SmartFruit: an Embedded AI System to Detect Fruit Ripeness and Prevent Food Waste

Food waste is one of the most important problems in the agri-food supply chain. Because fruits are especially perishable, ripeness is frequently determined by empirical or chemical methods, which can be imprecise, expensive, and challenging to automate on a big scale. Artificial intelligence tools are essential for streamlining the agricultural supply chain and reducing these inefficiencies. To identify the maturity of fruit in real time, this study proposes an embedded vision system based on a CNN model implemented on a low-cost, low-power platform. A free dataset, from Mendeley, with 16 classes was chosen, and three models - VGG-19, MobileNetV2 and a custom one, SmartFruit – were tested. Thanks to fine-tuning and transfer learning, all architectures showed encouraging results. The SmartFruit model was the most successful, with a testing accuracy of 90% and a memory occupancy of 0.57 MB. This makes it appropriate for deployment on an OpenMV Cam H7 Plus, a low-cost embedded device that could be used to implement a fruit ripeness detection system prototype.