Construction of apricot variety search engine based on deep learning

Abstract

Apricot has a long history of cultivation and has many varieties and types. The traditional variety identification methods are time-consuming and labor-consuming, posing grand challenges to apricot resource management. Tool development in this regard will help researchers quickly identify variety information. This study photographed apricot fruits outdoors and indoors and constructed a dataset that can precisely classify the fruits using a U-net model (F-score: 99%), which helps to obtain the fruit's size, shape, and color features. Meanwhile, a variety search engine was constructed, which can search and identify variety from the database according to the above features. Besides, a mobile and web application (Apricotview) was developed, and the construction mode can be also applied to other varieties of fruit trees. Additionally, we have collected four difficult-to-identify seed datasets and used the VGG16 model for training, with an accuracy of 97%, which provided an important basis for Apricotview. To address the difficulties in data collection bottlenecking apricot phenomics research, we developed the first apricot database platform of its kind (ApricotDIAP, http://apricotdiap.com/) to accumulate, manage, and publicize scientific data of apricot.

Headnote

ABSTRACT

Apricot has a long history of cultivation and has many varieties and types. The traditional variety identification methods are time-consuming and labor-consuming, posing grand challenges to apricot resource management. Tool development in this regard will help researchers quickly identify variety information. This study photographed apricot fruits outdoors and indoors and constructed a dataset that can precisely classify the fruits using a U-net model (F-score: 99%), which helps to obtain the fruit's size, shape, and color features. Meanwhile, a variety search engine was constructed, which can search and identify variety from the database according to the above features. Besides, a mobile and web application (Apricotview) was developed, and the construction mode can be also applied to other varieties of fruit trees. Additionally, we have collected four difficult-to-identify seed datasets and used the VGG16 model for training, with an accuracy of 97%, which provided an important basis for Apricotview. To address the difficulties in data collection bottlenecking apricot phenomics research, we developed the first apricot database platform of its kind (ApricotDIAP, http://apricotdiap.com/) to accumulate, manage, and publicize scientific data of apricot.

Keywords: Apricot; Variety; Convolutional neural network; Deep learning; Database platform; Mobile application; Image retrieval

1. Introduction

Apricot belongs to the Rosaceae family and Prunoideae sub-family. It has a wide plantation range and is mainly distributed in temperate areas of the northern and southern hemispheres (Groppi et al., 2021). After a long history of cultivation, many apricot varieties have evolved, with unique and stable morpho-logical appearances (Krichen et al., 2009). For example, 'Long-wangmao' is oblate-circle-shaped, 'Xiaobaixing' is yellowish-white, and 'Pepper' looks like pepper. Morphological characterization has proven to be useful for cultivar identification, selection, differentiation, quality and germplasm management (Krichen et al., 2014; Wang et al., 2021; Zhu et al., 2022).

In the traditional variety identification, researchers measure the fruit size using a digital calliper (Giménez et al., 2010) and then compare the measured data against hundreds of data resources to determine the variety. This method is timeconsuming, labor-intensive, and inconvenient for outdoor work. The convolutional neural network (CNN) is widely used in image segmentation and classification (Yang and Xu, 2021). The content-based image retrieval (CBIR) technology can transform objective measurement into size, color, and shape features (Mandlik and Sagare, 2014) and search with databases. In recent years, researchers have studied the application of these technologies in identifying horticultural crop varieties. Zhang et al. (2021) used a multiscale and robust method to extract topological features of leaf shapes, textures, and vein details, and then combined these features with a CNN to classify 88 cherry varieties with an accuracy of 83.52%. Liu et al. (2021) trained CNN models to classify grape leaves from 21 varieties and found that the best model was GoogLeNet with an accuracy of 99.91%.

