Benchmarking deep and hybrid quantum-classical models for Gallo-Roman ceramic sherd classification: a reproducible evaluation framework

The classification of Gallo-Roman ceramic fragments remains time-consuming and subject to variation between experts. Using a corpus of images from northern France, we compare two computer vision approaches: classification of isolated object crops and end-to-end detection on full images. We evaluate ResNet-18, MobileNetV3-Small, YOLOv8-s, YOLOv11-s, RT-DETR-L, QCNN and QCNN-VQE in unbalanced and rebalanced training scenarios. Classification on cropped images yields the most stable results, with MobileNetV3-Small achieving macro-F1 scores of 0.964 and 0.942 in unbalanced and balanced contexts, respectively, whilst QCNN-VQE remains competitive with a very low memory footprint. End-to-end detectors are more sensitive to class imbalance and the choice of confidence threshold. Terra sigillata is consistently the most recognisable class, whilst residual errors mainly reflect confusion between common oxidised and reduced ceramics. These results show that task formulation strongly conditions the interpretation of AI performance in heritage science.