Convolutional Neural Networks in Granulated Kissel Production

Artificial intelligence can be used to monitor production parameters in the food industry. Kissel is a jelly-like fruit or berry starch drink. Instant kissel usually consists of granules. Neural networks may help to control the size of kissel granules. In this research, convolutional neural networks monitored the production parameters of granulated kissel powder by localizing granules in an image. Size is the most important parameter of kissel granules: it should remain between 2 and 5 mm. To detects larger granules (≥ 5 mm), the network was provided with a visual dataset of granules of varying sizes. The localization models were developed using Detectron2. The research yielded a set of optimal operating principles and quality metrics. The R50-FPN model achieved the best results. The AP50 metric had the highest value, followed by AP75 and AP. The models performed well in visual detection and successfully determined the coordinates of the bounding rectangle. The resulting dataset did not label objects for small (APs) and medium (APm) sizes because the study focused on localizing large granules. The APl metric values for all models were high. The approach to AI training and neural network architecture proved optimal for food production control. The trained model made it possible to develop a computer program based on convolutional neural networks that demonstrated good results in detecting large granules in instant kissel powder. The new program can be used in continuous production to monitor the size of finished products and their compliance with process parameters.