Plant leaf diseases constitute substantial issues in agriculture. The tomato, being one of the world's most vital crops, incur enormous economic losses for farmers and impair the healthy development of the tomato sector when impacted by illnesses. Therefore, accurate detection and classification of tomato leaf diseases have become crucial. In recent years, more and more academics have begun to use deep learning approaches in the field of plant disease identification and have produced good results. However, these approaches still have limited accuracy in situations with complicated backdrops or multiple interferences and are not lightweight enough, with significant requirements on computational resources. The present study proposes an improved network model for the purpose of recognizing tomato leaf diseases. Based on the original ResNet50 model, a novel network model, SDE-ResNet50, was constructed by adding an Efficient Channel Attention（ECA）module, depth-wise separable convolutions and modifications to the stem structure. The experimental findings indicate that the SDE-ResNet50 attained a classification accuracy of 98% in the identification of tomato leaf diseases, hence outperforming ResNet50 by a margin of 5.6%. Furthermore, the SDE-ResNet50 demonstrates a decrease in size of about 47.73% and a reduction in computational complexity of roughly 70.28% when compared to the ResNet50 model. The SDE-ResNet50 was evaluated against current widely-used classification networks using an identical dataset of tomato leaf diseases, and it exhibited greater performance. The present discovery provides evidence of the efficacy and viability of the proposed enhancement to the model as outlined in this research article.
Research Article
Tomato Leaf Disease Classification Based on Feature Enhancement and SDE-ResNet50
https://doi.org/10.21203/rs.3.rs-4373732/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Attention Mechanism
Deep learning
Image Classification
Resnet50
Tomato Leaf Disease Detection
The importance of food security has increased due to ongoing global population growth. According to the Food and Agriculture Organization of the United Nations (FAO), the global prevalence of hunger persisted among an estimated 691 to 783 million individuals as of 2022, with a discernible upward trend [1]. Plant diseases are severe natural disasters characterized by multiple categories, strong destructive power, difficulty in identification, and frequent outbreaks. They are significant factors leading to the decline in food quality and yield. Research has shown that nearly 40% of global crop yields are significantly reduced due to diseases and pests [2]. Hence, the timely identification of illnesses and pests by agricultural personnel can significantly mitigate crop output losses and minimize reliance on pesticide substances. The tomato is one of the most extensively consumed veggies globally. The substance in question is abundant in vitamin A, vitamin C, calcium, potassium, iron, and various other components. Additionally, it is rich in proteins, carbs, and dietary fiber [3]. The tomato category provides approximately 19% of the overall vegetable consumption in the United States, ranking second in prevalence after potatoes, which account for 23% of the total. These substances are extensively employed in the manufacturing of several culinary components, condiments, and drinks. The annual value of commercial tomato production in the United States exceeds 2 billion dollars [4]. Consequently, identifying and classifying of diseases tomato leaf diseases has become a crucial research area in recent years. The conventional approach to the identification of plant diseases entails the collection of afflicted leaves, which are subsequently dispatched to specialists in plant pathology to ensure precise diagnosis. Nevertheless, this methodology is characterized by its requirement of a significant amount of time and manual effort. The accuracy and effectiveness of the process are highly dependent on the competence, diagnostic acumen, and physical well-being of the professionals involved.
