A Comparative Study of Convolutional Neural Network Architectures for Enhanced Tomato Leaf Disease Classification Using Refined Statistical Features

To improve the quality of images and extract useful data for analysis, image pre-processing is a crucial stage in the disease detection process for tomato plant leaves. Pre-processing images in order to spot diseases on tomato plant leaves involves several steps as shown in Figure 1 Standardization and normalizing ensure image comparability. These methods standardize image scale, range, and color space for fair and meaningful comparisons. Standardization reduces illumination, camera, and image capture variances, making diseases classification algorithms more robust. Preprocessing approaches prepare images for feature extraction. Features like color histograms, texture descriptors, and form parameters are extracted from sick patches. Extracting discriminative characteristics helps classify damaged tomato leaves more accurately. Preprocessing pipeline data augmentation can enhance dataset diversity and size. Data augmentation boosts diseases classification model resilience and generalization by adding samples with different illumination, rotations, flips, or noise levels. Image preprocessing improves image quality, reduces noise, segments significant sections, and extracts useful characteristics. The preprocessing pipeline improves tomato plant leaf disease classification by improving data quality and relevance. It allows machine learning or deep learning models discover the patterns and correlations between image data and disease classes, enabling tomato disease control.

1.png

Figure 1. Pre-processing steps for image diseases classification

Data Acquisition: The dataset is made up of a set of different image classes and several images. These numbers were taken from the primary dataset that was collected from PlantVillage, ten classifications and 14529 images of tomato leaf diseases are included in the data collection. They are check leaves, stems, and fruits for discoloration, stains, wilting, and abnormalities. They document and image afflicted plants with notebooks, cameras, or phones. Tomato plant leaves are captured with digital cameras or sophisticated equipment. To see the leaves, several images are taken from different perspectives. Disease signs are highlighted in the images. Experts carefully outline impacted leaf parts. This stage helps deep learning models identify disease trends.

The tests' findings diseases, severity, and other findings are recorded. This data trains and validates disease detection models. The dataset includes field observations, annotated images, and laboratory test findings. This dataset shows tomato plant diseases visually and diagnostically. Data integration connects sources for analysis. Images, annotations, and lab results are organized. Depending on project size and needs, it can be stored in a database, file system, or cloud. Data organization and documentation improves management and retrieval. Rotating, flipping, resizing, and adding noise to images creates more diversified and larger datasets. Data augmentation exposes deep learning models to more disease variants, improving their reliability and generalizability. Pre-processed data is used to identify features, train models, and predict diseases. Using gathered data, deep learning algorithms automatically detect and classify tomato plant diseases based on visual signs. Disease detection models depend on data quality and representation.

Pre-Processing: This is done in order to change the data into a form that can be used by the feature extraction approach as well as the stages that come after it. During this step, augmentation and normalization of the data are carried out.

Feature Extraction: First, the important features are retrieved so that we may answer the categorization problem that we have. Color, form, and texture are all characteristics of an image. The texture feature is given a greater amount of attention by diagnostic algorithms that use photos of leaves to identify disorders.

Feature extraction for tomato disease identification uses data such as tomato leaf images to represent healthy and diseased plants. This describes tomato disease detection feature extraction. Image pre-processing improves image quality and standardization before feature extraction. Resizing, color normalization, noise reduction, segmentation, and other techniques from image pre-processing may be used. First, select tomato leaf areas of interest (ROIs) if the image has many leaves. Image segmentation or leaf boundary delineation can do this next, identify tomato leaf disease symptoms in each ROI. Image processing algorithms like deep learning can segment diseases spots, lesions, discoloration, and other apparent indicators of disease. Localized symptoms of disease can be analyzed using feature extraction approaches. Common tomato disease detection feature extraction methods include:

a. Colour Features: Color histograms, color moments, and color channel statistics can show color differences between healthy and unhealthy regions.

b. Texture-based Features: Use Gabor filters, LBP, or GLCM statistics to define healthy and diseased areas' texture patterns.

c. Shape-based Features: Calculate disease symptoms' geometric features such area, perimeter, circularity, and aspect ratio.

d. Statistical Features: Extract mean, standard deviation, skewness, and kurtosis from disease symptom pixel intensities to record their statistical distributions.

e. Deep Learning Features: Use VGG, ResNet, or Inception pre-trained CNNs to extract high-level features from localized illness symptoms. Feature vectors are CNN intermediate layers.

