基于元学习的甲骨文拓片识别研究
Research on Oracle Bone Inscription Recognition Based on Meta-Learning
卢凡 1赵宇明1
作者信息
- 1. 上海交通大学电子信息与电气工程学院,上海 201100
- 折叠
摘要
为提高甲骨文拓片的识别效率,针对甲骨文拓片识别过程中存在的数据集种类繁多但类内样本过少的问题,将元学习引入甲骨文拓片图像的识别,提出一种基于元学习的甲骨文拓片识别算法.首先,选择残差网络(ResNet)18作为基本网络结构,以更好地提取甲骨文数据集特征.然后,通过元学习方法对初始模型参数进行学习.试验结果表明,该算法学习到的初始模型参数对于学习新类别的识别有着很好的效果,优于与模型无关的元学习(MAML)等其他模型,并且对于少样本的甲骨文数据集的识别十分有效.该研究为其他少样本数据集的处理和识别提供了一种解决的思路.
Abstract
In order to improve the recognition efficiency of oracle bone inscription,for the problem that there are many kinds of datasets but too few samples within the class in the process of oracle bone inscription recognition,meta-learning is introduced into the recognition of oracle bone inscription images,and a meta-learning-based oracle bone inscription recognition algorithm is proposed.Firstly,the residual network(ResNet)18 is chosen as the basic network structure,which can realize better feature extraction for oracle bone dataset.Then,the initial model parameters are learned by the meta-learning method.The experimental results show that the initial model parameters learned by this algorithm have a good effect on learning new categories for recognition,which is better than other models such as the model-agnostic meta-learning(MAML),etc.,and it is very effective for the recognition of the oracle bone dataset with few samples.This research provides a solution idea for processing and recognizing other less sample datasets.
关键词
甲骨文拓片分类/深度学习/元学习/残差网络/卷积神经网络/与模型无关的元学习算法Key words
Oracle bone inscription classification/Deep learning/Meta-learning/Residual network(ResNet)/Convolutional neural network/Model-agnostic meta-learning(MAML)algorithm引用本文复制引用
出版年
2024
NETL
NSTL
万方数据