Abstract
The preservation of the Mogao Grottoes murals requires non-contact methods to detect "flaking," a condition where paint layers detach from their substrate. This paper presents an unsupervised pipeline using hyperspectral imaging (HSI) and the Ordering Points to Identify the Clustering Structure (OPTICS) algorithm. By extracting spectral features through Sparse Auto-Encoders (SAE) and applying density-based clustering, the method generates continuous severity maps of mural decay without requiring labeled training data. Experimental results demonstrate that the SAE-OPTICS approach effectively distinguishes between pigment variations and actual structural deterioration, outperforming traditional clustering techniques.
Introduction
The Mogao Grottoes in Dunhuang, China, represent a pinnacle of Buddhist art, yet their age makes them highly susceptible to environmental degradation. "Flaking" is one of the most prevalent "diseases" affecting these murals, characterized by the gradual detachment of the pigment layer. Traditional assessment relies on manual inspection, which is subjective and potentially invasive. While Hyperspectral Imaging (HSI) offers a non-destructive alternative by capturing hundreds of narrow spectral bands, the lack of annotated datasets and the continuous nature of flaking progression make automated classification challenging.
Methodology and Data Acquisition
The proposed workflow utilizes HSI data collected from the 126th kiln (Tang Dynasty) using a V10-PS hyperspectral camera, covering a range of 400nm to 1000nm.
The analysis follows a three-step process:
- Feature Extraction: To manage the high dimensionality of HSI data (728 bands), Sparse Auto-Encoders (SAE) are employed. SAEs are particularly effective at compressing data while preserving the unique spectral "step shapes" associated with flaking, helping to ignore irrelevant pigment-based color changes.
- Density-Based Clustering: The OPTICS algorithm is used to identify clusters within the feature space. Unlike K-means, OPTICS does not require a pre-defined number of clusters and is robust against the noise typical of ancient surfaces.
- Severity Mapping: By calculating the Euclidean distance between a identified "core" flaking point and other pixels in the feature space, the system generates a continuous distribution map representing the degree of flaking.
Results and Discussion
The SAE+OPTICS pipeline was compared against standard methods like PCA+OPTICS and various K-means iterations. A key finding was the SAE's ability to normalize spectral differences caused by different pigments, which often lead to false positives in simpler models.
As shown in the results, the SAE-based method produces a cleaner, more accurate distribution map that aligns with expert manual diagnoses. The final "heat maps" (red indicating severe flaking, blue indicating normal areas) provide conservators with an intuitive tool for prioritizing restoration efforts. The method proved stable across different mural sections and insensitive to minor hyperparameter adjustments.
Conclusion
This unsupervised approach provides a robust, non-contact solution for monitoring the health of cultural relics. By leveraging the density-based clustering of OPTICS and the feature-learning capabilities of SAE, the system overcomes the primary hurdles of data labeling and pigment interference. Future work will focus on automating the reference point selection and optimizing computational efficiency for large-scale surveys of the Grottoes.
Summary updated