In the early stages, scholars initiated an exploration into the prospective utilization of machine learning and computer vision technologies within the fields of plant disease and pest identification. The purpose of implementing this approach was to replace time-consuming manual identification operations, with the ultimate aim of establishing intelligent agriculture management. Zhu Leqing et al. [5] proposed a method for automatic classification by analyzing insect wing images using histograms and gray-level co-occurrence matrices. The method was tested on a database containing 100 insect species, achieving an accuracy of 71.1%. Barbedo et al. [6] presented a methodology for effectively distinguishing between asymptomatic tissues and disease-infected regions in plant leaves. This was achieved by utilizing the H-channel of the HSV color space and the a-channel of the L*a*b color space. The methodology uses the histogram of the color channels stated above for the purpose of classification. Mokhtar, Usama, and colleagues [7] proposed a technique that employs the Gabor wavelet transform and support vector machine (SVM) [8] for the purpose of identifying diseases in tomato plants. The methodology entails the utilization of the Gabor wavelet feature technique to extract significant features from tomato leaf images, followed by the application of SVM for the goal of classification. Revathi et al. [9] proposed an HPCCDD algorithm based on edge detection and image segmentation for image analysis to classify and identify cotton leaf spot disease lesions. Naikwadi et al. [10] proposed an automated method for identifying and classifying plant leaf diseases. The approach consists of two major steps. To begin, it finds the green pixels in the image and masks them with a specified threshold computed using Otsu's approach [11]. Pixels with zero values in the red, green, and blue channels, as well as pixels on the borders of infected zones, are totally deleted in the second stage. The detection accuracy of the proposed approach ranged from 83% to 94%. The aforementioned approaches for identifying plant diseases and pests typically commence with preprocessing operations such as edge detection, segmentation, and pixel manipulation on the input images. The subsequent stage involves the manual design of features. The method of feature engineering described above necessitates a substantial investment of both time and resources. Furthermore, these methods exhibit a significant drawback in terms of specificity, so their performance is typically less than optimal when faced with different kinds of plants or different kinds of pests.
In recent years, there has been a gradual rise in the prevalence of deep learning, a technique that avoids the necessity for laborious manual feature extraction. Deep learning exploits the capabilities of neural networks and massive training to autonomously acquire feature representations. The aforementioned capacity enables the utilization of deep learning in classification tasks across many domains, showcasing strong generalization capabilities. Moreover, deep learning effectively tackles the constraints encountered by conventional machine learning approaches when dealing with extensive and complex datasets, which frequently exhibit inadequate performance. Ramcharan, Amanda, et al. [12] trained a deep convolutional neural network (DCNN) to diagnose three illnesses and two pest diseases in cassava by using transfer learning. During training, they attained an overall accuracy of 93% on previously encountered datasets. Ma et al. [13] proposed a DCNN for the identification of four diseases in cucumber: anthracnose, downy mildew, powdery mildew, and target leaf spot. Their approach achieved a recognition accuracy of 93.4%. Zhang et al. [14] conducted trials on the recognition of nine maize leaf diseases using two modified models: GoogLeNet and Cifar 10. They altered the model parameters, changed the pooling combinations, and applied dropout regularization. The two improved models were trained and tested on the corn leaf disease images, obtaining average accuracy of 98.9% and 98.8%, respectively. Kawasaki et al. [15] presented a plant disease detection system based on CNN. They utilized 800 cucumber leaf images and trained the CNN system using a 4-fold cross-validation strategy. The system achieved an average accuracy of 94.9% in their experiments. Turkoglu et al. [16] proposed a multi-model long short-term memory (LSTM) strategy for plant disease and pest identification, with a pre-trained CNN serving as an ensemble majority voting classifier. The experimental results showed that the proposed combination structure outperformed pre-trained deep architectures in terms of performance. Esgario et al. [17] constructed a CNN-based program for segmenting, classifying, and evaluating the degree of various forms of leaf damage caused by biological causes in coffee leaves. The program accurately approximated the severity values, which were extremely close to the actual values, according to the experimental data, with a classification accuracy surpassing 97%. Karthik R. et al. [18] presented two distinct deep architectures with the aim of identifying disease kinds in tomato leaves. The initial architectural design employed residual learning as a means to effectively capture significant classification elements. The subsequent architectural design integrated an attention mechanism into the pre-existing residual network. The accuracy of the second architecture in classifying three forms of tomato leaf diseases reached 98%.
