Abstract
The national cultural pattern is a precious treasure of the Chinese nation. Semantic annotation and analysis of national cultural pattern is useful in cultural heritage and modern recreation. This article presents a multi-label dictionary learning algorithm called similar coefficient multi-label incoherent dictionary learning (SCMIDL) based on a multi-class dictionary learning algorithm. SCMIDL combines the incoherence of the dictionary and the similarity of coefficients, which significantly improves the performance of multi-label annotation. The superior annotation ability of the algorithm was confirmed on three kinds of national cultural pattern datasets constructed for evaluation purposes.
Introduction
Traditional ethnic patterns—found in textiles, carpets, and architecture—represent a visual language rich with symbolic meaning. As digital archives of these patterns grow, manual labeling becomes an impossible bottleneck. The challenge is twofold: first, the multi-label nature of these patterns, where a single piece of embroidery might feature a "dragon," "clouds," and "flowers" simultaneously. Second, the scarcity of data in niche cultural heritage domains makes training massive deep learning models like Convolutional Neural Networks (CNNs) difficult due to overfitting.
The SCMIDL Algorithm
To address these challenges, researchers developed the Similar Coefficients & Multi-label Incoherent Dictionary Learning (SCMIDL) algorithm. Unlike standard classification, dictionary learning represents data as a sparse combination of basic elements (atoms), which is highly effective for smaller, high-dimensional datasets.
Key innovations in SCMIDL include:
- Dictionary Incoherence: The algorithm enforces distinctness between dictionaries of different classes. This ensures that the "dragon" dictionary does not overlap with the "flower" dictionary, leading to clearer classification.
- Label Correlation: Recognizing that certain symbols often co-occur (e.g., dragons and clouds), SCMIDL uses cosine similarity to ensure that related labels have similar reconstruction coefficients.
- Intra-class Scatter Penalty: A penalty term minimizes the variance of reconstruction coefficients for samples sharing the same label, effectively clustering similar patterns in the feature space.
Experimental Datasets
The study evaluated SCMIDL on three newly constructed, manually segmented datasets:
- Ming and Qing Court Dress Patterns: 899 images with 5 labels (e.g., butterfly, flower, dragon, phoenix, cloud).
- Traditional National Dress Patterns: 782 images with 6 labels focusing on Miao, Mongolian, and Uyghur styles.
- Traditional Carpet Patterns: 536 images with 3 labels (animal, plant, and geometry).
Results and Insights
SCMIDL was compared against established multi-label learning algorithms such as ML-KNN, Rank-SVM, and LLSF.
- Superior Precision: On the Traditional Carpet dataset, SCMIDL achieved an Average Precision of 0.9530.
- Lower Error Rates: In the complex Ming and Qing dress dataset, SCMIDL significantly reduced the "One-error" rate compared to standard dictionary methods.
- Stability: The inclusion of the intra-class scatter term provided greater stability across different data subsets.
The core insight is that mathematical integration of label relationships—acknowledging that a "phoenix" is likely to appear with "flowers"—allows the algorithm to use the presence of one motif to increase confidence in another.
Conclusion and Future Work
While SCMIDL performs excellently on specialized datasets, the authors note that scaling to larger archives remains a goal. Future work may involve hybrid models combining deep learning's feature extraction with the interpretability of dictionary learning, as well as handling "in-the-wild" photos where patterns may be obscured or shadowed.
Summary updated