Comparison performance of the CNN-based deep learning models for the distinguishing ultrasound pretreated and microwave dried jujube fruits

Abstract

Classifying dried fruits with economic importance and high nutritional content using novel techniques is crucial for achieving uniformity and practicality. It also plays a key role in identifying and distinguishing dried products, benefiting end consumers and the food processing industry. In this study, jujube slices were microwave-dried with and without ultrasound pretreatment at 100, 200, 300, and 600 W (watt) power. The classification models were explored based on an image data set using ConvNeXt_Tiny, ResNet-18, Densenet-121, ConvNeXt-Base, and EfficientNet-B1 deep learning models, which are widely used in the Fastai library and developed based on the transfer learning technique. Considering model accuracy and computational cost, using images with an input image size of 224*224 is efficient. Experimental results revealed that at the end of 20 iterations, the accuracy results of the models reached 95 %, 98 %, 99 %, 98 %, and 99 % for ResNet-18, ConvNeXt-Tiny, DenseNet-121, and ConvNeXt-Base and EfficientNet-B1 algorithms, respectively, and the models showed a tendency to converge. It is observed that the DenseNet-121 and EfficientNet-B1 models had the best accuracy rate. The precision, recall, and F1-score also support these results. The proposed models can potentially be used for non-invasive, effective, and rapid classification of dried fruits on the embedded system in a related application.