In conclusion, deep learning techniques have made significant strides in addressing limitations observed in conventional machine vision methods, such as low recognition efficiency and accuracy. However, these approaches continue to encounter obstacles including intricate model structures, computationally demanding processes, insufficient exploitation of spatial and channel information within images, and reduced accuracy when confronted with complex backgrounds. This research presents a novel network model, namely SDE-ResNet50, as an approach to address the aforementioned issues. The primary objective of this model is to achieve both high efficiency and accuracy in the classification of tomato leaf diseases.
The main contributions of this paper are as follows:
- By altering the stem structure of the ResNet50 model, including the ECA module, and replacing the standard convolution with a depth-wise separable convolution, a novel network model, SDE-ResNet50, was built in this research to recognize tomato leaf diseases under complicated backdrops and disturbances.
- The paper applies various data augmentation techniques to the dataset, including level flipping, random occlusion, Gaussian noise, motion blur, and brightness changes. The aim of these operations is to reinstate the intricate backgrounds and disruptions that would often arise throughout a realistic acquisition process.
- The SDE-ResNet50 model proposed in this study demonstrates a notable accuracy of 98% on the augmented dataset, exhibiting a substantial enhancement of 5.6% compared to the original model. In comparison to other state-of-the-art network models, SDE-ResNet50 demonstrates the most significant overall benefit in terms of both performance and its lightweight
2.1 Dataset
The dataset pertaining to tomato leaf disease utilized in this study was obtained from the Kaggle platform. The dataset comprises a total of ten distinct categories, encompassing nine distinct classifications of tomato leaf diseases, namely bacterial spot, early blight, late blight, and target spot, and a class denoting healthy leaves. The 10 types of tomato are shown in Figure 1.
The dataset consists of 11,000 photos in JPG format, each having dimensions of 256x256 pixels. The photos were partitioned into a training set and a test set using a random allocation method, with a ratio of 10:1. Table 1 displays the distribution of the dataset. The training set has 10 classes, with each class containing 1,000 images, resulting in a cumulative total of 10,000 images. Similarly, every class within the test set is composed of 100 photographs, resulting in a cumulative total of 1,000 images.
Table 1 The distribution of ten different types of tomato leaf diseases in the training and testing datasets.
|
Disease |
Number of images for training |
Number of images for testing |
|
Bacterial spot |
1000 |
100 |
|
Early blight |
1000 |
100 |
|
Late blight |
1000 |
100 |
|
Leaf mold |
1000 |
100 |
|
Septoria leaf spot |
1000 |
100 |
|
Two spotted spider mite |
1000 |
100 |
|
Target spot |
1000 |
100 |
|
Yellow leaf curl virus |
1000 |
100 |
|
mosaic virus |
1000 |
100 |
|
Health |
1000 |
100 |
|
Total |
10000 |
1000 |
2.2 Data augmentation
Data augmentation is the process of generating new training samples by applying a series of random transformation operations to the original data. The goal of this strategy is to enhance the amount and variety of the dataset, thereby improving the model's generalization and robustness. Collected images in practical applications frequently include complicated backgrounds and numerous sorts of interference. Therefore, maintaining high accuracy in such challenging situations is critical for a model. To overcome this, data augmentation can be used to boost the dataset's diversity and difficulty. This study employed a range of data augmentation techniques, such as the addition of Gaussian noise, manipulation of image brightness, application of motion blur, implementation of horizontal flipping, and introduction of random occlusion. The purpose of these procedures was to replicate the intricacy and variety observed in real-life situations. Enhance the model's capacity to effectively acclimate to intricate surroundings and diverse disturbances. Figure 2 and Figure 3 illustrate the visual changes in the dataset before and after applying data augmentation techniques.
