Scalable, Parameterizable Model Generation of Fruits for Machine Learning of their Life Cycles DigitalFruitTwin

The food industry faces growing demands for higher quality standards, especially con-cerning the freshness of fruits and vegetables. Additionally, there is a strong emphasis on reducing the wastage of food that remains suitable for consumption. Legislation across all EU countries is being implemented to drastically reduce food waste, empha-sizing the importance of sustainable practices and the responsible use of resources in the food industry. In order to contribute with such goals, multiple AI-solutions are being currently developed. However, modeling a complex metabolic process such as fruit ripening is very difficult and, in the case of apples, the following problems are encountered in general: (1) Models usually require multimodal inputs, such as ethylene concentration, hardness, and thermograms taken by a thermal imaging camera obtained through sensors, which are more difficult to obtain.
(2) Open-source datasets documenting the complete ripening cycle of apples from pick-ing to rotting are not yet available, and all data must be collected on our own. (3) The model's predictions should have structural similarity to the inputs, on top of which it must show the differences between different ripening stages. To address this set of problems, we use Deep Learning (DL) to develop a scalable model that maps the ripening cycle of apples using both real and synthetic data. The fol-lowing are the solutions:
(1) Only traditional photography is utilized to capture apple surface information, and unpaired images are classified into different ripening stages for training. (2) To reduce the complexity of the model, only one model is utilized to accomplish the prediction of different ripening stages. (3) The generated images are processed to ensure that they are closer to real images. In summary, our proposed deep learning solution is able to effectively model the ripen-ing process of apples despite limited data access. By combining real and synthetic data, adopting a unified model structure, and optimizing the generated results with post-pro-cessing, we have not only achieved visual prediction of apple ripening stages, but also provided feasible ideas and methods for intelligent monitoring and quality assessment of other agricultural products. This research is expected to contribute to the digital trans-formation of the food industry, the reduction of food waste, and the realization of sus-tainable development goals.