Prerpints.org logo
Preprint
Article

Classification of Appearance Quality of Red Grape Based on Transfer Learning of Convolution Neural Network

Altmetrics

Downloads

169

Views

45

Comments

0

A peer-reviewed article of this preprint also exists.

Submitted:

17 May 2023

Posted:

17 May 2023

You are already at the latest version

Alerts
Abstract
Grapes are a globally popular fruit, with grape cultivation worldwide being second only to citrus. This article focuses on the low efficiency and accuracy of traditional manual grading of red grape external appearance and proposes a small-sample red grape external appearance grading model based on transfer learning with convolutional neural networks (CNNs). Initially, the CNN transfer learning method was used to transfer the pre-trained AlexNet, VGG16, GoogleNet, InceptionV3, and ResNet50 network models on the ImageNet image dataset to the red grape image grading task. By comparing the classification performance of the CNN models of these five different network depths with fine-tuning, ResNet50 with a learning rate of 0.001 and a loop number of 10 was determined to be the best feature extractor for red grape images. Moreover, given the small number of red grape image samples in this study, different convolutional layer features output by the ResNet50 feature extractor were analyzed layer by layer to determine the effect of deep features extracted by each convolutional layer on SVM classification performance. This analysis helped to obtain a ResNet50+SVM red grape external appearance grading model based on the optimal ResNet50 feature extraction strategy. Experimental data showed that the classification model constructed using the feature parameters extracted from the 10th node of the ResNet50 network achieved an accuracy rate of 95.08% for red grape grading. These research results provide a reference for the online grading of red grape clusters based on external appearance quality and have certain guiding significance for the quality and efficiency of grape industry circulation and production.
Keywords: 
Subject: Computer Science and Mathematics  -   Computer Vision and Graphics

1. Introduction

Red globe grapes, also known as red earth grapes, are a popular variety of fresh eating grapes in China due to their large size, vibrant color, and ability to be stored for extended periods of time. They are the second most widely cultivated variety of fresh eating grapes in the country[1]. Fruit appearance is the most intuitive quality characteristic of fruit and accounts for 60% of the evaluation of fruit quality[2], directly affecting market prices and consumer purchasing desire. Currently, the grading of red globe grapes is mainly based on subjective human classification, which is prone to subjective factors, high cost, low efficiency, and lacks effective testing methods[3]. Therefore, there is an urgent need to achieve rapid and non-destructive grading of the appearance quality of red globe grapes to improve their commercial yield.
To address the aforementioned issues, domestic and foreign researchers have explored the use of machine vision technology. Chen Ying et al.[4] used three methods, namely the fruit surface coloring rate calculated in the RGB color space, the projection area method for calculating the size of grape clusters, and the projection curve method for calculating shape parameters of grape axes, to grade the appearance quality of 20 clusters of Kyoho grapes. The grading accuracy rates were 88.3%, 90.0%, and 88.3%, respectively. Li Junwei et al.[5] used machine vision technology to predict and grade the quality and diameter of individual grapes from seedless white and red globe grapes in Xinjiang, achieving a grading accuracy rate of over 85%. Xiao Zhuang et al.[6] used the random least-squares ellipse detection method to extract grape size and grade 42 clusters of red globe grapes, with a grading accuracy rate of 90.48%. The aforementioned studies all used single features as grape grading indicators, requiring image preprocessing and the selection of artificially extracted shallow features, which lacked robustness.
In recent years, deep learning has emerged as a modeling method based on computers[7]. Among them, convolutional neural networks are directly driven by data, enabling self-learning and avoiding the complex operation of manual feature extraction[8]. They also have good adaptability to image displacement, scaling, distortion, and the ability to combine low-level features into high-level features. As a result, they have been widely applied in crop identification[9,10], agricultural product quality detection[11,12,13], and crop disease identification fields[14,15]. Geng et al.[16] designed a dual-branch deep fusion convolutional neural network (DDFnet) for the classification of dried red dates. Li et al.[17] conducted comprehensive analysis based on the intact degree, fruit size, and color of green plum fruits, and developed a deep learning-based classification method for green plum grades with powerful feature extraction and recognition capabilities. They also established a cognitive error entropy based on the generalized entropy theory to reflect the credibility of the classification results. Momeny et al.[18] achieved recognition and grading of cherry fruit based on whether the fruit shape was regular by improving the traditional CNN model. Sozzi M et al. [19] compared three YOLO series algorithms for automatically detecting and counting white grape clusters, as an alternative to using object detection to estimate crop yields.
The deep feature learning and extraction of convolutional neural networks rely heavily on a large amount of data. In the case of a small sample size, overfitting can be a serious problem. Transfer learning reduces the amount of training data and computational power required to build deep learning models through knowledge (weight) sharing techniques, relaxing the conditions of the sample size and identical probability distribution, and can effectively solve the problem of overfitting on complex network structures with small samples. Behera SK [20] compared classification methods using machine learning and transfer learning, and ultimately achieved 100% accuracy in accurately classifying papaya ripeness using the VGG19 model based on transfer learning.Xue Yong et al. [21] used the GoogLeNet deep transfer model to detect apple defects. Initializing the network with transfer features can improve the network's generalization performance. However, the features extracted by each layer of the network model are different and have varying effects on classification performance. Yosinski et al. [22] discussed the feature extraction ability of different convolutional layers in the network structure under different data set sizes by freezing the parameters of each convolutional layer separately. The study showed that the classification performance of the transferred network is not necessarily increased with the increasing number of layers in the network.