2.3 Network model
In 2015, Kaiming He et al. and other researchers from Microsoft Research Asia proposed ResNet [19] as a deep neural network architecture. Residual blocks are incorporated in order to improve the training process of the network and allow for more depth in its architecture. The inclusion of shortcut connections that span many convolutional layers within the residual blocks enables the neural network to efficiently learn the residual between the current layer's output and input. This mechanism effectively mitigates the challenges of vanishing and exploding gradients that are commonly seen in excessively deep networks. Hence, the utilization of ResNet in the domain of plant leaf detection exhibits the ability to significantly enhance the precision and efficiency of leaf disease image identification. For the purposes of this study, the ResNet-50 model was chosen as the foundational network due to its suitable depth and complexity. The network architecture of ResNet-50 is seen in Figure 4. The architecture comprises a network stem module, followed by four consecutive bottleneck groups, and culminates in a completely connected layer.
2.4 Replacement of the stem structure
The stem module denotes the beginning of a network model. A max pooling layer follows a 7x7 convolutional layer in the ResNet-50 stem module. However, when it comes to photographs of tomato leaf disease, it is seen that the affected areas tend to comprise a very minor fraction of the overall image. The utilization of a 7x7 convolutional kernel, characterized by its expansive receptive field, has the potential to result in the omission of important information throughout the process of feature extraction. In this study, we replace the original stem structure with two stacked 3x3 convolutional kernels and an extra reduction module [20]. The stem structure is replaced by smaller convolutional kernels with higher local feature extraction and detail capture capabilities. Stacking numerous tiny convolutional kernels creates more non-linear responses, increasing the model's expressive ability. The extracted features are inputted into two parallel branches within the reduction module. One of the branches is comprised of convolutional layers, which employ convolutional operations to extract local features from the input feature maps. The second branch proceeds by passing pooling layers, wherein higher-level global features are extracted via downsampling operations. Ultimately, the aforementioned branches are combined through the action of concatenation, utilizing the concat function. This results in a more enriched and varied representation of the information. Simultaneously, the novel stem architecture compresses the input feature maps to a reduced size while preserving abundant feature information. This enhancement results in a more lightweight model without sacrificing its powerful classification performance. The revised structure is depicted in Figure 5.
2.5 Attention mechanism and depth-wise separable convolution
The appearance of disease spots on tomato leaves is distinguished by irregular and varied shapes. Furthermore, different diseases might present visually similar symptoms on the leaf surface. These difficulties make model recognition challenging. Attention mechanisms ease this by assigning different weights to input features, allowing the model to selectively pay to important information while ignoring unimportant details. Notably, this method has a low parameter count, which improves the model's performance and efficiency.
During the 2019 Conference on Computer Vision and Pattern Recognition (CVPR), Wang et al. proposed the ECA module [21]. The approach employs one-dimensional convolution to enable the exchange of information between channels. The size of the convolutional kernel (k) is controlled by an adaptive function with the following formula:
The ECA module utilizes global average pooling (GAP) on the input feature map, resulting in the production of a one-dimensional vector. The vector is subsequently submitted to processing by a one-dimensional convolutional layer, where the size of the convolutional kernel is chosen using an adaptive function. The weights obtained are multiplied element-wise with the input feature map, resulting in a new feature map that incorporates weighting specific to each channel.
The depth-wise separable convolution [22] contains two distinct components: depth-wise convolution and point-wise convolution. In the depth-wise convolution, individual channels undergo convolution with distinct kernels, and afterwards, the resulting outputs are concatenated. This type of point-wise convolution involves the application of a 1x1 convolutional kernel to each channel of the input data. The former method demonstrates superior capability in capturing the spatial properties of the feature map, whereas the latter method is more proficient in capturing features specific to individual channels. When comparing standard convolution to depth-wise separable convolution, it becomes evident that the latter demonstrates enhanced abilities in extracting intricate features and capturing correlations within the feature space.
In this study, the final 1x1 regular convolution within the bottleneck architecture was replaced with depth-wise separable convolution, after which an ECA module was added. The enhanced bottleneck configuration is shown in Figure 6.
