Key Takeaways
- Chloros, a Tokyo-based crop imaging and AI company, published an SSRN preprint on a quantum-classical hybrid model for fine-grained tomato ripeness classification.
- The model pairs the YOLOv8 object-detection architecture with a four-qubit variational quantum circuit to sort tomatoes into four ripeness stages.
- In simulation, it reached 93.80% accuracy, versus 87.44% for a classical hybrid model, clearing the 90% benchmark Chloros identified for field use.
- Macro F1 rose from 58.92% to 73.63%, and recall for near-ripe tomatoes jumped from 21.33% to 70.51%.
- The study is simulation-based; future work will test quantum hardware and real field imagery.
Chloros Inc., a developer of crop imaging and AI-based analysis tools, has published a preprint presenting a quantum-classical hybrid approach to classifying tomato ripeness. The study, posted on the SSRN preprint server, was co-authored by Chloros co-founder and chief technology officer Yuichi Ito and Geetha K S of Yokohama National University, who conducted the work during an internship at the company.
How the Chloros model works
The approach combines the YOLOv8 detection architecture with a four-qubit variational quantum circuit that uses parameterized rotations and CNOT-based entanglement. After YOLOv8 detects the fruit, extracted features are encoded into the quantum circuit, whose outputs drive the final classification into four ripeness stages. Distinguishing adjacent stages is difficult because color and texture overlap, especially between half-ripe and near-ripe fruit, and the task grows harder with changes in lighting, occlusion and background common in real farm imagery.
Simulation results
In the simulation-based evaluation, the hybrid model achieved 93.80% classification accuracy, compared with 87.44% for a classical hybrid tested under the same conditions, exceeding the 90% threshold Chloros identified through customer interviews as a practical benchmark for field deployment. Macro F1, which reflects performance across all four classes, climbed from 58.92% to 73.63%. The largest gain came in the near-ripe class, where recall rose from 21.33% to 70.51%, an improvement of 49.18 percentage points.
“The improvement in near-ripe recall is particularly encouraging. While this study is simulation-based, it highlights the potential of quantum-classical AI for agricultural imaging,” said Yuichi Ito, co-founder and CTO of Chloros.
What comes next for Chloros
The company notes the work is an initial evaluation on a public dataset with a simulated circuit, and does not test performance on quantum hardware or in the field. Future research will assess reproducibility on quantum hardware and applicability to field imagery, with potential uses in ripeness assessment, harvesting support and yield estimation. Accurate, automated ripeness detection can help growers time harvests, reduce waste and grade produce more consistently. Founded in 2025 and headquartered in Tokyo, Chloros builds AI and data-driven farming tools that draw on precision agriculture imaging from drones and smartphones.
