Verification of historical sketches via one-class learning on compact feature representations

Abstract

Historical sketch authentication is challenging because securely attributed reference sets are often small, and stylistic evidence is carried primarily by line, texture, tonal variation, and mark-making. We present a reproducible framework for verifying historical sketches using artist-specific one-class autoencoders trained on compact handcrafted feature representations. Ten artist models were trained using authenticated sketches from six open-access cultural heritage collections. Each drawing was represented by five interpretable descriptors, namely, Fourier-domain energy, Shannon entropy, global contrast, Grey-Level Co-occurrence Matrix homogeneity, and box-counting fractal complexity. The system was evaluated using a biometric-style verification protocol in which each artist model was tested on genuine held-out works and impostor works by other artists. On the primary evaluation partition of 900 decisions, comprising 90 genuine and 810 impostor trials, the method achieved 87.6% balanced accuracy, 77.8% True Acceptance Rate, 2.6% False Acceptance Rate, 0.748 Matthews Correlation Coefficient, and 11.4% Equal Error Rate. Performance remained stable across 20 repeated random train/test splits. The proposed model also outperformed Gaussian and one-class SVM baselines, while pretrained ResNet50 and EfficientNet-V2 feature representations performed substantially worse in this data-scarce setting. Leave-one-feature-out ablation confirmed that all five descriptors contributed positively, with fractal complexity and GLCM homogeneity providing the strongest individual contributions. Error analysis revealed structured false-accept pathways to be consistent with stylistic proximity between artists. The framework provides transparent, reproducible, and interpretable quantitative evidence for historical sketch verification. It is intended to support, not replace, expert connoisseurship in attribution settings where available reference corpora are limited.