改进ResNeSt网络的拓片甲骨文字识别

PDF下载HTML

Title:

Rubbing oracle bone character recognition based on improved ResNeSt network

作者:

毛亚菲¹, 毕晓君²

1. 哈尔滨工程大学 信息与通信工程学院, 黑龙江 哈尔滨 150001;
2. 中央民族大学 信息工程学院, 北京 100081

Author(s):

MAO Yafei¹, BI Xiaojun²

1. College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;
2. Department of Information Engineering, Minzu University of China, Beijing 100081, China

关键词:

ResNeSt网络模型; 甲骨文字识别; 跳转连接; 坐标注意力机制; 分类器优化; OBC306; 深度学习; 神经网络

Keywords:

ResNeSt network model; oracle bone character recognition; jump connection; coordinate attention mechanism; optimization of classifier; OBC306; deep learning; neural network

分类号:

TP391

DOI:

10.11992/tis.202211041

摘要:

目前，拓片甲骨文字的识别方法存在局部细节特征提取能力弱，对部分高相似度的甲骨文字识别率较低的问题。为此，本文提出了一种基于改进ResNeSt网络的甲骨文字识别方法，通过设计跳转连接结构，逐步将网络浅层特征向网络深层传递并进行融合；同时结合甲骨文字“长条形”的特点，引入坐标注意力机制模块，从宽度和高度两个方向上对所得特征进行加权融合；最后通过去掉网络最后一层的激活函数和全连接层以及对最后一个卷积层输出通道数的重新设置，对网络分类器进行了有效优化。实验结果表明，本文提出的改进拓片甲骨文字识别模型在OBC306数据集上识别准确率达到93.53%，取得了目前最好的识别效果。

Abstract:

At present, the methods for recognition of rubbing oracle bone characters have some problems, such as weak ability to extract local details and low recognition rate for some oracle characters with highly similarity. Therefore, an oracle bone character recognition method based on the improved ResNeSt network is proposed in this paper. By designing jump connection structure, the shallow network features are gradually transferred to the deep network and merged. At the same time, combined with the Strip characteristics of the oracle bone characters, the coordinate attention mechanism module is introduced to carry out the weighted fusion of the obtained features in both width and height directions. Finally, the network classifier is optimized effectively by removing the activation function of the last layer and the full connection layer, and resetting the number of output channels of the last convolution layer. The experimental results show that the recognition accuracy of the improved rubbing oracle bone character recognition model proposed in this paper reaches 93.53% on OBC306 data set, having achieved the best recognition effect at present.