2. Materials and Methods

2.1. Image acquisition of red grape ear

The red seedless grape clusters were harvested from the 22nd company of the 121st regiment of the 8th Agricultural Division of Xinjiang Production and Construction Corps, located at 44°81′N, 85°59′E and an altitude of 245 meters. Professional personnel selected 65 first-grade clusters, 40 second-grade clusters, and 45 third-grade clusters based on the five appearance indicators of fruit powder integrity, cluster shape, berry diameter, and berry count, in accordance with the grading standard DB65/T2832-2007 for Red Globe grapes. The grape clusters were then photographed from multiple angles (clockwise rotation of 120 degrees) using a Canon EOS 550D camera at a distance of 600 mm, with an image resolution of 2,976 × 3,968 pixels.
Preprints 73890 g001
Figure 1. Example of “Red Globe” images data set of different grades.
Figure 1. Example of “Red Globe” images data set of different grades.
Preprints 73890 g001

2.2. Data augmentation of Red Grape Cluster Images.

Considering the randomness of grape cluster imaging and the necessity of data augmentation, the 450 images of the 150 grape clusters were randomly rotated by less than 15 degrees and Gaussian noise with mean 0 and variance 0.01, as well as salt and pepper noise with a noise density of 0.05, were added to increase the diversity of the training dataset, reduce overfitting during network training, and improve the generalization performance of the network. Finally, the dataset of red grape cluster images was expanded to 1800 images. Subsequently, the grape cluster image dataset was randomly divided into a training set and a test set at a ratio of 7:3, with 1260 images used for learning the weight parameters of the grading model and 540 images used to verify the classification ability of the constructed model. The numbers of expanded red grape cluster images and dataset partitions are detailed in Table 1.

2.3. Construction method of appearance quality grading model for red grapes.

2.3.1. Transfer Learning:

As deep convolutional neural networks, represented by VGG16, ResNet50, and GoogleNet, have been sufficiently trained on the ImageNet dataset and learned a large amount of knowledge required for image classification recognition, this paper uses pre-trained classic deep convolutional neural networks as the basis and adopts the transfer learning method of pre-trained network weight parameters. AlexNet, VGG16, ResNet50, GoogleNet, and InceptionV3 are used as feature extractors for red grape image classification recognition. Each convolutional neural network is composed of multiple convolutional and pooling layers. Among them, the convolutional layers load pre-trained weights and bias parameters to extract features and feature mapping from the input image, while the pooling layers are used to maintain the invariance of image features and obtain feature vectors for classification after the fully connected layer.

2.3.2. SVM Classifier

Support vector machine (SVM) is a model discriminant algorithm based on the idea of minimizing structural risk, which has superior classification performance compared to the end-classification function in traditional CNN, especially in solving small sample datasets, nonlinear classification problems, and high-dimensional pattern recognition, and can make better use of the knowledge learned by the pre-trained model [23]. Considering the small number of red grape clusters classification categories and images in this study, the Gaussian radial basis function (RBF) is selected as the kernel function of SVM to achieve the classification and recognition of red grape clusters grades.

