Convolutional Neural Network to Detect the Optimal Water Content of Cassava Chips During the Drying Process

Abstract

Cassava chips are used as raw materials to manufacture modified cassava flour. To produce high-quality modified cassava flour, a drying process for cassava chips is required to produce optimal water content in the range of 15-18% wb. This study aims to detect the optimal water content of cassava chips during the drying process in a hybrid hot-air tray dryer with computer vision using a convolutional neural network. Three categories of cassava chips' water content during the drying process are wet (water content of 55-70% wb), semi-dry (20-40% wb), and optimal dry (15-18% wb). In this study, the performance of four types of the pre-trained convolutional neural network, i.e., AlexNet, GoogLeNet, ResNet-50, and SqueezeNet, were compared by using different optimizers (SGDm, Adam, and RMSProp) and different learning rate values, 0.00005 and 0.0001, resulting in 24 types of experimental design. The results showed 12 convolutional neural network models with perfect validation accuracy. AlexNet with the SGDm optimizer and learning rate of 0.00005 was determined as the best model because of its stable training iteration process that experienced no fluctuations, perfect validation accuracy, specifically 100%, as well as perfect testing accuracy was 100%, and fastest training and validation process time, notably 32 minutes. This best convolutional neural network model will later be used to develop a rapid, real-time, and accurate hybrid hot-air tray dryer with computer vision to maintain cassava chip products with optimal water content.