Benchmarking deep and hybrid quantum-classical models for Gallo-Roman ceramic sherd classification: a reproducible evaluation framework
Year: 2026
Type: article
Abstract: Gallo-Roman ceramic sherd classification is time-consuming and sensitive to inter-operator variability, especially under class imbalance. Using an image corpus from an urban archaeological context in Amiens (northern France), we benchmark two paradigms: (i) conditioned classification on ground-truth crops and (ii) end-to-end detection on full images. We compare lightweight CNNs (MobileNetV3-Small, ResNet-18), hybrid CNN/parameterised quantum-circuit classifiers (QCNN, QCNN-VQE) and detectors (YOLOv8s, YOLOv11s, RT-DETR-L) under an imbalanced “field” scenario and a balanced training scenario (train-only rebalancing). On crops, MobileNetV3-Small achieves macro-F1 0.964 (imbalanced) and 0.942 (balanced), while QCNN-VQE reaches 0.950 and 0.926 with ~ 0.36 MiB weights-only. For detection, YOLOv8s/YOLOv11s reach macro-F1 0.866/0.860 in the imbalanced setting. Rebalancing does not improve validation globally and degrades OW, highlighting a diversity–bias trade-off. We finally illustrate a post-hoc detection+reclassification fusion relevant to embedded deployment.
Source: npj Heritage Science
Authors: Cyrille Chaidron, Hafsa Taiebi Imrani
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7162494992