2.3.3. Red Grape Appearance Quality Grading Model Construction Process

As CNN pre-trained networks are good at image feature extraction and SVM is good at small sample dataset classification, this study combines the CNN optimal network as the feature extractor with SVM classifier to construct a red grape grading model based on the CNN-SVM hybrid model. The specific construction process is shown in Figure 2. First, the red grape images in the training set are input into the pre-trained networks of AlexNet, VGG16, GoogleNet, InceptionV3, and ResNet50 for image feature information extraction. In this process, due to the different matching degrees between each network model and the image input size, the red grape cluster image size is automatically adjusted by randomly flipping along the vertical axis and randomly translating no more than 30 pixels in the horizontal and vertical directions, to adapt to the requirements of different network models for image input size. Then, by comparing the classification accuracy of red grapes by the SoftMax classifiers of the 5 CNN networks, the best CNN feature extractor is selected, and the optimal weight parameters of the selected network are determined by adjusting the learning rate and training times and other network hyperparameters. Finally, the red grape cluster image feature information extracted from different convolutional layers of the selected CNN network is input into SVM for red grape cluster grading, and model performance evaluation indicators are calculated to establish the best network depth of the feature extractor in the red grape grading model. The hardware device used in the model construction process is a SAMSUNG desktop computer, with an Intel(R) Xeon(R) CPU E5-2620@2.10 GHz processor, 16GB memory, Windows 7 operating system, and Matlab 2018b software for image processing and network training.

2.4. Model Optimization

In order to further improve the accuracy and real-time performance of image recognition of red grape ears, the established classification model needs to be optimized.This paper optimizes network performance by learning rate, training times, loss function and optimizer.The optimizer adopts a stochastic gradient descent optimizer with a first-order momentum factor;Since the sample size of the data set in this study is relatively small, the loss function Loss shown in formula (1) is used to impose constraints and penalties on the model parameters through the L2 norm regularization term to prevent the network model from over fitting and improve the generalization ability of the model.
L o s s = − 1 N ∑ j = 1 N ∑ i = 1 N y j i l o g ⁡ ( y ^ j i ) + 2 W 2 2
where, N represents the number of samples updated in one parameter update during model training, K represents the number of classification categories, y j i uses one-hot labels to represent the classification category labels for the ith image in the jth batch, y ^ j i   represents the predicted classification category labels for the ith image in the jth batch,  represents the regularization parameter, and W represents the weights.

2.5. Model Evaluation Metrics

Accuracy, recall, and F1-score were used to evaluate the classification performance of various network models for red grape bunches.
A c c u r a c y = T P + T N T P + T N + F P + F N × 100 %
R e c a l l = T P T P + F N × 100 %
F 1 = 2 T P 2 T P + F N + F P × 100 %
where, TP is the number of positive samples correctly determined as positive samples, and TN is the number of negative samples correctly determined as negative samples; FP is the number of negative samples wrongly determined as positive samples, FN is the number of positive samples wrongly determined as negative samples.

3. Results and Analysis

3.1. Performance Analysis of Pretrained Network Models

Table 1 shows the classification results of different pretrained network models on red grape cluster images with the same hyperparameter settings, including learning rate (0.01), iterations (100), momentum factor (0.9), batch size (10), and dropout rate (0.5). As shown in Table 1, the detection speed of the AlexNet network model is the fastest, but its training and testing accuracy are relatively low, which may be due to the relatively shallow depth of the AlexNet network, leading to poor feature extraction ability of its convolutional layers for grape cluster images. With the increase of network layers, the classification performance of each model on red grape clusters is improved to varying degrees. Among them, the InceptionV3 network model has the best classification performance on the training set, while the ResNet50 network model has the highest classification performance on the testing set. Although the InceptionV3 and ResNet50 network models have their own strengths in the classification performance on the training and testing sets, respectively, it can be seen from the average detection time that the detection efficiency of the ResNet50 network model is much higher than that of the InceptionV3 network model, which can better meet the real-time requirements of online sorting in the future. Therefore, considering the classification performance of various network models on red grape clusters, this study uses ResNet50 as a feature extractor for deep exploration of red grape cluster images.
Table 2. Comparison of migration learning performance of different network models.
Table 2. Comparison of migration learning performance of different network models.
Feature
extractor		 Network depth Training set Test set Mean detection time/ms
Accuracy/% Recall/% F1-score/% Accuracy/% Recall/% F1-score/%
AlexNet 8 75.62 75.86 74.97 72.31 70.09 72.35 97.00
VGG19 16 81.35 82.24 80.81 78.15 75.46 77.67 128.64
GoogleNet 22 82.14 83.35 82.42 78.79 76.02 78.16 236.14
InceptionV3 48 88.43 89.94 88.11 80.31 78.86 80.99 480.20
ResNet50 50 86.74 87.43 85.98 82.85 80.31 82.69 241.20