Depending on the collected features' dimensionality and modelling needs, feature selection or reduced dimensionality may reduce feature space complexity. This eliminates unnecessary data and improves model efficiency and generalization. Numerical feature vectors represent extracted features. Each tomato leaf ROI has traits that describe its disease symptoms. Machine learning or deep learning models are trained using extracted feature vectors and their labels signifying health or sickness. These models discover diseases class patterns using extracted features. Based on extracted attributes, trained algorithms may forecast disease in new tomato leaf samples. Feature extraction from tomato leaf images distinguishes healthy and sick plants. It allows the creation of precise and reliable disease detection models for tomato agriculture disease control.

Classification: Image classification refers to the process of determining what something depicted in an image symbolizes. It is possible to train a model to recognize different kinds of images through the use of a categorization model. You may train a model, for instance, to recognize photographs depicting three distinct species of leaves.

The image pre-processing methods that have been described so far can be used to carry out the same procedure on tomato leaf leaves and categorize them in accordance with their traits.

Deep learning algorithms are used in studies to detect and manage diseases in fruits, vegetables, and field crops early. CNN models are more sensitive to key features and can analyze leaf diseases [11], Due to its several feature extraction processes, Accurate categorization is produced using CNN deep learning. The most recent Deep CNNs and meta-architectures, including VGG Nets, Le-nets ResNet, R-CNN, Mask R-CNN, FCN, SSD, and other well-known object identification and segmentation architectures, were used to forecast tomato leaf disease. But they moved too slowly to produce better outcomes. These developments made the Convolutional Neural Network faster and more accurate. CNN leaf disease detection architecture is shown in Figure 3.

4.1 LeNet

After a first fully linked layer, the LeNet design moves onto convolutional and pooling layers. By adopting a condensed connection layer known as max-pooling to reduce picture size between convolutional layers, overfitting can be avoided and CNNs can train more effectively [12]. It has been used to recognize handwritten numbers, traffic signs, and even human faces [13]. The LeNet CNN is a straightforward but efficient model. Even though LeNet's design hasn't evolved much since it was built more than 20 years ago, it is still in broad usage as seen in Figure 4.

In order to classify images, the LeNet convolutional neural network (CNN) architecture is frequently utilized. It is less useful for tomato leaf disease detection than more advanced designs. This article explains LeNet and provides pertinent examples of its use:

LeNet design is LeNet has convolutional, pooling, and classification layers. For preliminary image categorization, the architecture is simple and efficient. Modern CNN architectures have more layers and parameters. It also classifies plant diseases and identifies plant stress. Such as: In "A Review on Detection of Plant Diseases using Image Processing Techniques," tomato diseases were classified using LeNet. The researchers diagnosed several illnesses using LeNet on tomato leaf photos.

3.png

Figure 3. Deep convolutional neural networks architecture

4.png

Figure 4. LeNet architecture

Limits and Challenges:

LeNet may miss subtle patterns and subtleties due to its shallowness. Tomato leaves might hide modest diseases signals.

LeNet's simplicity may prevent it from representing all tomato leaf disease features and variants, resulting in lower accuracy than more complicated structures.

Lighting, camera position, and disease type can complicate tomato leaf disease data. LeNet could find this complexity challenging. Pre-trained models for tomato leaf disease detection may be harder to find because LeNet is an older design than other models. Recent CNN architectures like ResNet, VGGNet, and GoogleNet have become popular due to their deeper structures and improved performance. These patterns help researchers detect complex diseases and improve precision.

Thus, while LeNet can still be utilized for tomato leaf disease identification, academics and practitioners prefer more contemporary systems tailored for disease detection and other complex image classification tasks.

4.2 GoogleNet

There are a total of 27 layers in the GoogleNet Model in Figure 5. Others, such as the Convolutional and Fully-Connected layers, are parameterized while others, such as the Max-Pooling layer, are not. This module is built using a variety of tiny convolutions that greatly minimize the number of parameters. Their method is based on a 22-layer deep CNN even though they reduced the number of parameters from 60 million (Alex Net) to merely 40 million.

GoogleNet, a CNN architecture that pioneered Inception modules, categorizes images quickly and accurately. It has been used in comparable conditions but not to detect tomato leaf infections. GoogleNet is introduced and used to identify tomato leaf diseases. Inception modules distinguish GoogleNet Architecture. These modules' several filter sizes (1×1, 3×3, and 5×5) and parallel convolutions collect features at different scales. It maximizes calculation speed and precision. GoogleNet, a member of the Inception family, has been used to detect tomato leaf diseases. This architecture has been used for plant disease classification, but tomato leaf disease detection is rare. Some examples: The researchers [14] used GoogleNet to classify diseases of plants including tomato leaf diseases.

