| 蒋语昕*,陈玲洁*,郭晨阳**.基于GOQPSO-Swin Transformer的画作风格分类模型研究[J].高技术通讯(中文),2025,35(11):1201~1212 |
| 基于GOQPSO-Swin Transformer的画作风格分类模型研究 |
| Painting style intelligent recognition using GOQPSO-Swin Transformer hybrid approach |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 11. 005 |
| 中文关键词: 广义反向量子粒子群算法； Swin Transformer； 画作风格分类； 智能化； 美术 |
| 英文关键词: generalized opposition quantum-behaved particle swarm optimization, Swin Transformer, painting style classification, intellectualization, fine art |
| 基金项目: |
| 作者 | 单位 | | 蒋语昕* | (*云南大学艺术与设计学院昆明 650093)
(**太原理工大学安全与应急管理工程学院太原 030024) | | 陈玲洁* | | | 郭晨阳** | |
|
| 摘要点击次数: 84 |
| 全文下载次数: 61 |
| 中文摘要: |
| 智能、快速、精准的画作风格分类技术对推动艺术教育、艺术鉴赏、绘画创作和文化遗产保护等领域的智能化发展具有重要意义。卷积神经网络（convolutional neural network，CNN）作为当前画作风格分类的主流算法，在计算原理上的不足严重影响分类性能。计算机视觉领域表现优异的Swin Transformer在画作风格分类任务中展现了新的潜力，但其性能高度依赖于超参数设定。因此，本文以广义反向量子粒子群算法（generalized opposition quantum-behaved particle swarm optimization，GOQPSO）优化Swin Transformer，构建了GOQPSO-Swin Transformer画作风格分类模型。实验结果表明，该方法具备优异的参数优化能力，能够自动发现并利用有效的非常规超参数组合，在参数优化阶段取得了85.63%的适应度，并在画作风格数据集上最终取得了87.81%的首位准确率，与对比模型相比，各项指标提升了1.42%~13.93%。实验结果证明较主流CNN算法，Swin Transformer在画作风格分类任务中更具优势；同时证明本文所构建的GOQPSO-Swin Transformer模型能够充分挖掘模型在画作风格分类任务中的潜力，为该任务提供了新的智能技术路径。 |
| 英文摘要: |
| Intelligent, fast, and accurate painting style classification technologies are of great significance for advancing the intelligent development of art education, art appreciation, painting creation, and cultural heritage preservation. Convolutional neural network (CNN), as the mainstream algorithm for painting style classification, suffers from computational limitations that significantly hinder classification performance. In recent years, Swin Transformer, which has demonstrated outstanding performance in computer vision, has shown new potential in painting style classification tasks; however, its performance is highly dependent on hyperparameter settings. Therefore, this study proposes a GOQPSO-Swin Transformer model, in which Swin Transformer is optimized by generalized opposition quantum-behaved particle swarm optimization (GOQPSO). Experimental results show that the proposed method exhibits superior parameter optimization capability, automatically discovering and leveraging effective unconventional hyperparameter combinations. During the parameter optimization phase, the model achieved a fitness of 85.63%, and ultimately attained a Top-1 accuracy of 87.81% on the painting style dataset, outperforming comparative models by 1.42%~13.93% across various metrics. These findings demonstrate that compared to mainstream CNN algorithms, Swin Transformer has greater advantages in painting style classification tasks. Moreover, they validate that the proposed GOQPSO-Swin Transformer model can fully exploit the potential of Swin Transformer in this domain, providing a novel intelligent technical pathway for painting style classification. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