Content-based image retrieval (CBIR) is a rapidly developing forward-looking technology (Mandlik and Sagare, 2014). Users can quickly find target information from the database using CBIR, which is convenient for the management of fruit germplasm. CBIR first obtains the features of images including color, shape, and size, and then matches the features of the input image with the image features in the database to search for similar images (Piras and Giacinto, 2017). However, the image retrieval process is affected by background noises, and this problem can be addressed using the convolutional neural network models (Vijayan et al., 2021).

The convolutional neural network (CNN) is a type of selflearning feedforward neural network with convolution calculation. It is one of the representative algorithms of deep learning (Gao et al., 2021). CNN has been extensively applied to image classification and segmentation (semantic segmentation and instance segmentation) (Yang and Xu, 2021). Image classification divides images into different categories according to their characteristic information. The first image classification model AlexNet was proposed in 2012. Compared with traditional methods, AlexNet significantly improves the accuracy of image classification (Krizhevsky et al., 2012). Inspired by AlexNet, a variety of network models have been developed, including VGGNet (Tu et al., 2018; Naşiri et al., 2019), GoogLeNet (Brahimi et al., 2017), and many others. Image segmentation can extract the object's feature information from the image and help to remove background noises. The commonly used modeling algorithms include U-net (Ronneberger et al., 2015) and Mask-RCNN (He et al., 2017). U-net can segment image details, especially for small data sets, and its effect is better than traditional segmentation methods (Tassis et al., 2021). At the same time, it also has high speed and high resolution of images (Li et al., 2021), which is suitable for the development and use of tools.

In this study, a large number of photos of apricot fruits and seeds were collected. Then, a method for constructing variety search engine is proposed based on the fruit size, shape, and color features in combination with the U-net deep learning model, and the VGG16 deep learning model was used for seed classification research. Finally, the apricot variety retrieval tool ApricotView, and the apricot dataset platform ApricotDIAP are developed.

2. Materials and methods

2.1. Plant materials

Apricot fruits were collected in experimental farms and orchards from Xinjiang (Yining, Huocheng, and Xinyuan), Inner Mongolia (Chifeng), Gansu (Dunhuang), Qinghai (Haidong), and Henan (Xinxiang and Mengzhou) in China, from 2018 to 2020. The fruits (six directions) and seeds were placed on a black-, white- or blue-background sheet for taking photos indoors, and the fruits were also photographed from six directions outdoors with a digital camera (SONY DSC-HX400, China) (Fig. 1). In addition, we also collected four seed datasets including 'Yiwofeng', 'Xiao-bianxing', 'Aketoyong x Pearl' and 'Pi'naizi x Pearl' (which are available via ApricotDIAP).

2.2. Image processing and model construction

2.2.1. Semantic segmentation

A total of 1588 fruit images of 207 germplasms were manually annotated using the labelme annotation tool. The pixels corresponding to the fruit were obtained using this tool, and the foreground was annotated. A python script was used to convert the annotations from labelme into a JSON file. All images were randomly split into a training dataset (90% of images) and a test dataset (10% of images). Then, the dataset was used to train the U-net model, which contained two stages: a contraction stage and an expansion stage. The contraction stage was constructed using the 16-layer deep convolutional neural network, and the overall structure of the U-net is shown in Fig. 2. The following parameters were set: the learning rate of 10 4, the epochs of 50, and the transfer learning strategy.

2.2.2. Classification training

A total of 1842 images from the four seed datasets were used for classification training. First, the image is set to 224 x 224 size by PIL's Image function and then normalized, and then trained using the VGG16 model. The VGG16 model contains 13 convolutional layers with 3x3 kernels and five 2x2 max-pooling layers (Fig. 3). Divide all datasets into 70% training set and 30% validation set. The following parameters were set: the learning rate of 10 4, the epochs of 50, and the transfer learning strategy.

