Convolutional Neural Networks in Commercial Grading of Cherry

Manual quality control of fruit and berries remains a major challenge in the food industry. Existing visual assessment methods are subjective, labor-intensive, and dependent on operator skills, which limits their effectiveness in modern production environments. However, convolutional neural network models of the YOLO type can be used to classify cherries by commercial grade at the raw material acceptance stage.The study focused on cultivated cherries ( Prunus subg. Cerasus ) harvested in the Kemerovo Region (Russia) in 2024. The research was conducted at the Zdorovye Pitanie Research and Production Association (Kemerovo, Russia). The comparative analysis evaluated images across four quality categories using five YOLO models (YOLOv8m, YOLOv9m, YOLO11m, YOLO12m, YOLO26m) and three YOLO12 architectural variants of different sizes. Hyperparameter optimization was performed for the selected YOLO12m model, assessing eight optimizers, four batch sizes, and five Dropout regularization coefficient values.The medium-sized YOLO12 model demonstrated the highest accuracy. Among the optimizers, SGD yielded the best performance metrics. The optimal training batch size was determined, while adjusting the Dropout parameter had no significant effect on the results. Increasing the number of final training epochs improved recognition performance across all four categories. The model’s performance was validated on an industrial sample. Based on this model, a prototype software application with a graphical user interface (GUI) was developed for image loading, detection visualization, and report generation.The results confirm the suitability of YOLO models for automating incoming cherry quality control, providing objective and reproducible analysis instead of subjective manual assessment, thereby paving the way for advanced industrial quality control systems.