Challenges and Limitations:

There is a dearth of material dedicated to the detection of tomato leaf diseases using GoogleNet or Inception architecture. This points to the necessity for specialized study in this area.

Fewer pre-trained models may be available for tomato leaf disease detection using GoogleNet than other popular architectures such as ResNet or VGGNet. This can make it harder to learn and adopt new skills.

The availability, breadth, and diversity of datasets on tomato leaf diseases varies widely. An enormous quantity of labeled data is typically required to successfully train a deep architecture like GoogleNet. Performance and generalization may be affected if the dataset is small or unbalanced.

Deeper designs, such as GoogleNet, can be computationally intensive due to their intricate design. It can be difficult to conduct training and inference in contexts with limited resources due to the potential need for increased processing resources, such as high-end GPUs.

5.png

Figure 5. Google net architecture

4.3 AlexNet

Although AlexNet is larger and more complex than LeNet, the network architecture shown in Figure 6 is remarkably comparable to the use of numerous Convolutional Layers. Soon after its first release, the AlexNet architecture demonstrated excellent performance when used with big image datasets. AlexNet includes five convolutional layers in total, including two dropout layers, three fully connected layers, and two max-pooling levels.

In 2012, Alex Krizhevsky & colleagues announced AlexNet, a revolutionary CNN architecture. In order to identify plant diseases and generally classify images, it has been widely used. Use AlexNet to detect tomato leaf disease. The AlexNet architecture consists of convolutional, maximum pooling, and totally linked layers. Deep learning became well-known in computer vision thanks to the ground-breaking result of the ImageNet Large-Scale Visual Recognition Challenge.

Diagnostics for Plant Disease Although AlexNet has achieved success in more challenging plant disease detection tasks, there are few examples of tomato leaf disease detection. Some instances: According to the study "Plant Disease Recognition Using a Convolutional Neural Network," Alex Net was able to identify tomato plant diseases such citrus canker and bacterial leaf spot. The classification of disorders was excellent. The study [15] used AlexNet as a feature extractor.

6.png

Figure 6. AlexNet architecture

Issues and Limitations: Applying AlexNet to tomato leaf disease detection or similar activities has various issues and limitations.

4.4 VGGNet

Input images up to 224×224 pixels in size can be processed using its 4096 convolutional features. While VGGNet excels at particular tasks, Large-filter CNNs are not the ideal solution for most image classification problems; CNN designs like Google Net's AlexNet architectures. This is particularly the case for input photographs ranging in size from 100×100 to 350×350 pixels [16] as seen in Figure 7. The VGGNet CNN architecture competed and won the ILSVRC 2014 classification competition, demonstrating its usefulness in real-world applications. When it comes to computer vision, many applications can benefit from the use of VGG CNN because of its processing efficiency [17]. It can be employed, among other things, for object detection. Its comprehensive feature representations are useful for many kinds of neural network architectures (YOLO, SSD, etc.).

VGGNet is a deep convolutional neural network design. A well-known image classification is VGGNet. tool and plant disease detection, although tomato leaf disease detection applications are scarce. VGGNet may identify tomato leaf disease. The uniform VGGNet Architecture has numerous convolutional layers with modest 3x3 filters followed by max pooling layers.

Plant diseases, in particular leaf diseases, have been successfully identified by VGGNet. There aren't many instances, despite the fact that VGGNet's ability in identifying plant diseases suggests it might be used to identify diseases in tomato leaves. Some instances: In "Deep Learning Approaches for Plant Disease Detection and Diagnosis," scientists examined tomato leaf disease using the VGGNet architecture.

In the study [18] classified using VGGNet, successfully diagnosed disorders of tomato leaves. Transfer learning was used to enhance a pre-trained VGGNet model for precise disease classification on a dataset of tomato leaf diseases.

Challenges and Limitations: Using VGGNet for tomato leaf disease detection or related tasks, such as Due to its complex nature, VGGNet has more parameters than other architectures. In order to train and infer with a VGGNet, powerful GPUs and CPUs are required.

4.5 ResNet-101

Simple two of the deep learning problems that ResNet was used to tackle at Microsoft Research Asia in 2016 and 2017 were statement completion and machine understanding [19] as seen in Figure 8. Disciplines in ResNet Both ResNet-50 and ResNet-101 are supported. Microsoft's machine comprehension system uses CNNs to respond to one hundred thousand inquiries across twenty different domains. To match the processing power of GPUs, ResNet's CNN design may be scaled to meet your needs.