2.3. Fruit size

Fifty apricot resources (Fig. 4) were used for size analysis. Three linear dimensions of apricots, length (L), width (W), and thickness (T), were measured with the digital calliper (0-200 mm ± 0.03 mm, DL311200, China). For each apricot resource, 5 samples were measured. Based on the three linear dimensions, nine indicators [Arithmetic mean diameter (Da/mm), Geometric mean diameter (Dg/mm), Equivalent diameter (De/mm), Square mean diameter (Se/mm), Shape index (SI), Sphericity (ф), Aspect ratio (Ra/%), Surface area (S/mm2) and Volume (V/mm3)] (Souty and Caillavet, 1950; Mohsenin, 1986; Jain and Bal, 1997; Altuntas et al., 2005; Mirzaee et al., 2008)were transformed. All data of the indicators were saved in the MySQL database.

2.4. Screening of indicators/parameters

In order to screen valuable indicators and parameters and remove redundancy, we define the "sample overlap", that is, the sum of the number of overlaps between the maximum and minimum values of samples under single or multiple indicators and parameters. The first, each sample in the database is used to establish a community according to the indicator to collect other samples that overlap with it. Second, make a dictionary for each indicator and collect the community of each sample under that indicator. Finally, the number of indicators is categorized into different levels, ranging from low to high, and they are randomly combined. The minimum sample overlap for each level is then calculated. When the sample overlap is no longer reduced, the corresponding indicators are used as the valuable indicator.

sample^ = [sample^,...]

samplef represents the community of sample i under the indicator k, and the samples within the community have overlap with sample i.

indicator = {samplei : [samplea,...], samplej : [sampleb,...],...}

Indicator represents the dictionary of indicator k, which contains the community of each sample under this indicator.

(ProQuest: ... denotes formula omitted.)

summin is the minimum value of the "sample overlap" calculated by the random combination of a specific number of indicators, and the corresponding indicators are the effective indicator under this number.

2.5. Construction of fruit image search engine

The fruit size, shape, and color were used for variety search. The fruit length, width, and thickness were obtained from fruit feature comparison after U-net model segmentation. The color pictures and single-channel pictures of fruits were obtained using python through the U-net model. The contours of the fruit were obtained using findcontours (Bradski and Kaehler, 2000). Then, the fruit size and related parameters were compared with the database, and the interval between the maximum and minimum values and the intersection between the parameters of each variety were taken as the selection criteria. The similarity of the contours was compared using matchshapes, which compares the similarity of the two shapes by calculating the Hu invariant moments distance of the image. Hu invariant moments (Hu, 1962) are robust against rotation, translation, and scaling. The similarity of fruit color was compared using color histograms (Widiyanto et al., 2019), which is not affected by image rotation and translation.

2.6. Evaluation criterion

Comprehensive evaluation index (F-score) (Afonso et al., 2020) and loss were used to evaluate the U-Net model results. Two pictures were randomly selected from each type, and each picture was expanded to five pictures through rotation, scaling, movement, and brightness (Fig. 5; Fig. SI) to test color and shape features. Meanwhile, new pictures of five new resources and five existing resources were used to test the search engine.

Precision = TP/(TP+FP).

Recall = TP/(TP+FN).

F-score = 2 x Precision x Recall/(Precision + Recall).

TP is the number of true positives, FP is the number of false positives, and FN is the number of false negatives.

2.7. The development tool and dataset platform

To help researchers better manage apricot germplasm resources and promote the development of apricot phenomics, this study developed a variety retrieval application called Apricot-View based on fruit images and an apricot dataset platform called ApricotDIAP. The Apricotview was designed and built using MUI, which can be downloaded from ApricotDIAP (http://apricotdiap. com/tools). The user-friendly interfaces of ApricotDIAP were developed using JavaScript, HTML5, and CSS3. The Flask of Python Web framework was used to manage the back-end, and Python programming was used to process images and output results. The cloud server was provided by Ali Cloud (Ali, Hong Kong, China), which is configured with Ubuntu Sever, 4 vCPU, and 8 GB memory.