3.1 Experimental Setup
The experiments were conducted in a Linux operating system using the PyTorch deep learning framework with Python 3.8 as the programming language. The GPU used was NVIDIA Tesla V100. The CPU used was Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz, The data analysis tool used in this experiment was matplotlib.
3.2 Evaluation Metrics
This study introduces three evaluation metrics, namely accuracy, precision, and F1 score, which are employed to evaluate the classification performance of our model. Accuracy is a metric that describes the ratio of accurately identified samples to the overall number of samples. Precision is a measure that describes the ratio of correctly detected positive samples by the classifier relative to all samples labeled as positive by the classifier. The F1 score is a metric that quantifies the balance between precision and recall by calculating their harmonic mean. A higher F1 score signifies superior classification performance for the model. The formulas utilized in the computation of these metrics are as follows:
3.3 Experimental content
The research involved doing experimental comparisons between the SDE-ResNet50 model and the original ResNet50 model. During the training process, both models employed the identical cross-entropy loss function and Adam optimizer while ensuring the maintenance of consistent hyperparameters. The training procedure encompassed 100 epochs, with a batch size of 32 and an initial learning rate of 0.01. Moreover, we conducted a comparative analysis of the classification performance of the SDE-ResNet50 model with three other widely-used CNN models, namely VGG16 [23], Inception-ResNet, and MobileNetV3 [24], using an identical dataset of tomato leaf diseases.
3.4 Analysis of experimental results
Figure 7 shows the test set loss and accuracy outcomes for both the SDE-ResNet50 model and the ResNet50 model, subsequent to their training on the training and test sets. The examination of the loss curve indicates that the SDE-ResNet50 curve achieves stability and converges to a significantly lower minimum in comparison to the baseline model after 10 iterations. The accuracy curve demonstrates the consistent superiority of the SDE-ResNet50 model over the original model. The peak accuracy of the SDE-ResNet50 model was 98%, exhibiting a substantial improvement of 5.6% when compared to the old model's accuracy of 92.4%. Table 2 presents the alterations observed in the parameters of the improved front-end model. The improvement in the model resulted in a reduction of computational complexity by 70.28 percent and a decrease in model size by 47.73 percent, while exhibiting only a marginal rise in the parameter.
Table 2 The parameters of the ResNet50 model before and after improvement
|
Model |
parameters |
Floating-point operations |
Model size |
|
Resnet50 |
23,528,522 |
3.89 Gmac |
364.82 MB |
|
SDE-Resnet50 |
23,557,434 |
1.16 Gmac |
190.73 MB |
[1]
Figure 8 shows the dynamic alterations in the evaluation metrics and test set loss across the training iterations for the SDE-ResNet50 model, along with three additional widely utilized CNN. The SDE-ResNet50 and Inception-ResNet models exhibit superior performance in terms of rapid convergence, stability, and optimization of loss throughout the training period. They also outperform the other two models in terms of accuracy, precision, and F1 scores. The SDE-ResNet50 model obtains peak accuracy, precision, and F1 score scores of 98.0%, 98.02%, and 97.99%, respectively, whereas the Inception-ResNet model achieves 98.2%, 98.23%, and 98.2%, respectively. The SDE-Resnet50 model, as shown in Table 3, offers excellent classification capabilities but a lightweight network structure. Its parameters, calculations, and model sizes are substantially smaller than those of the Inception-Resnet model.
Table 3 The parameters of the SDE-ResNet50 model and other deep models for comparison
|
Model |
parameters |
Floating-point operations |
Model size |
|
SDE-Resnet50 |
23,557,434 |
1.16 Gmac |
190.73 MB |
|
MobileNet_V3 |
4,214,842 |
0.23 Mmac |
122.06 MB |
|
VGG16 |
134,309,962 |
15.53 Gmac |
835.06 MB |
|
Inception-Resnet |
54,321,834 |
13.22 Gmac |
897.77 MB |
[1] “Gmac” is a unit that expresses the amount of floating-point operations, which means one billion floating-point operations; “MB” is a unit of computer storage capacity.
