一种基于多任务学习的文化创意陶瓷产品视觉特征提取方法

A Multi-Task Learning Based Method for Visual Feature Extraction of Cultural and Creative Ceramic Products

摘要: 文化创意陶瓷产品视觉特征提取，即从图像中自动输出釉色、纹饰、器型及工艺风格等特征的类别标签，是图像素材库自动标注以及基于特征的图像生成与检索等应用的基础性工作。传统方法依赖人工对这四类视觉特征逐一判断，效率低且主观性强。现有深度学习方案采用单任务学习范式，对每类特征分别训练独立的分类模型，推理效率较低。本文提出一种基于多任务学习的文创陶瓷产品视觉特征提取方法(MTL-CVFE)，将釉色、纹饰、器型、工艺风格四项任务整合于同一模型，以单次前向传播同步输出类别标签。该方法以ResNeXt50-SSD提取基础特征，引入扩张卷积支路增强纹饰感受野，并设计四任务并行输出头，实现四类特征的端到端同步识别。同时构建了相应标注数据集用于实验验证。结果表明，本方法在测试集上综合平均精度均值达到85%以上，在保持各任务识别精度的前提下，推理效率较现有单任务深度学习方案有明显提升。本文方法输出的结构化特征向量可直接服务于设计素材库的自动标注与生成式设计模型的条件控制等下游应用场景。

Abstract: Visual feature extraction of cultural and creative ceramic products, which automatically outputs category labels of features including glaze color, decorative pattern, form, and craft style from images, is a foundational task for applications such as automatic annotation of image material libraries and feature-based image generation and retrieval. Traditional methods rely on manual judgment of these four types of visual features, which is inefficient and highly subjective. Existing deep learning schemes adopt a single-task learning paradigm, training independent classification models for each feature type, resulting in low inference efficiency. This paper proposes a multi-task learning based method for visual feature extraction of cultural and creative ceramic products (MTL-CVFE), which integrates the four tasks of glaze color, decorative pattern, form, and craft style into a single model and outputs category labels synchronously through a single forward propagation. The method employs ResNeXt50-SSD for backbone feature extraction, introduces dilated convolution branches to enhance the receptive field for decorative patterns, and designs a four-task parallel output head to achieve end-to-end synchronous recognition of the four feature types. A corresponding annotated dataset was constructed for experimental validation. Results show that the proposed method achieves a comprehensive mean average precision (mAP) of over 85% on the test set, and under the premise of maintaining recognition accuracy for each task, the inference efficiency is significantly improved compared with existing single-task deep learning schemes. The structured feature vectors output by the proposed method can be directly applied to downstream scenarios such as automatic annotation of design material libraries and conditional control of generative design models.