3. Results

3.1. The structure of the apricot identification engine

The structure of the apricot identification engine is shown in Fig. 6. First, the background information of the input picture is removed using the U-Net model. Then, the color and shape features of the fruit are obtained, and the length, width, and thickness of the fruit are obtained by comparison with the reference, nine indicators are calculated. Next, the indicators are filtered and compared with the database to screen varieties. Subsequently, the color and shape features of the fruit are compared with the screened varieties to obtain the names of the sort and sent to the mobile phone(Android 8.0+). Finally, the detailed information on the apricot fruit is displayed interactively in the ApricotView.

3.2. Image semantic segmentation by deep learning

In this study, the U-net model was used to train semantic segmentation through fine-tuning transfer learning. The transfer learning strategy was chosen because it can accelerate training and improve accuracy at the same time (Alom et al., 2019). Fig. 7 shows the F-score and loss values of training and validation at each epoch. The F-score reached 97% after one epoch. Then, it continued to increase with epochs and reached a plateau after approximately 21 epochs. Overall, the semantic segmentation of the fruit images achieved an F-score of 99.43%. The semantic segmentation of the fruit images with the size reference object also achieved an F-score of 99.46%.

3.3. Determination of the indicators related to fruit size

Twelve indicators of 50 fruit resources were obtained (Fig. 8, C; Fig. SI), the fruit length (L) is 20.42-54.03 mm, the fruit width (W) is 18.62-53.48 mm, and the thickness (T) is 15.22-56.87 mm. There is a high correlation among indicators [length (L), width (W), thickness (T), Square mean diameter (Se), Surface area (S), Volume (V), Geometric mean diameter (Dg) and Equivalent diameter (De); arithmetic mean diameter (Da), Sphericity (ф) and Shape index (SI)], but the overlap between individuals changes (Fig. 8, A, C; Fig. S2). With the increase of the number of indicators, the minimum value of sample overlap showed a downward trend, and stabilized when the number of indicators reached 6 (Fig. 8, B). Finally, we identified six valuable indicators: L, W, T, aspect ratio (Ra), Da, and ф.

3.4. Effectiveness evaluation of the search engine

The main metric used for performance evaluation is the top-N retrieval rates, indicating whether the correct plant type is among the top-N returned images (Mandlik and Sagare, 2014). The color and shape were evaluated using the top-ten retrieval rate. The results indicate that, except for one surface image which could not be retrieved within the top 10, the others could be successfully located (Fig. SI). Meanwhile, five germplasms that do not measure the fruit size and five new germplasms were selected to evaluate the search engine. The results showed that it can make accurate identifications in the existing resources (four topi and one top2), and similar resources were retrieved about new resources (three were matched) (Fig. S3).

3.5. ApricotView performance

The development of convenient tools can help researchers collect and manage apricot resources. In this study, the apricot variety retrieval App called ApricotView was designed and developed (Fig. 9). Users can upload fruit pictures by taking photos or locally select the type for retrieval. ApricotView will return the name or the number of the top ten matches. Other detailed information, including the pictures of fruits, flowers, tree shapes, and color comparison, can be obtained by clicking the name or the number.

3.6. Variety identification based on VGG16 model

For variety identification of VGG16 model, the accuracy and loss values of training and validation at each epoch are shown in Fig. 10. The curves of validation accuracy and loss during training and validation reached a plateau after approximately 26 epochs. The final accuracy value reaches 97.79% (Fig. 10, A), and the loss value is 0.0721 (Fig. 10, B). The visualization of the prediction of classified seed images was obtained using Grad-CAM. The results show that VGG16 model distinguishes different cultivars mainly through the middle and base regions of the seeds (Fig. 10, C-F).

3.7. Comparison with traditional variety identification

We extracted the fruit, seed and kernel pictures of 50 resources from the search engine database for researchers' identification, and have extracted 10 pictures (not participating in training) from each of the 4 seed datasets for manual judgment and model discrimination. In the judgment of 50 resources, researchers can only determine the information of 5-10 resources, which reflects the advantages of search engines with large resources and saving rare resources. In the identification of different varieties of seeds (Fig. 11), both achieve an accuracy of 100%. However, in terms of difficult-to-identify materials, manual judgments have great fluctuations, while the deep learning model has the advantages of high accuracy and stability (Fig- И).