3.2. Analysis of the factors affecting the performance of the ResNet50 model

During the training process, we used an exponential ruler to set the learning rate and conducted comparative experiments with different learning rates and iteration parameters to determine the optimal values of these parameters and further improve the transfer learning performance, using the average accuracy on the test set as the standard. According to the results in Table 3, under the same number of training iterations, when the learning rate is 0.001, the ResNet50 model performs better on the test set than other values in terms of average accuracy. As the number of training iterations increases, the average accuracy also increases to varying degrees. When the learning rate is 0.001, the average accuracy increases by 11.04%, 0.33%, and 0.13% as the number of training iterations increases. However, when the number of training iterations reaches 30, the average accuracy decreases by 0.99%, indicating that the ResNet50 model has overfit. After the number of training iterations exceeds 10, the improvement in the network's performance on the test set is limited. Considering the training time of the model, we ultimately set the learning rate to 0.001 and the number of training iterations to 10.

3.3. Performance Analysis of Feature Extraction Based on ResNet50 Network Model

The ResNet50 network model is divided into four stages, as shown in Figure 3. Each stage includes a residual mapping module and a varying number of identity mapping modules, enabling the network to become deeper while maintaining precision and controlling speed.

3.3.1. Performance Analysis of ResNet50 Network Feature Extraction Based on Feature Visualization

Figure 4 shows the feature visualization of the four stages of the ResNet50 network. As can be seen from the figure, the shallow convolution layers near the network have smaller receptive fields for learning low-level features, such as color features, while the deeper convolution layers near the end of the network have larger receptive fields for learning higher-level combinations of low-level features to extract more advanced features.
We selected the feature extracted by the 39th convolution kernel in the first convolution layer of ResNet50 and used it to activate the input image of a red grape. The resulting heatmap is shown in Figure 5(b). The pixels closer to red indicate a stronger positive activation area. It can be seen from Figure 5(b) that this convolution kernel is a color filter that mainly extracts the feature of lost fruit powder on grape grains. For areas with strong light, it can better extract the region of lost fruit powder, while for areas with weak light, the feature extraction effect of lost fruit powder is poor.

3.3.2. Analysis of Network Architecture and Classification Performance of Red Grapes

Based on the characteristics of the ResNet50 network model structure, we selected the feature outputs of the convolution layers in the first convolution layer, each residual mapping module, and each identity mapping module (a total of 17 nodes), and trained an SVM classifier. The classification accuracy on the training set and test set is shown in Figure 4.
Using SVM to train the feature parameters of the 17 nodes extracted by the ResNet50 network model, the average accuracy of each node on the SVM is shown in Figure 6. With the increasing depth of the network, the average accuracy of the features extracted by each node on the SVM continued to increase. The highest average accuracy on the test set was achieved at node 10 (the first identity mapping module in Stage3), reaching 95.08%. At this point, the average accuracy on the training set was 96.88. When the depth of the network continued to increase, the growth trend of the average accuracy on the training set slowed down, while the average accuracy on the test set showed a fluctuating downward trend. This is because the small size of the dataset and the complex deep network structure increased the likelihood of overfitting. The features extracted by nodes 1-10 are universal and can extract common feature information, while the features extracted by nodes 11-17 are specific to different datasets, so even if the network depth increases, the model's capability does not continue to increase. Therefore, we selected the output at node 10 of the ResNet50 network model as the feature for different levels of red grapes.
Overall, this section analyzed the performance of the ResNet50 network model in terms of feature extraction and provided insights into the selection of appropriate features for different levels of red grapes.

3.3.3. Improved Model Performance Analysis.

Based on the previous research, the Resnet50 network model was combined with SVM to construct a grape quality grading model for red table grapes based on their appearance. The strongest feature extraction node in the output of the network model was selected as the feature extraction network, which was combined with SVM to build the model. Using this network model, the appearance of 50 red table grape clusters were classified. Among them, 27 clusters of Grade I and Grade III were correctly classified according to the judgement of professional personnel, while 3 clusters of Grade II were misclassified as Grade III, resulting in an accuracy rate of 94%. Further analysis of the misclassified grape clusters revealed that in the three images taken of each cluster, there was a difference in shape or grade requirements on one side of the grape cluster. In the future, further research will be conducted on the comprehensive evaluation of grape cluster grades based on multiple images.

