on three major datasets of handwriting: QUWI, IFN/ENIT, and IAM for the recognition performance of writers. Our results indicate that the proposed hybrid method outperforms the basic UNet model and similar techniques, which achieved the highest recognition rates of 98.96% on the IFN/ENIT database, 96.25% on the IAM database, and 90.56% on the QUWI database. We conclude from this work that the integration of deep learning with traditional feature extraction raises the accuracy of handwriting recognition, provides robustness to normal variations in handwriting, and sets new benchmark standards for author identification.