基于元学习模型优化器的小样本手写西波文字识别

Meta-Learned Optimizer for Few-Shot Handwritten Xibo Characters Recognition

摘要: 本文针对手写西波文字因样本稀缺而难以使用传统神经网络进行有效识别的问题，提出了一种基于元学习的优化框架。该框架以长短期记忆神经网络(LSTM)作为元优化器，来学习如何优化一个基础的卷积神经网络(CNN)或深度神经网络(DNN)。其核心创新在于结合了两种策略：1) 通过在MNIST数据集上进行预训练，实现跨任务的知识迁移；2) 通过LSTM优化器实现动态的梯度调控策略，自适应地调整基础网络的参数更新。实验结果显示，该方法在10类手写西波文字的小样本测试集上，将CNN的识别准确率从24%显著提升至73%，并将训练时间缩短至368秒，远优于传统的数据增强方法。研究旨在为西波文等少数民族濒危文字的数字化保护提供一种高效、可行的小样本学习技术方案。

Abstract: This paper proposes a meta-learning-based optimization framework to address the challenge of effectively recognizing handwritten Xibo characters using traditional neural networks under extreme sample scarcity. The framework employs a Long Short-Term Memory (LSTM) network as a meta-optimizer to learn how to optimize a base Convolutional Neural Network (CNN) or Deep Neural Network (DNN). Its core innovation lies in combining two strategies: 1) Cross-task knowledge transfer achieved through pre-training on the MNIST dataset; 2) A dynamic gradient regulation strategy implemented by the LSTM optimizer to adaptively adjust the parameter updates of the base network. Experimental results demonstrate that on a few-shot test set of 10-class handwritten Xibo characters, this method significantly boosts the recognition accuracy of a CNN from 24% to 73%, while reducing the training time to 368 seconds, far outperforming traditional data augmentation methods. This research aims to provide an efficient and feasible few-shot learning technical solution for the digital preservation of endangered minority scripts like Xibo.