Detection of Mulberry Ripeness Stages Using Deep Learning Models

Abstract

Ripeness classification is one of the most challenging tasks in the postharvest management of mulberry fruit. The risks of microbial contamination and human error in manual sorting are significant; it may result in quality degradation and wasting of processed products. Due to advanced developments in computer vision and machine learning, automated sorting became possible. This study presents the results of developing and testing a computer vision-based application using convolutional neural networks (CNNs) for the classification of mulberry fruit ripening stages. To reduce the training cost and improve the accuracy of classification, transfer learning was used to fine-tune the CNN models. The CNN models in the test include DenseNet, Inception-v3, ResNet-18, ResNet-50, and AlexNet. Transfer learning was used to fine-tune the models and improve the accuracy of classification. The AlexNet and ResNet-18 networks exhibited the best performance with 98.32% and 98.65% overall accuracy for classifying the ripeness of white and black mulberries, respectively. Moreover, the performance of the models did not change when the data sets of both genotypes were mixed. The ResNet-18 was able to classify both genotype and ripeness from 600 fruit images in 2.36 min with an overall accuracy of 98.03%, which was superior to other architectures. It indicates that the model could be used for precise classification of the ripening stages of mulberries and other horticultural products, as a part of an automated sorting system.INDEX TERMS Convolutional neural network, computer vision, online detection, ripening classification, transfer learning.

Discrimination of foreign bodies in quinoa (Chenopodium quinoa Willd.) grains using convolutional neural networks with a transfer learning approach

“…6 . The accuracy reached for all models during training was over 95%, results comparatively similar to other works ad Miraei Ashtiani et al (2021) when reduced size products or their contaminants are discriminated. In the same figure, it is observed, with the exception of AlexNet , that there was no effect of color space on the statistical metric of model training; this could be due to improvements in the generalization capacity of new models such as Movilenet-V2 and DenseNet-201 as a comment ( Espejo-Garcia et al, 2020 ).…”

“…The latter is slightly better in training with less than 300 epochs, subsequently equaling his training metrics. What is more, according to the convergence of the precision and loss curves, it follows that they are inversely related to the complexity of the model, consistent with the proposed by Too et al (2019) , Chen et al (2021b) , Ciocca, Napoletano & Schettini (2018) and Miraei Ashtiani et al (2021) .…”

“…Finally, each combination CNN—color space was trained thirty times, in a hold-out cross-validation strategy, and F-score saved in concordance with the performed by Chen et al (2020) dividing 70% for training and 30% for validation. According to Miraei Ashtiani et al (2021) , the reason for splitting the dataset into two subsets is that in small datasets, the additional split might lead to a smaller training set which may be exposed to over-fitting. The effect of variables in F-score was considered when P value < 0.05, comparing it through Tukey -test.…”

Classification of jackfruit and cempedak using convolutional neural network and transfer learning

“…CNN was applied in a work that classified the ripeness of mulberry fruit with some fine-tuning to help improve the classification's accuracy [17]. From the five CNN models used, the AlexNet and ResNet-18 networks appeared to have the best performance, with ResNet-18 showing the most superiority.…”

Adulteration detection in minced beef using low-cost color imaging system coupled with deep neural network

“…The samples were divided into three subsets, the training set was 70% of all samples stratified by the labels, the validation set was 10% of the training set, and the testing set was 30% of all samples. The data set was divided into three subsets because, in small data sets, an additional split could result in a smaller training set that is more susceptible to overfitting (Miraei Ashtiani et al, 2021). The training and validation datasets are used together to develop the classification model and the testing is used solely for testing the final results.…”