Classical, Deep, and Quantum Machine Learning for Automated Ulluco Quality Classification

Introduction

Ullucus tuberosus (ulluco) is an Andean tuber currently graded by subjective visual inspection, with no international quality standard. This study addresses two objectives: preventing defective tubers from contaminating healthy batches, and proposing a five-category grading framework to support standardization. We benchmarked classical machine learning, deep convolutional neural networks (CNNs), and quantum models for automated visual classification.

Methods

A dataset of 2,390 images was captured inside a lightbox using two cameras to capture different lighting conditions. Samples were manually labeled into five categories (Extra, Category I, Category II, Bad 1, Bad 2) using Label Studio. A background-removed version was generated via image segmentation. We extracted 40 features covering color, texture, shape, frequency, and wavelet domains. Three CNNs (Custom, MobileNetV3-Large, ResNet-50) and nine classical classifiers were trained, alongside quantum models: angle-encoding QNNs (pure 12‑qubit and hybrid with a pre‑network) and QSVMs with PCA to 8 qubits, all evaluated with 5‑fold cross‑validation.

Results and Discussion

MobileNetV3-Large achieved the best performance: 79.4% multiclass and 92.6% binary accuracy (good vs. bad), with fast inference (3.9 ms/img, 12.4 MB). The hybrid QNN reached 72.9% multiclass and 92.1% binary accuracy, outperforming all classical baselines, with 19 KB and 1.1 ms per inference. ResNet-50 scored 76.2% but was slower and larger. Logistic Regression, the best among the nine classical classifiers tested, attained 67.0%. The pure QNN performed near random (34.8%), highlighting the importance of the pre‑network in hybrid designs.

Conclusions

Deep CNNs remain the top choice for ulluco grading, but hybrid quantum‑classical models offer competitive accuracy with an extremely small footprint, suitable for edge deployment. Both objectives were met: reliable good/bad separation (92.6%) and five‑category grading (79.4%). This work provides the first quantum ML benchmark for Andean crops and a reproducible basis for formal ulluco quality standards, paving the way for automated quality control in Andean agriculture.