How AI Is Transforming Ceramic Artifact Classification and Market Valuation

Research Overview

A joint team from Universiti Putra Malaysia and the University of New South Wales Sydney proposes an intelligent framework that combines computer‑vision and economic reasoning to automatically classify ceramic artifacts and predict their market value. The framework uses an enhanced YOLOv11 model for visual feature extraction and a random‑forest regression model trained on multi‑source auction data.

Data Collection and Annotation

A curated dataset of 8,213 high‑resolution images was assembled from three channels: auction houses (42.6 %), museum collections (24.3 %), and online marketplaces (33.1 %). The images cover 20 kiln‑specific styles and decorative techniques. Annotation follows a mixed workflow: an AI pre‑labeler (a pretrained YOLO detector) generates bounding boxes, then domain experts refine the labels with LabelImg, assigning three‑level categories (style, shape, pattern). The data split is 7:2:1 for training, validation, and testing.

YOLOv11 Model Enhancements

The baseline YOLOv11 architecture (ResNet‑50 backbone) is augmented with three attention‑focused modules:

C3k2‑EIEM : captures edge information, preserves spatial detail, and fuses features to improve detection of fine decorative elements such as carving and inscriptions.

Fast Spatial Pyramid Pooling (SPPF) : performs multi‑scale pooling to extract diverse visual cues.

C2PSA (Cross‑Stage Local Attention) : adaptively re‑weights background versus key regions, enhancing sensitivity to underglaze painting and intricate patterns.

Training employs 5‑fold cross‑validation, K‑means anchor optimisation, a cyclic learning‑rate scheduler, early stopping, and model checkpointing. Image augmentations include Mosaic, GridMask and MixUp to improve robustness and generalisation.

Random‑Forest Regression for Valuation

Visual features extracted by the enhanced YOLOv11 are fed into a random‑forest regressor. Categorical attributes (decorative patterns, style) are one‑hot encoded; quantitative attributes (shape metrics, complexity scores) are min‑max scaled. Auction prices from six major houses (2000‑2024) are normalised to 2024 USD using historical exchange rates and CPI data, and outliers are removed via the inter‑quartile‑range (IQR) method.

The regression pipeline consists of:

Data preprocessing (normalisation, outlier removal).

Feature extraction (YOLOv11 visual embeddings, one‑hot encoded categorical variables).

Ensemble training (random‑forest with multiple decision trees to reduce variance).

Prediction and evaluation.

Feature‑importance analysis shows that craft complexity and decorative style are the strongest predictors of market value, aligning with historical auction trends.

Results

Compared with the original YOLOv11, the enhanced model improves mean average precision (mAP) and F1‑score by ~2 % while slightly reducing raw precision, achieving a better balance between false positives and false negatives.

The random‑forest regression model attains 99.65 % accuracy, precision, recall and F1 on the training set, and 98.91 % on an independent test set, demonstrating stable and reliable price prediction.

Overall, the integrated framework provides a scalable, interpretable solution for automatic ceramic classification and market‑value estimation, illustrating how AI can modernise heritage‑object appraisal.

Paper:

https://www.nature.com/articles/s40494-025-01886-6