3.8. Development of the apricot dataset platform

The first apricot datasets sharing platform called ApricotDIAP (http://apricotdiap.com/) was developed to collect and publish apricot-related phenotypic datasets, which will promote the development of apricot phenomics (Fig. 12). ApricotDIAP consists of the tools module, download module, upload module, and help module. The tools module collects and releases the latest phenotypic tools and models of apricot. The download module contains apricot datasets that users can download for free. The upload module can facilitate researchers to share the datasets of apricot.

4. Discussion

Apricot has a long cultivation history, and many varieties have evolved (Krichen et al., 2009). The morphological research is useful for apricot germplasm management. In this study, the photos of the apricot fruit (six directions) were collected, and a fruit image search engine was built according to the fruit size, shape, and color features. To eliminate the interference of background, the U-net deep learning model was used for fine segmentation of the apricot fruit. Five unmeasured fruit sizes and five new resource images were randomly selected to test the accuracy of the search engine. The results show good accuracy and robustness (four topi and one top2 in unmeasured fruits; three are matched in new resources).

Although there are morphological differences between varieties, the morphology of apricot fruit still varies slightly within the cultivar. Therefore, keep multiple images for each cultivar can help improve retrieval accuracy. Considering the size of the fruit can significantly improve retrieval accuracy. CNN models such as U-net can be used to accurately segment and obtain the size features of apricot fruit in complex environments. In order to screen valuable indicators/parameters and remove redundancy, we defined "sample overlap" to extract effective indicators and added it to ApricotView. With the increase of materials and indicators, the ApricotView will automatically analyze and optimize indicators. Compared with traditional methods, the developed convenient tools using Content-based image retrieval and deep learning technology will help to obtain phenotypic data and determine varieties quickly and accurately, reduce the consumption of human and material resources, and minimize errors caused by human factors.

The deep learning classification models work well in fruit identification. Osako et al. (2020) used the VGG16 model to achieve fine identification of four litchi varieties. Masuda et al. (2021) used four CNN models to conduct a non-invasive study on the seedless fruit of persimmon and found that the VGG16 model achieved the highest accuracy. In this study, we used VGG16 model to classify three apricot varieties and two families of the same variety, which achieved accurate classification(accuracy, 97.79%) and provided a basis for ApricotView improvements. Although we have collected a lot of data about apricots, there is still less for each variety. In future, we will collect more datasets on apricot varieties, and manage the data by detailed division and management based on tree age, region, mutation material and other information. As the data continue to increase, we will improve the ApricotView, set up different material combinations, reduce material usage while maintaining high precision, and establish stable indicators.

In this study, a variety search App ApricotView was developed. ApricotView can help researchers manage apricot resources easily and discover new varieties. Apricot genomics has developed rapidly in recent years (Jiang et al., 2019; Groppi et al., 2021; Zhang et al., 2021), but the advancement of phenomics has been slow. The development of phenomics is limited by the difficulty of obtaining datasets. ApricotDIAP will assist researchers in obtaining apricot-related datasets for scientific research and promote the development of apricot phenomics.

5. Conclusions

In this study, a search engine was developed for apricot fruit varieties to achieve accurate detection results. Meanwhile, ApricotView App was developed to quickly search varieties indoors or outdoors, which can help the growers to manage apricot germplasms. Besides, the apricot dataset sharing platform (ApricotDIAP) was created to facilitate the development of phenomics. ApricotDIAP will keep providing updated resources of various datasets, tools, and models, serving as an active platform for apricot researchers.

Declaration of interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This study was supported by the Fundamental Research Funds for the Central Non-profit Research Institution of the Chinese Academy of Forestry (Grant No. CAFYBB2020ZY003), the Key S&T Project of Inner Mongolia (Grant No. 2021ZD0041-001002) and the Central Public-interest Scientific Institution Basal Research Fund (Grant No. 11024316000202300001).

