Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (22): 151-157.DOI: 10.3778/j.issn.1002-8331.2208-0108

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Handwritten Character Recognition Model Based on Discriminant Convolutional Neural Network

QU Xiwen, WU Xiang, HU Mianjun, HUANG Jun   

  1. School of Computer Science and Technology, Anhui University of Technology, Maanshan, Anhui 243000, China
  • Online:2023-11-15 Published:2023-11-15



  1. 安徽工业大学 计算机科学与技术学院,安徽 马鞍山 243000

Abstract: At present, the deep learning models for online and offline handwritten character recognition are generally shallow in depth, and they are trained by using softmax loss function, so they lack the ability to learn discriminative information from different classes of samples. In order to build a deeper convolutional neural network model, a convolutional neural network model combining encoder and decoder is proposed, which can effectively increase the network depth and avoid the problem of training difficulties caused by the sharp increase of model parameters. For further improving the recognition accuracy, this paper presents a discriminative loss function. The loss function trains the neural network model by minimizing the distance between the feature of training sample extracted by neural network and the optimized prototype. Since the optimized prototype is learned by the minimization classification error algorithm, the constructed convolutional neural network model in this paper trained with the discriminant loss function can obtain strong discriminative ability. Comparative experiments are carried out on the in-air handwritten Chinese character dataset IAHCC-UCAS2016 and the offline handwritten character dataset MNIST. The results show that compared with softmax loss function, the recognition rate of discriminative loss function is improved by 1.04 percentage points on in-air handwritten dataset. The recognition rates of the proposed discriminant convolution neural network model on the two datasets are 95.53% and 99.73% respectively. Compared with the existing traditional models and depth network models for handwritten character recognition, the proposed model achieves higher recognition accuracy.

Key words: handwritten character recognition, deep learning, softmax loss function, optimized prototype, minimization classification error algorithm

摘要: 目前用于离线和联机手写字符识别的深度学习模型一般深度较浅,且使用softmax损失函数进行训练,缺乏从不同类样本中学习判别信息的能力。为构建更深层卷积神经网络模型,提出了一种编码器与解码器相结合的卷积神经网络模型,该模型可以有效增加网络深度,避免模型参数剧增导致训练困难的问题。为了进一步提高深度学习模型的识别精度,提出了一种判别损失函数,该损失函数通过最小化训练样本在全连接层的输出与该样本同类别优化原型之间的距离来训练神经网络模型。由于优化原型由最小化分类误差算法学习得到,使用判别损失函数训练的卷积神经网络模型具备了更强的判别能力。在公开的联机手写汉字数据集IAHCC-UCAS2016和离线手写字符数据集MNIST上进行对比实验,实验结果表明,在联机手写数据集上与softmax损失函数相比,判别损失函数的识别率提高了1.04个百分点。在两个数据集上,与现有的用于手写字符识别的传统模型和深度模型相比,提出的判别卷积神经网络模型识别率分别达到95.53%和99.73%,获得了更高的识别精度。

关键词: 手写字符识别, 深度学习, softmax损失函数, 优化原型, 最小化分类误差算法