The paper proposes a method based on the ResNet50 model. The SDE-ResNet50 model is developed by including ECA modules and depth-wise separable convolution while also making modifications to the stem structure of the network. In comparison to the original model and three alternative convolutional neural network models, the proposed approach exhibits superior comprehensive performance in terms of classification ability and lightweight when applied to the tomato leaf disease dataset that has been augmented with data augmentation. The model demonstrated a test set accuracy rate of 98%, exhibiting a notable enhancement of 5.6% in comparison to the original model. Concurrently, it exhibited a noteworthy decrease in model size and computational complexity. In contrast to the Inception-ResNet model, which attained a comparable accuracy rate of 98.2%, the SDE-ResNet50 model exhibits significantly reduced parameter count, model size, and computational complexity, amounting to 43%, 21%, and 9%, respectively. The results of this research demonstrate that the enhancements implemented in the model have proven to be extremely effective, effectively mitigating the challenges associated with inflated CNN models and compromised accuracy in complex backdrops.
Authors in this paper declare that there are no conflicts of interest in any product, service or company regarding the publication. There are also no financial and personal relationships with other researchers or institutions that can inappropriately influence our work
Author Contribution
Jinyi Wu conceptualized the research idea, led the overall execution and coordination of the study, and was responsible for communication and manuscript revisions with the journal.Jiaxin Dong spearheaded the experimental work and research design, and took on the primary role in drafting the manuscript. Shengwei Chen was in charge of the literature review, data collection, and preliminary analysis.Jie Wang assisted in experimental design and performed the statistical analysis of the results. Xin He provided crucial technical support and created the figures. All authors participated in the discussion, review, and editing of the manuscript.
- FAO, IFAD, UNICEF, WFP and WHO.: In Brief to The State of Food Security and Nutrition in the World 2023. FAO, IFAD, UNICEF, WFP and WHO. https://www.fao.org/3/cc6550zh/cc6550zh.pdf (2023). Accessed 26 June 2023
- R. Gandhi, S. Nimbalkar, N. Yelamanchili, and S. Ponkshe.: Plant disease detection using CNNs and GANs as an augmentative approach. In 2018 IEEE International Conference on Innovative Research and Development (ICIRD). pp. 1-5(2018)
- G. R. Beecher.: Nutrient Content of Tomatoes and Tomato Products. Exp. Biol. Med., 218(2), 98–100 (1998), https://doi.org/10.3181/00379727-218-44282a
- K. J. Reimers and D. R. Keast, et al.: Tomato Consumption in the United States and Its Relationship to the US Department of Agriculture Food Pattern. Nutrition Today. 51(4), 198–205(2016), http://doi.org/ 10.1097/NT.0000000000000152
- Zhu L Q, Zhang Z.: Auto-classification of insect images based on color histogram and GLCM. In 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery. Yantai, China. pp. 2589–2593(2010). https://doi.org/10.1109/FSKD.2010.5569848
- J. G. A. Barbedo, et al.: A novel algorithm for semi-automatic segmentation of plant leaf disease symptoms using digital image processing. Tropical Plant Pathology. 41(4), 210–224(2016). https:// doi.org/10.1007/s40858-016-0090-8
- U. Mokhtar, M. A. S. Ali, A. E. Hassenian, and H. Hefny.: Tomato leaves diseases detection approach based on Support Vector Machines. In 2015 11th International Computer Engineering Conference (ICENCO). pp. 246-250. https://doi.org/10.1109/ICENCO.2015.7416356
- C. Cortes and V. Vapnik.: Support-vector networks. Machine Learning. 20(3), 273–297(1995). https://doi.org/10.1007/BF00994018
