Neural Network to Check Agricultural Product Quality
Scientists and students at Novosibirsk State Technical University NETI have developed an AI system that detects defects in fruits and vegetables and assesses their freshness.
Quality and safety inspection is a critical task for agribusinesses, food processors and retail chains. They need quality control systems that can keep substandard goods out of the supply chain and prevent food from spoiling during processing, storage and delivery to store shelves. The process is complex and demands human attention. At high volumes, identifying every defective product becomes extremely difficult. That is why the repetitive work of inspecting food products is increasingly being automated, with digital technologies making the process more effective. Artificial intelligence is now joining the effort.
A Smart Sorter
Fruits and vegetables are particularly vulnerable. Experts say they can easily be damaged at any point in the logistics chain, whether by the producer or the transportation company. Improper harvesting, impacts and deviations from required temperature conditions during storage and transportation can all cause produce to spoil rapidly.
In August 2026, third-year students Dmitry Shipunov, Alexander Sedelnikov and Semyon Simonyak of the Faculty of Automation and Computer Engineering at Novosibirsk State Technical University NETI (NSTU NETI), working under Yegor Antonyants, an assistant in the university’s Department of Automated Control Systems, presented an intelligent system for automatically assessing the quality of agricultural products. They developed a “smart sorter” that initially specializes in bananas, oranges, strawberries and tomatoes. The computer vision software classifies produce from images and can process photographs taken under different lighting conditions and from different angles. The system can assess fruit and vegetable quality without destructive testing or complex equipment.
The developers say they had to solve a difficult problem: catch every spoiled item without mistakenly rejecting good produce.
A Novel Network Architecture
To find the optimal approach, the team conducted a detailed comparison of fundamentally different methods for training neural network algorithms. This allowed the researchers to understand how AI identifies defects under each training approach and why one model is better at detecting rotten produce while another identifies underripeness. The hardest task was teaching a single neural network to distinguish specifically between rotten and unripe fruits and vegetables. Project leader Yegor Antonyants said the main difficulty is that these conditions can be hard to classify visually in many photographs because both can involve greenish or brownish hues.
“We used a specialized triplet network architecture that considers not simply pairs of images but several examples at once – an anchor, a positive example and two negative examples. This allowed us to create more flexible boundaries between classes and substantially reduce the number of errors,” Yegor Antonyants said.
The developers are now completing work on a research prototype. Three neural network architectures – CNN, SNN and TNN – have been trained and tested, achieving accuracy of up to 93.6% on an image dataset. The next stage will involve analyzing a video stream from a moving conveyor belt and testing the platform at an operating facility, followed by expanding the range of products the neural network can handle.
Smart Supply Chain Inspection
Russia is already developing industrial food sorting systems that monitor product quality. Lenta Group, for example, has introduced remote acceptance of fruits and vegetables at its distribution centers, where it is testing machine vision combined with neural networks. The system has performed well, with its assessments matching those of quality inspectors in more than 90% of cases. Digitalization also reduces the cognitive burden on employees and makes receiving and processing shipments several times faster.
The Russian market for such systems is expected to expand as the country’s agricultural sector becomes more digital. Over the long term, these technologies could become an important part of integrated smart supply chains stretching from the farm to the store shelf. They will allow producers to monitor storage and transportation conditions at every stage, improving transparency and reducing losses. Once integrated with Russian smart farming systems, retailers, in turn, will be able to follow production, harvesting, transportation, warehousing, storage and delivery to retail networks, ultimately helping guarantee quality for consumers.
Such technologies could also help increase Russian food exports by allowing smart supply chains to guarantee quality at every stage.