4. Conclusions

This article investigated a grape quality grading system network model based on transfer learning under small sample conditions. The main conclusions are as follows:
  • By using a model-based transfer learning method, compared with five pre-trained network models, namely, Alexnet, VGG16, Googlenet, ResNet50, and InceptionV3, ResNet50 network model is more suitable for the Red Globe Grape dataset with the same hyperparameter settings. On the test set, the average classification accuracy of the ResNet50 network model can reach 82.85%, and the F1 value is 82.69%.
  • By optimizing the hyperparameters, when the learning rate is set to 0.001 and the number of iterations is 10, the grading accuracy of the ResNet50 network model on the red table grape dataset can reach 93.89%.
  • By analyzing the feature output of the intermediate convolutional layers in the ResNet50 network model layer by layer, and combining SVM technology, a grape quality grading model for red table grape clusters based on their appearance was constructed and tested on the dataset. The average classification accuracy of the constructed network model can reach 95.08%. The experimental results indicate that the deep network transfer learning algorithm proposed in this article has certain application value for grape cluster quality grading.In summary, this study proposes a deep network transfer learning algorithm for grape cluster quality grading, which has achieved promising results on the red table grape dataset.

Author Contributions

Conceptualization, J.W.; Methodology, J.W., Z.Z. and D.S.; Software, Z.Z.; Validation, Z.Z.; Formal analysis, Z.Z.; Investigation, D.S.,H.S. and X.C; Resources, J.W. and Z.Z.; Data curation, Z.Z.,X.C. and D.S.; Writing—original draft preparation, Z.Z. and D.S.; Writing—review and editing, Z.Z.; Manuscript revising, J.W. and H.S.; Study design, Z.Z., X.C. and J.W.; Supervision, J.W., H.S.and D.S.; Project administration, Z.Z and J.W.; Funding acquisition, J.W. All authors have read and agreed to the published version of the manuscript.

Funding

This work was funded by the National Natural Science Foundation of China (grant number 31860466).