- P. Revathi and M. Hemalatha.: Classification of cotton leaf spot diseases using image processing edge detection techniques. In 2012 International Conference on Emerging Trends in Science, Engineering and Technology. pp. 169-173. https://doi.org/10.1109/INCOSET.2012.6513900
- S. Naikwadi and N. Amoda.: Advances in image processing for detection of plant diseases. International Journal of Application or Innovation in Engineering & Management. 2(11) (2013)
- Otsu, Nobuyuki.: A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics. 9(1), 62-66(1979)
- A. Ramcharan, K. Baranowski, P. McCloskey, B. Ahmed, J. Legg, and D. P. Hughes.: Deep learning for image-based cassava disease detection. Front Plant Sci. 8, 1852(2017). https://doi.org/10.3389/fpls.2017.01852
- J. Ma, K. Du, F. Zheng, L. Zhang, Z. Gong, and Z. Sun.: A recognition method for cucumber diseases using leaf symptom images based on deep convolutional neural network. Computer and Electronics in agriculture. 154, 18-24(2018). http://doi.org/10.1016/j.compag.2018.08.048
- X. Zhang, Y. Qiao, F. Meng, C. Fan, and M. Zhang.: Identification of maize leaf diseases using improved deep convolutional neural networks. IEEE Access. 6, 30370–30377(2018). https://doi.org/10.1109/ACCESS.2018.2844405
- Kawasaki, Yusuke, et al.: Basic study of automated diagnosis of viral plant diseases using convolutional neural networks. Advances in Visual Computing: 11th International Symposium. pp. 638-645(2015). https://doi.org/10.1007/978-3-319-27863-6_59
- Turkoglu M, Hanbay D, Sengur A.: Multi-model LSTM-based convolutional neural networks for detection of apple diseases and pests. Journal of Ambient Intelligence and Humanized Computing. 13(7), 3335-3345(2019). https://doi.org/10.1007/s12652-019-01591-w
- Esgario JGM, de Castro PBC, Tassis LM, Krohling RA.: An app to assist farmers in the identification of diseases and pests of coffee leaves using deep learning. Information Process in Agriculture. 9(1), 38–47(2022). https://doi.org/10.1016/j.inpa.2021.01.004
- Karthik, R., et al.: Attention embedded residual CNN for disease detection in tomato leaves. Applied Soft Computing. 86, 105933(2020). https://doi.org/10.1016/j.asoc.2019.105933
- K. He, X. Zhang, S. Ren, and J. Sun.: Deep Residual Learning for Image Recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778(2016). https://doi.org/10.1109/CVPR.2016.90
- Szegedy, Christian, et al.: Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning. Proceedings of the AAAI conference on artificial intelligence. 31(1), 2017
- Wang Qilong, et al.: ECA-Net: Efficient channel attention for deep convolutional neural networks. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11534-11542(2020). http://arxiv.org/abs/1910.03151
- Howard, Andrew G., et al.: MobileNets: efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv: 1704.04861(2017)
- K. Simonyan and A. Zisserman.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv: 1409.1556(2014). https://doi.org/10.48550/arXiv.1409.1556
- Howard, Andrew, et al.: Searching for mobilenetv3. Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324(2019). https://doi.org/10.1109/ICCV.2019.00140
- Liu J, Wang X.: Plant diseases and pests detection based on deep learning: a review. Plant Methods. 17, 1-18(2021). https://doi.org/10.1186/s13007-021-00722-9
- Li L, Zhang S, Wang B.: Plant Disease Detection and Classification by Deep Learning—A Review. IEEE Access. 9, 56683–56698(2021). https://doi.org/10.1109/ACCESS.2021.3069646
- Nagaraju, Mamillapally, and Priyanka Chawla.: Systematic review of deep learning techniques in plant disease detection. International journal of system assurance engineering and management. 11, 547-560(2020). https://doi.org/10.1007/s13198-020-00972-1
No competing interests reported.