A Multi-Stage Deep Learning Framework for Automatic Pigment Classification and Filter Prediction in Rock Art Paintings

Rock art paintings are a fragile but invaluable component of global cultural heritage. Digital enhancement methods such as the DStretch algorithm are widely used to reveal faded motifs, yet their effectiveness depends on the manual selection of filters and scales, a process that is subjective, time-consuming, and difficult to reproduce. This paper introduces the first automated framework for pigment classification, filter prediction, and scale selection in rock art enhancement. The system follows a three-stage pipeline: (i) pigment classification using HSV histograms and a Random Forest, (ii) filter prediction with multi-label MLP and CNN models tailored to red-only versus mixed pigment cases, and (iii) scale prediction usingVision Transformer embeddings (DINO-ViT) to capture global contextual information. The framework was evaluated on a dataset of 4, 900 images derived from 866 photographs, annotated by domain experts. The results show that pigment classification achieves 99% accuracy, Just-Red filter prediction reaches 93.67% accuracy, BW&Red filter prediction attains 71.5% accuracy, and scale prediction achieves 85.38% top-2 accuracy, significantly reducing the number of manual trials required by experts. Custom evaluation metrics, including top-k and ’at-least-two-correct’ measures, were introduced to align performance with archaeological practice. By reducing subjectivity and increasing reproducibility, this approach supports scalable digital documentation of rock art and demonstrates how artificial intelligence can be integrated into archaeological workflows.