Tomato type classification through images: a case study using convolutional neural networks

The aim of this work is to use artificial neural networks to classify different types of tomatoes (variety or cultivar) on sale in establishments in the western region of Paraná. The classification process was carried out using models based on convolutional neural networks, and included image acquisition and processing, training, validation and testing of the classification models. A smartphone camera was used to build the tomato image database. The models chosen for classification were: ResNet50V2, MobileNet, MobileNetV2, Xception, EfficientNetV2B0, EfficientNetV2B1, EfficientNetV2B2 and EfficientNetV2B3, which were implemented in the Python language using the Keras API of the TensorFlow framework. The results obtained from the eight models were compared using accuracy and loss. For a test set of 12 images, significantly different from the training set, the MobileNet model performed best, with a test accuracy of 83.33% and an error of 0.622.