Neural Network Trained to Detect Rotten Fruit and Vegetables From Ordinary Photos
The technology has already undergone laboratory testing.
Researchers at Novosibirsk State Technical University have developed a neural network for assessing the quality of agricultural products using ordinary photographs. The scientists used a triplet network architecture that can distinguish between rot and immaturity. The algorithms achieve up to 89% accuracy even on products that were not included in the training dataset.
The neural network performs significantly better than standard computer vision technologies that require ideal imaging conditions.
"We did not simply create another classifier but conducted a detailed comparison of fundamentally different training approaches. This allowed us to understand how each method identifies defects and why one model performs better at detecting rotten produce while another is more effective at identifying immaturity," the developers explained.
The neural network has completed laboratory testing. The team plans to adapt the system to process real-time video streams at agricultural enterprises.
