计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (3): 182-188.DOI: 10.3778/j.issn.1002-8331.1911-0117

• 模式识别与人工智能 • 上一篇    下一篇

Trans-Net:基于迁移学习的手写简笔画识别

姚可欣,曹卫群   

  1. 北京林业大学 信息学院,北京 100087
  • 出版日期:2021-02-01 发布日期:2021-01-29

Trans-Net:Stick Figure Recognition Based on Transfer Learning

YAO Kexin, CAO Weiqun   

  1. School of Information Science and Technology, Beijing Forestry University, Beijing 100087, China
  • Online:2021-02-01 Published:2021-01-29

摘要:

手写输入可通过少量的书写进而传递丰富的文本信息,如何准确地对手写简笔画进行识别越来越引起了各界研究者们的关注。传统的简笔画识别算法多基于简笔画相对固定的结构特性进行识别。此种方法对于笔迹清晰、结构相对简单的简笔画具有较高的识别率,但是随着分类数以及简笔画自身结构复杂度的增加这种方法存在一定局限性,往往会造成误分类。为取得更好的识别效果,该研究以具有固定参照模板的简笔画作为研究对象,使用图像生成算法对手写笔迹进行预处理,并提出了一种基于卷积神经网络的简笔画识别模型(Trans-Net),其中运用迁移学习技术解决了样本库中数据量小的问题。实验结果表明,该方法能够对输入的简笔画笔迹进行有效地特征提取,并且对样本库中150类简笔画对象的平均识别精度达到了94.1%。

关键词: 简笔画识别, 卷积神经网络, 图像生成, 迁移学习

Abstract:

Handwritten input can transfer rich text information through a small amount of writing. How to accurately recognize hand written stick figure has attracted more and more attention of researchers from all walks of life. The traditional recognition algorithm of stick figure is based on the relatively fixed structure characteristics of stick figure. This method has a high recognition accuracy for simple stick figures with clear handwriting and relatively simple structures. However, with the increase of the object of classifications and the complexity of the structure of the stick figure, such methods have some limitations, which often cause misclassification. In order to obtain a better recognition effect, this paper takes stick figure with a corresponding template as the research object. The image generation algorithm is used to preprocess the handwriting input, and a simple recognition model(Trans-Net) based on convolutional neural network is proposed to recognize. Particularly, the use of transfer learning solves the deficient of dataset in the sample library. Experimental results show that the proposed method can effectively extract the features of the handwriting input, and the average recognition accuracy of 150 types of stick figure objects in the sample library reaches 94.1%.

Key words: stick figure recognition, convolutional neural network, figure generation, transfer learning