卷积神经网络DenseNet在阅读障碍手写图像分类中的应用

摘要

阅读障碍是一种与词汇层面阅读困难相关的特定学习障碍（SLD）。由于语音加工与正字法加工存在共同缺陷，该障碍常在儿童手写中表现为间距不规则、字母大小不一致。早期识别至关重要，但传统诊断流程耗时较长，不适用于大规模筛查。本研究旨在利用DenseNet121卷积神经网络（CNN）模型开展段落级手写分析，为资源受限的教育场景提供一种低成本的阅读障碍筛查工具。本文对100幅英文手写图像进行预处理并标准化为200个样本，其中70%的数据采用4折交叉验证进行评估，其余30%用于测试。该模型取得90%的测试准确率与92.86%的训练准确率，显著优于随机森林基线模型（训练准确率83.57%、测试准确率63.33%），其统计显著性经McNemar检验确认。本研究的主要贡献在于证明：采用段落级分析的轻量化单一架构DenseNet121，能够在计算资源需求显著更低、流水线更为简化的前提下，取得与依赖复杂混合模型和字符级分析的既有研究相当的性能。上述发现表明，DenseNet121可为资源有限的教育环境中的阅读障碍初步筛查提供稳健而低成本的解决方案。

Abstract

Dyslexia is a specific learning disability (SLD) associated with word-level reading difficulties and often manifests in childhood handwriting through irregular spacing and inconsistent letter sizing, due to shared phonological and orthographic processing. Early identification is critical; however, traditional diagnostic procedures are time-consuming and unsuitable for large-scale screening. This study aimed to develop a handwriting analysis at the paragraph-level using a DenseNet121 convolutional neural network (CNN) model as a low-cost dyslexia screening tool for resource-constrained educational settings. One hundred English handwriting images were preprocessed and standardized into two hundred samples, with 70% of the dataset evaluated using 4-fold cross-validation and the remaining 30% used for testing. The model achieved 90% test accuracy and 92.86% training accuracy, significantly outperforming a random forest baseline that reached 83.57% train accuracy and 63.33% test accuracy, with statistical significance confirmed by McNemar's test. The main contribution of this study is the demonstration that a lightweight, single-architecture DenseNet121 using paragraph-level analysis can achieve competitive performance compared to prior studies that relied on more complex hybrid models and character-level analysis, while requiring substantially lower computational resources and a simplified pipeline. These findings indicate that DenseNet121 provides a robust and low-cost solution for preliminary dyslexia screening in resource-limited educational environments.