基于改进ResNet-18的京剧旦角服饰分类研究

摘要

为建立京剧服饰数据库、保护戏曲服饰的文化遗产价值，本文针对分类准确率不高的问题，提出一种基于改进ResNet-18的京剧旦角服饰分类模型（ResNet18-APSS）。首先，收集京剧旦角服饰样本图像，构建包含18个类别的数据集；其次，对预训练的ResNet-18模型进行改造，以APSS模块替换其第四层卷积模块；再次，采用Optuna进行超参数寻优以获得最优配置，并引入迁移学习与自动混合精度训练策略以进一步提升模型性能，同时将损失函数调整为交叉熵损失函数以加快收敛速度；最后，使用寻优得到的最优超参数完成训练与验证。在自建京剧旦角服饰数据集上的对比结果表明，改进后的模型能够有效实现京剧旦角服饰分类，平均准确率达到95.41%。本研究为京剧旦角服饰的分类难题提供了有效解决方案，对推进其数字化进程具有重要意义。

Abstract

To establish a Peking opera costume database and preserve the cultural heritage of opera costumes, an improved ResNet-18-based classification model (ResNet18-APSS) for classifying Peking opera Dan costumes is proposed, addressing the issue of classification accuracy. First, Peking opera Dan costume sample images were collected to construct a dataset comprising 18 categories. Then, the pre-trained ResNet-18 model was modified by replacing the fourth convolutional layer with an APSS module. Hyperparameter optimization was performed using Optuna to find the optimal configuration. Transfer learning and automatic mixed precision training strategies were employed to enhance model performance further, and the loss function was adjusted to CrossEntropyLoss to accelerate convergence. Finally, the identified optimal hyperparameters were used for training and validation. Comparative results on the self-constructed Peking opera Dan costume dataset demonstrate that the improved model effectively classifies Peking opera Dan costumes, achieving an average accuracy of 95.41%. This study provides an effective solution to the classification challenges of Peking opera Dan costumes and significantly contributes to advancing their digitization.