Microsoft Research developed ResNet-101, a CNN architecture. ResNet-101 is frequently used in image despite the limited applications for tomato leaf disease detection, categorization and plant disease detection are both possible. Using ResNet-101, tomato leaf disease may be found. ResNet-101 Architecture is 101-layer ResNet, among others Residual connections let the network learn residual mappings and solve the vanishing gradient problem. These connections help train deep networks and collect complicated information. ResNet-101 has effectively detected plant diseases, including leaf diseases. Although ResNet-101's efficacy in detecting plant diseases supports its usage in detecting tomato leaf diseases, there aren't many examples. As an instance, consider [20] to identify tomato leaf disease, employed ResNet-101. With the use of a skilled ResNet-101 model, tomato leaf diseases could be distinguished with a high degree of accuracy [20, 21].

Challenges and Limits:

ResNet-101 has some drawbacks when used for tomato leaf disease detection:

ResNet-101 is a deep architecture with several levels and parameters. ResNet-101 requires high-end GPUs for training and inference.

7.png

Figure 7. VGGNet architecture

8.png

Figure 8. ResNet101 architecture

9.png

Figure 9. ResNet152 architecture

4.6 ResNet-152

By learning the residual representation functions rather than the signal representation directly, ResNet-152 can have an extraordinarily deep network with up to 152 layers, as seen in Figure 9. ResNet-152 uses a skip connection, also known as a shortcut connection, to transport data directly from one layer to the next without altering it in any way [20-22].

While both AlexNet and ResNet-152 use over 60M parameters, the top-five accuracy between the two networks is roughly 10% different. ResNet-152 training, however, necessitates a great deal of calculations (about 10 times that of AlexNet), necessitating longer and more taxing training sessions. With more parameters and FLOP than ResNet-152, VGGNet also performs worse. Training an inaccurate VGGNet requires more time.

Table 1. Statistical features comparison of different CNN architectures

Deep convolutional neural network (CNN) ResNet-152 makes use of ResNet (Residual Network). For general image categorization and plant disease detection, ResNet-152 has been extensively employed. The 152-layer ResNet architecture is known as ResNet. The network can learn residual mappings and resolve the vanishing gradient problem thanks to residual connections. These connections aid in the training of deep networks and the acquisition of complex data. Plant Disease Detection ResNet-152 has been used to detect plant diseases, although tomato leaf disease detection instances are scarce. However, the architecture's plant disease detection capabilities suggest its use for tomato leaf diseases. It is [23, 24] used a ResNet-152 version to detect plant diseases. The study showed ResNet-152's accuracy in diagnosing plant diseases, not just tomato leaf diseases.

Challenges and Limitations: When using ResNet-152 for tomato leaf disease detection or related tasks, consider the following:

ResNet-152 is a deep architecture with several layers and parameters. High-end GPUs are needed for ResNet-152 training and inference.

The Table 1 represents the comparison different CNN architectures with respect to statistical features.

For the proposed study, we developed a convolutional neural network (CNN)-based model for disease identification in tomato crops. The suggested CNN-based design uses three convolution and maximum pooling layers with different numbers of filters. The PlantVillage dataset provided us with data on tomato leaves that we utilized in our tests. The dataset includes nine distinct diseases and a tenth "healthy" category. We employed data enhancement methods to ensure that each student had access to an equivalent number of high-quality images, and we evaluated CNN models and calculated corresponding evaluation parameters based on each disease class

The PlantVillage Dataset provided the input data. One healthy label and nine disease labels (which include information about bacterial spot, black leaf mold, gray leaf spot, late blight, and powdery mildew) make up the 10 classes or labels in the data set. As seen in Table 2, the original dataset contains 14,529 images of tomato leaves in various arrangements.

As per Table 2, out of the total of 14,529 pieces of information, 80% will be set aside for use in the training phase, and 20% will be used in the testing phase, respectively. The validation data section will make up 80% of the whole, whereas the testing data component will only make up 20%.

ResNet-152's performance is compared to that of ResNet-101, VGGNet, GoogleNet, AlexNet, and LeNet using a variety of criteria, including Accuracy and Precision, Recall, F1-Score. The metrics for the comparable models were gathered and are explained below.

