1. Introduction
There is a precise correlation between the maturity and fundamental quality of the fruit. The maturity of the fruit at harvest has an important impact on the softening and texture change of the fruit [
1]. For the appearance of the fruit, height, weight and color changed with the progress of maturity [
2]. For the content of compounds contained in the fruit, the maturity of the fruit will affect the content of total soluble solids and total acid [
3]. For the above reasons, the fundamental quality of fruit can be effectively maintained by choosing the optimal harvest maturity [
4]. The researchers selected different indicators as evaluation criteria for different fruit maturity to grade the fundamental quality. Zhu et al. proposed following three maturity levels for grading
Camellia oleifera fruit samples: unripe, ripe, and overripe, and the results indicate that there were significant differences in seed oil content, seed soluble protein content, seed soluble sugar content, seed starch content, dry seed weight, and moisture content among the maturity stages graded [
5]. Iqbal and Hakim utilized appearance features such as shape, texture, color, and size to grade eight different cultivars of harvested mangoes for extra class, class I and class II [
6]. Ma et al. used the color and sweetness values of the three different parts of banana to classify the maturity of the fruit for six stages [
7].
Deep learning has brought about a significant transformation in conventional internet-based industries, including web search and advertising. In addition, it has enabled the development of new products and businesses by assisting people in various domains such as healthcare, education, agriculture, and even autonomous driving [
8]. For instance, deep learning has significantly improved the accuracy of predicting subcellular protein localization [
9]. Notably, deep learning has been instrumental in advancing computer vision, which has witnessed rapid progress in recent times. He et al. introduced the influential Residual Networks, which have remained the gold-standard since the inception of the ResNet architecture [
10]. Residual Networks are considered easy to optimize and can alleviate the problem of gradient disappearance that is caused by increasing depth in deep neural networks. As a result, ResNet is often used as the default architecture in studies or as a baseline for comparison when new architectures are proposed [
11]. Further, Duta et al. proposed an improved version of ResNet. The proposed enhancements tackle the three key elements of a ResNet—namely, the flow of information across network layers, the residual building block, and the projection shortcut. These enhancements can ensure improved network accuracy and learning astringency [
12].
The convolutional neural network is commonly utilized in the field of agriculture [
13]. Mamat et al. utilized YOLO net for automated annotation of fruit images. The mean Average Precision achieved for oil palm fruit was 98.7%, while the accuracy for fruit variety classification reached 99.5% [
14]. Gulzar implemented TL-MobileNetV2 for the classification of forty different types of fruits. The model achieved an accuracy of 99%, a recall of 99%, and a f1 score of 99% [
15]. Osako et al. used a pre-trained VGG16 model to develop a cultivar classification system for litchi fruit. The model achieved an accuracy of 98.33% [
16]. Chen et al. photographed images of apricot fruits in both outdoor and indoor settings, and created a dataset that enables accurate fruit classification using a U-net model. The model achieved an impressive F-score of 99%. Furthermore, the researchers collected four datasets containing seeds that are challenging to identify, and trained a VGG16 model to classify them, achieving an accuracy of 97% [
17]. Suzuki et al. proposed the use of a deep neural network to predict the occurrence of a severe fruit disorder in persimmon, namely rapid over-softening, using simple RGB images. Their research revealed that all the CNN models examined were successful in binary classification of the rapidly over-softened fruits and controls, reaching an accuracy of over 80% across various criteria [
18]. Unal and Aktas used Inception V3 model and EfficientNet model to accurately classify hazelnut kernels, with an accuracy of 97.85% and an accuracy of 99.28% respectively [
19].
Winter jujube (
Ziziphus jujuba Mill. cv. Dongzao) is an excellent late maturing fresh-eating variety in China [
20]. Tree-ripened winter jujube is highly appreciated by consumers for its taste, flavor and sweetness [
21]. Regrettably, such fruits have a brief postharvest lifespan, primarily due to rapid softening, resulting in increased vulnerability to mechanical damage and the development of decay [
22]. The traditional maturity classification methods such as mechanical testing can easily cause damage to the surface of winter jujube [
23]. Hence, the classification of winter jujube based on deep learning is an efficient and non-destructive way to solve this problem. Lu et al. used YOLOv3 algorithm to train models, with an accuracy of 97.28%. Additionally, they designed an automatic winter jujube classification robot based on computer vision [
24]. Al-Saif et al. devised a technique for distinguishing between various cultivars of Indian jujube fruits by utilizing a single fruit’s color and morphological characteristics and training an artificial neural network classifier. Their approach achieved an accuracy rate of 97.56% [
25]. Feng et al. used hyperspectral imaging with pixel-wise deep learning method to detect subtle bruises on winter jujube. Their study found that the CNN model based on all geographical origins performed the best, with most accuracy surpassing 85% [
26]. Despite some progress in research in this field, there are still challenges to overcome, including insufficient accuracy and a limited number of categories. In this study, we classified the maturity of winter jujube into five levels [
27,
28]. The puncture force, TSS, and TA content of winter jujube was tested at each maturity level to determine the optimal consumption period of winter jujube [
29]. We used the automatic recognition and classification of each maturity of winter jujube images by deep learning model. We took image data to train ResNet-50, iResNet-50, and designed the improved network model base on iResNet-50 to increase the accuracy of winter jujube classification. By implementing a more refined categorization approach, our model has acquired the capability to automatically classify the maturity levels of fruits and identify preliminary defects.
5. Conclusions
Our study confirms the importance of proper maturity grading in the production and marketing of winter jujube fruit. By selecting the optimal picking period and implementing effective post-harvest management techniques, farmers and other stakeholders in the industry can ensure that winter jujube fruit meets consumer expectations for flavor, texture, and shelf life. Five automatic maturity detection models were developed for winter jujube using three different algorithms. After thorough comparison, we found that the model based on the improved iResNet-50 algorithm with double residuals superimposed in the first Main Stage achieved the best performance. Specifically, the model achieved an accuracy of 98.35%, an average precision of 98.40%, and an average recall of 98.35% on the test set of this experiment. Moreover, the average F1 score of the model was 98.36%, which demonstrated the feasibility and effectiveness of the improved iResNet-50 based algorithm for automatic grading of winter jujube maturity. Our model facilitates the automated classification of winter jujube after harvesting. By incorporating the classification of “softened fruits” into the model, it expands beyond maturity classification and enables preliminary identification of damaged fruits. This enhancement enhances the model’s capabilities and improves its utility in post-harvest fruit classification. Overall, the findings of this study offer important insights into the use of deep learning techniques for automatic grading of agricultural products, and can provide valuable guidance for the development of similar applications in the future. However, it should be noted that the performance of the models may vary depending on the dataset and the specific task at hand, and further optimization and validation may be necessary to achieve optimal results in practical scenarios.