An intelligent system for assessing the quality of agricultural products has been developed at Novosibirsk State Technical University (NSTU). The technology uses ordinary photographs to identify external defects and the freshness of fruits and vegetables.
The developers explained that traditional computer vision systems require ideal shooting conditions or tens of thousands of examples for training. NSTU researchers compared different approaches and identified the optimal one for each task: in some cases, it is important not to miss a spoiled fruit, in others – not to reject a quality one.
Project leader Egor Antonyants explained that the difficulty lay in training the neural network to distinguish between rot and immaturity, which often look similar in photos. For this, a triplet network architecture was used, taking into account several examples at once. This made it possible to reduce the number of errors.
The system does not require a huge number of images and works with ordinary photographs under different lighting conditions. Recognition accuracy reaches 89% even on unfamiliar products. This allows the system to be quickly adapted to new types of products.
The prototype successfully passed laboratory tests on static photographs. The plans include adapting the system to work with video streams from a conveyor belt. So far, training has been conducted on bananas, oranges, strawberries, and tomatoes. The developers plan to attract partners for joint testing on real production lines.