The KI‑UltraHaltbarkeit project, carried out from 1 April 2019 to 31 March 2023 under cost centre 13FH566IX6, aimed to develop artificial‑intelligence methods for non‑destructive assessment of fruit ripeness and shelf life. The research was conducted by a consortium that included the Fraunhofer Institute for Industrial Engineering and a network of academic partners, with funding provided through a German federal research programme.

A central technical contribution was the evaluation of convolutional neural networks (CNNs) for classifying Galia melons into three ripeness classes. After 100 training epochs the model achieved a test‑set accuracy of 80.27 %. The training accuracy reached 80.80 %, while recall and precision values were 80.27 % and 89 % respectively. Loss curves for training and testing data showed a gradual decrease, indicating stable learning, although the final accuracy fell slightly short of the team’s target performance. The same CNN architecture was adapted for colour‑image analysis of apples, watermelons, and Galia melons. Images were captured with a Canon EOS 450D at 4272 × 2848 px and pre‑processed with data‑augmentation techniques such as random flipping, zoom, and puzzle transformations. The resulting model reported a test‑set accuracy of 90.60 %. However, over‑fitting was evident from the divergence between training and testing loss, prompting further fine‑tuning of hyper‑parameters and network depth to improve generalisation.

Ultrasound‑based assessment was explored as a complementary non‑destructive technique. Experiments employed phased‑array transducers initially operating in the 1–3 MHz range, later shifting to single‑array configurations below 500 kHz. The team investigated the micro‑structural impact of water content and air pockets on acoustic penetration. Sampling strategies were compared using k‑Detrimental Point Processes (k‑DPP) against other point‑process methods. Table 4 of the report shows that k‑DPP achieved F‑score, accuracy, precision, and recall values comparable to the alternatives, indicating its suitability for robust feature extraction in ultrasound data.

The project also investigated few‑shot learning for fruit classification. Using the Fruit360 dataset, a Siamese network was trained to recognise watermelon varieties with only a handful of examples. The limited scope to a single fruit type led to pronounced over‑fitting, and the model’s generalisation to unseen data was poor. The authors noted that few‑shot learning remains an emerging field and that recent advances in active learning and low‑compute inference could mitigate these issues.

Throughout the project, the consortium maintained a close collaboration between the Fraunhofer Institute’s engineering expertise and the academic partners’ research capabilities. Regular workshops and joint publications ensured that methodological advances were rapidly translated into prototype systems. The project’s outcomes—high‑accuracy CNN models for visual ripeness detection, a validated ultrasound assessment protocol, and insights into few‑shot learning—provide a solid foundation for future commercialisation of portable, contamination‑free fruit quality monitoring tools.