The marble industry relies on proprietary commercial names rather than objective visual categories, creating market inefficiencies for stakeholders who select stones based on appearance. Supervised classification perpetuates this problem by replicating inconsistent commercial labels instead of discovering intrinsic visual structure. We propose an unsupervised pipeline combining a two-stage training strategy: A pure self-supervised pretraining followed by cluster-aware fine-tuning of a DINO Vision Transformer, with empirically selected dimensionality reduction and agglomerative hierarchical clustering. Systematic ablation studies on 1480 marble images spanning 10 commercial varieties validate each design choice: cluster-aware training at k = 10 yields geometrically improved embeddings over the self-supervised baseline (mean Silhouette Score 0.693 ± 0.053 vs. 0.660 ± 0.030; mean Davies–Bouldin Index 0.386 ± 0.075 vs. 0.569 ± 0.012; N = 9 independent evaluations across 3 data partitions × 3 training initializations). The resulting taxonomy reveals three phenomena invisible to commercial classification: cross-category merging of visually indistinguishable stones carrying different market names, intra-category splitting of heterogeneous sub-populations within single varieties, and coherent grouping where commercial and visual boundaries coincide, with all three confirmed in every independent run. We further demonstrate that standard extrinsic metrics are misaligned with unsupervised taxonomy objectives when reference labels encode the inconsistencies the method aims to resolve. Validating this methodology across diverse stone types, larger datasets, and varied acquisition conditions represents a natural and necessary next step toward establishing its cross-domain generalizability.
Generative Adversarial Networks (GANs) have demonstrated remarkable capabilities for synthesizing photorealistic textures, yet deploying conditional GANs (cGANs) in industrial settings faces two barriers: the prohibitive cost of annotating proprietary data and the uncertain alignment between automated metrics and human perception. This study addresses both challenges for marble texture synthesis. We adapt an unsupervised segmentation pipeline combining Simple Linear Iterative Clustering (SLIC) superpixels, Gaussian Mixture Models (GMMs), and Graph Cut optimization to extract vein structures from 289 industrial scans without manual annotation. We then benchmark four cGAN architectures, a baseline cGAN, Pix2Pix, BicycleGAN, and GauGAN, using a dual-evaluation protocol contrasting automated assessment via pixel-based metrics, structural metrics, statistical metrics, and learned distributional metrics with human-centered assessment. Results reveal a significant metric–perception discrepancy: Pix2Pix achieved the best FID yet received the lowest human ratings due to checkerboard artifacts, whereas GauGAN produced textures statistically indistinguishable from real marble (Visual Turing Pass Rate, VTPR: 0.533; Mean Opinion Score on Marble Authenticity, MOS-MA: 2.89) despite inferior FID (87.3). These findings establish three contributions: (1) an unsupervised annotation-free segmentation pipeline, (2) empirical evidence that automated metrics alone are insufficient for architecture selection, and (3) a dual-evaluation framework validating human-in-the-loop assessment as essential for industrial deployment.
The marble industry relies on proprietary commercial names rather than objective visual categories, creating market inefficiencies for stakeholders who select stones based on appearance. Supervised classification methods perpetuate this problem by replicating inconsistent commercial labels instead of discovering intrinsic visual structure. We propose an unsupervised pipeline combining a two-stage training strategy, pure self-supervised pretraining followed by cluster-aware fine-tuning of a DINO Vision Transformer, with UMAP dimensionality reduction and Ward's agglomerative hierarchical clustering. Systematic ablation studies on 1,540 marble images spanning 10 commercial varieties validate each design choice: cluster-aware training at k=10 yields superior embeddings over the self-supervised baseline (Silhouette Score 0.778 vs. 0.761; Davies–Bouldin Index 0.293 vs. 0.364), UMAP compression to five dimensions resolves high-dimensional noise pathologies, and Ward's linkage produces the most compact partitions. The resulting taxonomy reveals three phenomena invisible to commercial classification: cross-category merging of visually indistinguishable stones carrying different market names, intra-category splitting of heterogeneous sub-populations within single varieties, and coherent grouping where commercial and visual boundaries coincide. We further demonstrate that standard extrinsic metrics are misaligned with unsupervised taxonomy objectives when reference labels encode the inconsistencies the method aims to resolve. This work provides a validated methodology for data-driven visual classification in the natural stone industry and a transferable template for domains with unreliable labelling conventions.
Deploying conditional Generative Adversarial Networks (cGANs) for industrial texture synthesis faces two barriers: the prohibitive cost of manual data annotation and the uncertain alignment between automated evaluation metrics and human perception. This study addresses both challenges for marble texture synthesis using 289 high-resolution industrial scans. We adapt an unsupervised segmentation pipeline combining Simple Linear Iterative Clustering (SLIC) superpixels, Gaussian Mixture Models (GMMs), and graph cut optimization to extract vein structures without manual annotation. Four cGAN architectures—baseline cGAN, Pix2Pix, BicycleGAN, and GauGAN—are benchmarked using a dual-evaluation protocol contrasting ten automated metrics with structured human-centered assessment. The results reveal a significant metric–perception discrepancy. Pix2Pix achieved the best Fréchet Inception Distance (FID = 85.3) yet received the lowest human ratings due to periodic texture artifacts. GauGAN produced textures statistically indistinguishable from real marble, achieving a Visual Turing Pass Rate (VTPR) of 0.533 and a Mean Opinion Score on Marble Authenticity (MOS-MA) of 2.89, despite an inferior FID (87.3). These findings make three contributions: an annotation-free segmentation pipeline, empirical evidence that automated metrics alone are insufficient for architecture selection, and a dual-evaluation framework that establishes human-in-the-loop assessment as essential for quality-critical industrial deployment.