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. MU Weisong, FENG Jianying, TIAN Dong, et al. The international trade and domestic demand of the table grape industry in China[J]. Chinese Fruit Tree, 2019(02): 5-10.
  2. Anuja Bhargava, Atul Bansal. Fruits and vegetables quality evaluation using computer vision: A review[J]. Journal of King Saud University-Computer and Information Sciences, 2021, 33: 243-257. [CrossRef]
  3. TIAN Youwen, WU Wei, LU Shiqian, et al. Application of Deep Learning in Fruit Quality Detection and Classification[J]. Food Science, 2020: 1-13. [CrossRef]
  4. CHEN Ying, LIAO Tao, LIN Chukao, et al. Grape Inspection and Grading System Based on Computer Vision[J]. Transactions of the Chinese Society for Agricultural Machinery, 2010,41(03): 169-172. [CrossRef]
  5. Li Junwei, Guo Junxian, Hu Guanghui, etcResearch on prediction and grading of single grape quality and fruit diameter based on machine vision technology [J]Xinjiang Agricultural Sciences, 2014, 51 (10): 1862-1868. [CrossRef]
  6. XIAO Zhuang, WANG Qiaohua, WANG Bin, et al. A Method for Detecting and Grading ‘Red Globe’ Grape Bunches Based on Digital Images and Random Least Squares[J]. Food Science, 2018, 39(15): 60-66. [CrossRef]
  7. MISHRA R K, REDDY G Y S, PATHAK H. The Understanding of Deep Learning: A Comprehensive Review[J]. Mathematical Problems in Engineering, 2021: 1-15. [CrossRef]
  8. KAMILARIS A, PRENAFETA-BOLDÚ F X. Deep learning in agriculture: A survey[J]. Computers and Electronics in Agriculture, 2018, 147: 70-90. [CrossRef]
  9. MAO S, LI Y, MA Y, et al. Automatic cucumber recognition algorithm for harvesting robots in the natural environment using deep learning and multi-feature fusion[J]. Computers and Electronics in Agriculture, 2020,170: 105254. [CrossRef]
  10. MAJEED Y, ZHANG J, ZHANG X, et al. Deep learning based segmentation for automated training of apple trees on trellis wires[J]. Computers and Electronics in Agriculture, 2020,170: 105277. [CrossRef]
  11. Zhang Siyu, Zhang Qiuju, Li Ke. Detection of peanut kernel quality based on machine vision and adaptive convolution neural network[J]. Transactions of the Chinese Society of Agricultural Engineering, 2020, 36(04): 269-277. [CrossRef]
  12. ZHENG Kai, FANG Chun, SUN Fuzhen. Quality classification of green pepper based on deep learning[J]. Journal of Shandong University of Technology(Natural Science Edition), 2020,34(04): 18-23.
  13. Long Mansheng, Ouyang Chunjuan, Liu Huan, et al. Image recognition of Camellia oleifera diseases based on convolutional neural network & transfer learning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(18): 194-201. (in Chinese with English abstract). [CrossRef]
  14. WANG G, SUN Y, WANG J. Automatic Image-Based Plant Disease Severity Estimation Using Deep Learning[J]. Computational Intelligence and Neuroscience, 2017,2017: 1-8. [CrossRef]
  15. Liu Yang, Gao Guoqin. Identification of multiple leaf diseases using improved SqueezeNet model[J]. Transactions of the Chinese Society of Agricultural Engineering, 2021,37(02): 187-195.
  16. GENG L, XU W, ZHANG F, et al. Dried Jujube Classification Based on a Double Branch Deep Fusion Convolution Neural Network[J]. Food Science and Technology Research, 2018,24(6): 1007-1015. [CrossRef]
  17. Li Weitao1.2, Cao Zhongda1, Zhu Chenghui. Intelligent feedback cognition of greengage gradeb based on deep ensemble learning[J]. Transactions of the Chinese Society of Agricultural Engineering, 2017,33(23): 276-283.
  18. MOMENY M, JAHANBAKHSHI A, JAFARNEZHAD K, et al. Accurate classification of cherry fruit using deep CNN based on hybrid pooling approach[J]. Postharvest Biology and Technology, 2020, 166: 111204. [CrossRef]
  19. Sozzi, M. and S. Cantalamessa, et al. (2022). "Automatic bunch detection in white grape varieties using YOLOv3, YOLOv4, and YOLOv5 deep learning algorithms." Agronomy 12(2): 319. [CrossRef]
  20. Behera S K, Rath A K, Sethy P K. Maturity status classification of papaya fruits based on machine learning and transfer learning approach[J]. Information Processing in Agriculture, 2021,8(2):244-250. [CrossRef]
  21. XUE Yong, WANG Liyang, ZHANG Yu, et al. Defect Detection Method of Apples Based on GoogLeNet Deep Transfer Learning[J]. Transactions of the Chinese Society for Agricultural Machinery, 2020,51(07): 30-35.
  22. YOSINSKI J, CLUNE J, BENGIO Y, et al. How transferable are features in deep neural networks[J]. International Conference on Neural Information Processing Systems, 2014. https://dl.acm.org/doi/10.5555/2969033.2969197.
  23. NTURAMBIRWE J F I, OPARA U L. Machine learning applications to non-destructive defect detection in horticultural products [J].Biosystems Engineering, 2020, 189: 60-83. [CrossRef]
Preprints 73890 g002
Figure 2. Flow chart of study.
Figure 2. Flow chart of study.
Preprints 73890 g002
Preprints 73890 g003
Figure 3. Structure of ResNet50 network.
Figure 3. Structure of ResNet50 network.
Preprints 73890 g003
Preprints 73890 g004
Figure 4. Example of feature extraction.
Figure 4. Example of feature extraction.
Preprints 73890 g004
Preprints 73890 g005
Figure 5. Example of feature extraction: (a) is the original image and (b)is the example of feature extraction.
Figure 5. Example of feature extraction: (a) is the original image and (b)is the example of feature extraction.
Preprints 73890 g005
Preprints 73890 g006
Figure 6. Classification accuracy of each node on test set.
Figure 6. Classification accuracy of each node on test set.
Preprints 73890 g006
Table 1. Number of “Red Globe” sample images in each category.
Table 1. Number of “Red Globe” sample images in each category.
Cluster grade Original image Data augmentation Dataset division
Training set Test set
Grade I 195 585 546 234
Grade II 120 360 336 144
Grade III 135 405 378 162
Table 3. Results of models training with different Epochs and learning rates (%).
Table 3. Results of models training with different Epochs and learning rates (%).
Learning rate Epochs
5 10 15 20 30
0.1 55.77% 65.46% 67.31% 69.23% 88.46%
0.01 82.85% 93.89% 94.22% 94.35% 93.36%
0.001 85.12% 92.31% 93.99% 92.31% 92.46%
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.
Prerpints.org logo

Preprints.org is a free preprint server supported by MDPI in Basel, Switzerland.

facebook logotwitter logolinkedin logo
MDPI Initiatives
Important Links
Subscribe

Disclaimer

Terms of Use

Privacy Policy

© 2024 MDPI (Basel, Switzerland) unless otherwise stated