Supplementary materials

Supplementary data to this article can be found online at https://doi.Org/10.1016/j.hpj.2023.02.007.

Sidebar

Received 4 August 2022; Received in revised form 22 September 2022; Accepted 2 February 2023

Available online 18 February 2023

Footnote

* Corresponding authors.

E-mail addresses: [email protected]; [email protected]

Peer review under responsibility of Chinese Society of Horticultural Science (CSHS) and Institute of Vegetables and Flowers (IVF), Chinese Academy of Agricultural Sciences (CAAS)

References

REFERENCES

Afonso, M., Fonteijn, H., Florentin, F.S., Lensink, D., Mooij, M., Faber, N., Polder, G., Wehrens, R., 2020. Tomato fruit detection and counting in greenhouses using deep learning. Front Plant Sei, 11: 571299.

Alom, M.Z., Taha, T.M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M.S., Hasan, M., Essen, B.V., Awwal, A., Asari, V.K., 2019. A state-of-the-art survey on deep learning theory and architectures. Electronics, 8: 292.

Altuntas, E., Ozgoz, E., Taser, O.F., 2005. Some physical properties of fenugreek (Trigonella foenum-graceum L.) seeds. J Food Eng, 71: 37-43.

Bradski, G., Kaehler, A., 2000. Opencv library. Doctor Dobbs, 25: 120-126.

Brahimi, M., Boukhalfa, K., Moussaoui, A., 2017. Deep learning for tomato diseases: classification and symptoms visualization. Appl Artif Intell, 31: 299-315.

Gao, M., Qi, D., Mu, H., Chen, J., 2021. A transfer residual neural network based on resnet-34 for detection of wood knot defects. Forests, 12: 212.

Giménez, G., Dussi, M.C., Reeb, P., Zon, K., Nyéki, J., Szabó, Z., Racsko, J., 2010. Fruit growth and abscission pattern of 'williams' pear treated with benzyladenine. Acta Hortic, 884: 481-490.

Groppi, A., Liu, S., Cornille, A., Decroocq, S., Decroocq, V., 2021. Population genomics of apricots unravels domestication history and adaptive events. Nat Commun, 12: 3956.

He, K., Gkioxari, G., Dollár, P., Girshick, R., 2017. Mask R-CNN. IEEE Trans Pattern Anal Mach Intell, 42: 386-397.

Hu, M.K., 1962. Visual pattern recognition by moment invariants. Information theory, 8: 179-187.

Jain, R.K., Bal, S., 1997. Physical properties of pearl millet. J Agric Eng Res, 66: 85-91.

Jiang, F., Zhang, J., Wang, S., Yang, L., Wang, Y., 2019. The apricot (Prunus armeniaca L.) genome elucidates rosaceae evolution and beta-carotenoid synthesis. Hortic Res, 6: 128.

Krichen, L., Audergon, J.M., Trifi-Farah, N., 2014. Variability of morphological characters among tunisian apricot germplasm. Sei Hortic, 179: 328-339.

Krichen, L., Trifi-Farah, N., Marrakchi, M., Audergon, J.M., 2009. Evaluation of the current apricot variability in Tunisia- comparison with previously described cultivars. Acta Hortic, 814: 113-119.

Krizhevsky, A., Sutskever, I., Hinton, G., 2012. Imagenet classification with deep convolutional neural networks. Adv Neural Inf Process Syst, 25: 1097-1105.

Li, Q., Jia, W., Sun, M., Hou, S., Zheng, Y., 2021. A novel green apple segmentation algorithm based on ensemble U-Net under complex orchard environment. Comput Electron Agric, 180: 105900.

Liu, Y.X., Shen, L., Su, J.Y., Lu, N., Fang, Y.L., Liu, F., Song, Y.Y., Su, B.F., 2021. Development of a mobile application for identification of grapevine (Vitis vinifera L.) cultivars via deep learning. Int J Agric Biol Eng, 14: 172-179.