1. Accuracy: The ratio of accurate reports to all predictions is known as accuracy.

The number of times the model is accurate when the predicted label matches the actual label. A false negative occurs when a model predicts the incorrect label for a specific case. The number of cases where the model accurately predicted a false label despite the fact that the actual label was true is referred to as the "True Negative". The number of False Positives in Eq. (1) is the number of occasions where the model correctly predicted the label but the label was incorrect.

Accuracy $=\frac{T P+T N}{T P+F P+T N+F N}$    (1)

2. Precision: By dividing the total number of positive predictions by the number of correct forecasts in Eq. (2), as shown in Table 3, one can determine the accuracy of a model's class predictions.

Precision $=\frac{ {TRUE \, POSITIVE }}{{ TRUE\, POSITIVE }+ { FALSE\, POSITIVE }}$     (2)

The proportion of the test set's total examples that belong to each class is used to calculate that class's weight.

3. Recall: Correctly identified observations as a percentage of total observations are one definition of accuracy in Eq. (3). The problem is that class differences in performance may not be accurately reflected. As a result, in Table 3, accuracy is just one of several metrics used to rank models these days.

Recall $=\frac{T R U E \text { POSITIVE }}{\text { TRUE POSITIVE }+ \text { FALSE NEGATIVE }}$    (3)

4. F1-Score: If the F1 score is high, then means the output forecasts are accurate with few erroneous positive and negative results. When the F1 score drops below Table 3, the model is considered a complete failure in Eq. (4).

Recall $=\frac{ {TRUE \,POSITIVE }}{{ TRUE\, POSITIVE }+ { FALSE\, NEGATIVE }}$    (4)

The evaluated CNN model of ResNet-152, ResNet-101 and VGGNet calculated the corresponding evaluation parameters based on each disease class, as shown in Figures 10-12.

In conclusion, both the ResNet-152 and the ResNet-101, both of which are VGGNet neural networks, have been implemented. The following factors can be taken into account when comparing ResNet-152, ResNet-101, and VGGNet for tomato leaf disease detection:

Table 3. Accuracy classes summary of precision, F1-score and recall for ResNet-152, ResNet-101 and VGGNet

The computing requirements of ResNet-152 and ResNet-101 are high because of their depth and number of parameters. It's important to note that VGGNet, especially in its deeper iterations, can be very computationally demanding.

10.png

Figure 10. Performance evolution metrics for precision

11.png

Figure 11. Performance evolution metrics for F1-score

12.png

Figure 12. Performance evolution metrics for recall

The more sophisticated and abstract properties of tomato leaves can be captured by models like ResNet-152 and ResNet-101. Tomato leaf disease traits are among those that can be captured by VGGNet's uniform design and tiny convolutional filters.

The increased complexity of deeper architectures like ResNet-152, ResNet-101, and VGGNet makes it more difficult to evaluate the learnt features and their connection to disease patterns in tomato leaves. These designs are popular because they can use pre-trained models. Rather than spending instead of spending more time and effort learning from scratch, you can use these pre-trained models and refine them using transfer learning on particular datasets related to tomato leaf disease.

The available dataset, computational resources, and the intended trade-off between accuracy and efficiency are only a few of the elements that influence the selection of the design. VGGNet strikes a good mix between depth and interpretability, while ResNet-152 and ResNet-101 give greater depth and may be able to capture more nuanced features. When deciding on an appropriate architecture for the tomato leaf disease detection task, it is important to consider both the availability of pre-trained models and the task-specific needs.

Their accuracy on pre-trained networks has been tested using 50 and 80 epochs, and records of a range of events have been kept. ResNet-152 has been successfully trained to categorize Utilizing pictures of tomato plant leaves from the plantvillage dataset, a variety of diseases that can harm tomato plant leaves are shown. The best classification performance can be reached when ResNet-152 is trained with a variety of various batch sizes. Additionally, the experiment results show that it is possible to get the highest accurate categorization performance with a learning rate of 0.001 and a data division ratio of 80% to 20% between training and testing. In preparation for future work, each and every class in the Plant Village Data set will be tasked with identifying each and every sickness.

13.png

Figure 13. Instance I with 50-epochs

14.png

Figure 14. Instance-I with 80-epochs

15.png

Figure 15. Instance-II with 50-epochs

The graphs in Figures 13-16 are drawn based on the result obtained, that reveals, irrespective of the epochs and instances ResNet-152 is found to have the highest accuracy than ResNet-101 and VGGNet at most of the points.

16.png

Figure 16. Instance-II with 80-epochs