Introduction Jujube (Ziziphus jujuba Miller) is an economically and nutritionally valuable fruit crop, particularly in arid and semi-arid regions. Drying is a critical postharvest operation that extends shelf life, facilitates storage and transportation, and increases the commercial value of jujube. However, precise determination of the drying status is essential for achieving consistent product quality. Inadequate drying may result in excessive residual moisture and reduced storage stability, whereas over-drying can cause undesirable changes in product quality and unnecessary energy consumption. Conventional approaches for determining drying status, including moisture measurement, weight loss, and visual inspection, are often time-consuming, destructive, or dependent on operator expertise. Computer vision combined with deep learning offers a rapid and non-destructive alternative for automated assessment of agricultural products. Although previous studies have demonstrated the effectiveness of deep learning for jujube maturity grading, defect detection, quality classification, and non-destructive evaluation, relatively limited attention has been given to the direct classification of jujube drying statuses using RGB images. Therefore, this study aimed to develop a dedicated deep convolutional neural network, termed DCNN_Jujube, for automatic classification of different drying statuses of Ghaenat jujube kernels and to compare its performance with transfer-learning and conventional machine-learning approaches. In this study, drying was performed using a hot-air dryer at 70 °C with an air velocity of 1 m/s.
Materials and Methods A dataset comprising 1,840 images of Ghaenat jujube kernels was established and categorized into four drying-status classes: fresh, semi-dried, dried, and excessively dried. The dataset consisted of 1,280 training images and 560 validation images, while an independent test set of 160 images was used for final performance evaluation. All images were cropped and resized to 512 × 512 pixels. To increase training-data variability and mitigate overfitting, augmentation techniques, including flipping, cropping, translation, scaling, and rotation, were applied. The proposed DCNN_Jujube architecture was specifically designed for the target four-class classification problem. The network comprised ten convolutional blocks and two fully connected blocks and incorporated batch normalization, ReLU activation, max-pooling, and dropout layers. A Softmax layer was used for final classification. For comparative evaluation, two pretrained architectures, InceptionV3 and InceptionResNetV2, were fine-tuned using the jujube dataset. In addition, conventional machine-learning methods based on handcrafted HOG and SIFT features were evaluated using support vector machine (SVM), k-nearest neighbors (KNN), and logistic regression classifiers. Model performance was evaluated using accuracy, precision, recall, F1-score, confusion matrix, and receiver operating characteristic (ROC) analysis.
Results and Discussion The proposed DCNN_Jujube achieved the highest classification performance, with an overall test accuracy of 97.00%. The corresponding precision, recall, and F1-score were 96.91%, 96.88%, and 96.88%, respectively. These results demonstrate the strong capability of the proposed architecture to discriminate among the four drying statuses. The fine-tuned transfer-learning models also achieved high classification performance. InceptionV3 and InceptionResNetV2 obtained test accuracies of 95.63% and 96.25%, respectively. Although both pretrained models provided competitive results, their performance was slightly lower than that of DCNN_Jujube. This finding suggests that a task-specific architecture can learn visual representations more closely associated with the characteristics of jujube drying than general-purpose pretrained models. In contrast, the conventional machine-learning approaches based on HOG and SIFT features achieved lower classification accuracies, ranging approximately from 85.00% to 92.50%. The lower performance of these methods may be related to the limitations of handcrafted features in representing the complex and progressive changes in color, texture, shape, shrinkage, and surface appearance that occur during drying. In comparison, the proposed convolutional network automatically learned hierarchical and discriminative features directly from the images. The confusion matrix demonstrated that the majority of samples were correctly assigned to their corresponding drying statuses. Some misclassification between adjacent statuses was observed, which is reasonable because drying is a continuous process and the visual differences between consecutive statuses may be subtle. The ROC analysis further confirmed the strong discriminative capability of DCNN_Jujube. Overall, the findings demonstrate that RGB image-based deep learning can provide an effective and relatively simple approach for non-destructive classification of jujube drying statuses.
Conclusion A dedicated deep convolutional neural network, DCNN_Jujube, was successfully developed for the automatic classification of four drying statuses of Ghaenat jujube kernels using RGB images. The proposed model achieved a test accuracy of 97.00%, with precision, recall, and F1-score of 96.91%, 96.88%, and 96.88%, respectively, and outperformed both fine-tuned transfer-learning models and conventional machine-learning approaches based on handcrafted features. The results demonstrate that a task-specific deep-learning architecture can effectively extract discriminative visual characteristics associated with progressive jujube drying and provide accurate multi-class classification. The proposed RGB-based approach has potential as a rapid, non-destructive, and relatively cost-effective tool for drying-status assessment, online quality monitoring, and intelligent process control. Further research should evaluate the model using larger and more diverse datasets covering different cultivars, drying conditions, illumination environments, and imaging configurations, followed by validation under real industrial processing conditions.