Mandlik, V.S., Sagare, U.S., 2014. Nursery flower plant image retrieval system using color & shape feature. IJETCAS, 10: 2279, 0055.

Masuda, K., Suzuki, M., Baba, K., Takeshita, K., Akagi, T., 2021. Noninvasive diagnosis of seedless fruit using deep learning in persimmon. Horticulture J, 90: 2.

Mirzaee, E., Rafiee, S., Keyhani, A.R., Djomeh, Z.E., Kheiralipour, K., 2008. Mass modeling of two varieties of apricot (Prunus armenaica L.) with some physical characteristics. Plant Omics, 1: 37-43.

Mohsenin, N.N., 1986. Physical properties of plant and animal materials. J Biomed Mater Res B, 11: 83-89.

Naşiri, A., Taheri-Garavand, A., Zhang, Y.D., 2019. Image-based deep learningautomated sorting of date fruit. Postharvest Biol Technoi, 153: 133-141.

Osako, Y., Yamane, H., Lin, S.Y., Chen, P.A., Tao, R., 2020. Cultivar discrimination of litchi fruit images using deep learning. Sei Hortic, 269: 109360.

Piras, L., Giacinto, G., 2017. Information fusion in content based image retrieval: a comprehensive overview. Inf Fusion, 37: 50-60.

Ronneberger, О., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. Springer International Publishing, Cham, pp. 234-241.

Souty, J., Caillavet, H., 1950. Monographie des principales variétés de pêchers. Fruits d'Outre-Mer, 5: 379-381.

Tassis, L.M., Tozzi, J.E., Krohling, R.A., 2021. A deep learning approach combining instance and semantic segmentation to identify diseases and pests of coffee leaves from in-field images. Comput Electron Agric, 186: 106191.

Tu, S., Xue, Y., Chan, Z., Yu, Q., Liang, M., 2018. Detection of passion fruits and maturity classification using RedGreen-Blue Depth images. Biosyst Eng, 175: 156-167.

Vijayan, M., Raguraman, P., Mohan, R., 2021. A fully residual convolutional neural network for background subtraction. Pattern Recogn Lett, 146: 63-69.

Wang, Z.J., Luo, G.Q., Dai, F.W., Xiao, G.S., Lin, S., Li, Z.Y., Tang, C.M., 2021. Genetic diversity of 569 fruit mulberry germplasm resources based on eight agronomic traits. Acta Hortic Sin, 48: 2375-2384.

Widiyanto, W.W., Kusrini, K., Fatta, H.A., 2019. Searching similarity digital image using color histogram. Techno, 20: 53.

Yang, B., Xu, Y., 2021. Applications of deep-learning approaches in horticultural research: a review. Hortic Res, 8: 123.

Zhang, Q., Zhang, D., Yu, K., Ji, J., Liu, N., Zhang, Y., Xu, M., Zhang, Y.J., Ma, X., Liu, S., 2021. Frequent germplasm exchanges drive the high genetic diversity of Chinese-cultivated common apricot germplasm. Hortic Res, 8: 215.

Zhang, Y., Peng, J., Yuan, X., Zhang, L., Zhu, D., Hong, P., Wang, J., Liu, Q, Liu, W., 2021. Mfcis: an automatic leaf-based identification pipeline for plant cultivars using deep learning and persistent homology. Hortic Res, 8: 172.

Zhu, P., Meng, Y.J., Zhang, K.J., Wang, X.Y., Liang, K., Wang, T.T., Xu, J., Qin, X.D., Wu, Z., Cheng, C.Y., Yu, X.Q, Qian, C.T., Lou, Q.F., Chen, J.F., Li, J., 2022. Mapping of fruit apex shape related QTLs across multi-genetic backgrounds in cucumber (Cucumis sativus L.). Hortic Plant J, 8: 328-340.