In this section, we illustrated some experimental results using the standard Convolutional Neural Network. During the experiment, the researcher used 20% of the dataset for testing and 80% for training. The number of filters used by the researcher is 16, 32, and 64, in the first, second, and third convolution layers respectively. In this study, the researcher created a custom CNN model using a 6600 color image dataset. These datasets Labeled with their parts with their corresponding disease type of plants.
In recent years, deep convolutional neural networks have demonstrated improved performance in image classification. As a result, the experiments are carried out in the simple model of the three-layer CNN, and all experiments are carried out in hyper parameters. In our experiment, we feed a model with a set of down-sampled RGB images with dimensions of 128 x128 x3. To create a custom CNN model, the researcher used a 3x3 kernel size and a 2x2 max pooling parameter. The researcher conducted various experiments on the convolutional neural network model. During the experiment, the researcher divided the data set 70/30, 80/20, and 90/10. For training and testing purposes, respectively.
Table 3
Comparison of different batch sizes
Bach size | Training accuracy | Validation accuracy | Loss | Validation loss | Hamming loss |
64 | 92.8% | 90.08% | 0.32% | 0.24 | 0.4% |
32 | 99.5% | 99.2% | 0.012% | 0.023% | 0.0064% |
As we depicted in the above Table 1–3 the choice of the Bach size affects the performance of the classifier when training the neural network. The smaller the Bach size the more performance of the model. In Our findings using Bach size 32 the model performed 99.5% training accuracy, and 99.2% validation accuracy. On the other hand on Bach size 64 the model performed 92.8% in training and 80.08% for validation accuracy. From the results, it is clear that the model outperformed using Bach size 32.
As we depicted in the above Table 1–3 the choice of the Bach size affects the performance of the classifier when training the neural network. The smaller the Bach size the more performance of the model. In Our findings using Bach size 32 the model performed 99.5% training accuracy, and 99.2% validation accuracy. On the other hand on Bach size 64 the model performed 92.8% in training and 80.08% for validation accuracy. From the results, it is clear that the model outperformed using Bach size 32.
Table 4
Comparison of different optimizations
Optimizer | Training Accuracy | Validation accuracy | Training Loss | Validation loss | Hamming loss |
SGD | 99.6% | 92.6% | 0.01% | 0.024% | 0.0064% |
Adam | 99.5% | 99.2% | 0.012% | 0.023% | 0.0064% |
Adagrad | 99.5% | 99.3% | 0.01% | 0.02% | 0.005% |
Ada delta | 81% | 82.3% | 0.43% | 0.54 | 0.69% |
RMS Prop | 99% | 99.2% | 0.04% | 0.024% | 0.063% |
This section summarizes the finding of the study using different optimizers presented in the following Table 1–4 by using the classification accuracy, validation accuracy, hamming loss, validation loss metrics in the form of training data, and validation data. The proposed model is outperforming in Adam optimizer in iteration per epoch. Whereas the other optimizers Adagrad, Adadelta, and Rmsprop is fluctuating in iteration per epoch. The stochastic gradient descent is performing well when the learning rate is the same. From the results, it is clear that the Adam optimizer is performing well in terms of loss hamming loss and accuracy.
As we depicted in the above Table 1–4 the choice of the Bach size affects the performance of the classifier when training the neural network. The smaller the Bach size the more performance of the model. In Our findings using Bach size 32 the model performed 99.5% training accuracy, and 99.2% validation accuracy. On the other hand on Bach size 64 the model performed 92.8% in training and 80.08% for validation accuracy. From the results, it is clear that the model outperformed using Bach size 32.
Table 5
Comparison of different optimizations
Optimizer | Training Accuracy | Validation accuracy | Training Loss | Validation loss | Hamming loss |
SGD | 99.6% | 92.6% | 0.01% | 0.024% | 0.0064% |
Adam | 99.5% | 99.2% | 0.012% | 0.023% | 0.0064% |
Adagrad | 99.5% | 99.3% | 0.01% | 0.02% | 0.005% |
Ada delta | 81% | 82.3% | 0.43% | 0.54 | 0.69% |
RMS Prop | 99% | 99.2% | 0.04% | 0.024% | 0.063% |
This section summarizes the finding of the study using different optimizers presented in the following Table 1–5 by using the classification accuracy, validation accuracy, hamming loss, validation loss metrics in the form of training data, and validation data. The proposed model is outperforming in Adam optimizer in iteration per epoch. Whereas the other optimizers Adagrad, Adadelta, and Rmsprop is fluctuating in iteration per epoch. The stochastic gradient descent is performing well when the learning rate is the same. From the results, it is clear that the Adam optimizer is performing well in terms of loss hamming loss and accuracy.
Figure 6a shows the convolutional neural network model's training and validation accuracy; in the first epoch, the training accuracy is lower than the validation accuracy. The difference between training and validation accuracy is minimal after 25 iterations. Both the training and validation losses are constant for the first 20 epochs in Fig. 6b. The validation loss scaled up after 40 epochs and turned down after 20 epochs. The training loss and the validation loss eventually reach parity. As a result, neither overfitting nor under fitting exists in the suggested CNN model. In unseen data, the model performs well.
Table 6 Normalized Confusion matrix for CNN
A normalized multi-labeled confusion matrix for the convolutional neural network model is shown in Table 6. Using a confusion matrix, the best-verified outcome is hypothesized during the experimentation phase. The Root class label, which is one of 17 classes, has the highest accuracy based on plant parts, or 98%. For the remaining classes leaf, flower, seed, fruit, stem, gonorrhea, wound eye disease, anti-cancer, anti-oxidant, abdominal pain, lung disease, common cold, snakebite, stomach ache, and influenza the accuracy of the labels is 88%, 82%, 82%%58%, 82%, 66%, 90%, 75%, 72%, 62%, 85%, 71%, 79%, 82%, 71%, and 88%, respectively. The model's overall accuracy is 99.5%, and its validation accuracy is 99.2%.
B. Result analysis in Inception V3
The efficient deep learning architecture known as Inception V3, which was created by Google and Google Net, takes its name from a popular internet meme. The proposed architecture has many layers and many neurons, but only one layer performs computation. One thing needs to be taken into account when referring to a deeper network: as the network layers increase, the probability of the model being overfitting increases. To solve this, the inception model ads sparsely connected filters that have multiple layers in a single layer. By accepting 299 x 299 x 3 images as input and 1000 classes as output, the model was trained on the top of the image net dataset, which comprises about a million images from the image database. It is computationally faster than VGG and has 42 total layers. The proposed model for classifying aromatic medicinal plants was trained using 128 x 128 x 3 RGB images from our data set. Last but not least, we trained the inception model by changing the network's last layer to 17 hidden neurons and using a sigmoid as an activation function. The network received a total of 6600 images, and it produced a promising result without any overfitting.
Figures 7a represent the training accuracy of the model is slightly higher than the validation accuracy. The model training is good in 100 epochs. And Fig. 5–5b the training and validation loss is going well. The above graph shows the model is performing well when it reaches epoch 100. As we can see in Figure, the above Inception V3 model is not overfitting. Because there is no significant gap between training and validation accuracy, the validation loss is also decreasing concerning training loss. The model is generalizing all class labels well. Overall, our model inception v3 was the one that obtained the most promising results.
Table 7
Result of the Inception V3 model
Training Accuracy | Validation Accuracy | Training Loss | Validation Loss | Hamming loss |
99.8% | 98.7% | 0.001% | 0.05% | 0.0015% |
Because of the large number of network layers, and the filter concatenation process in the inception V3 model performs well and has adequate computational time and resources, as shown in Table 7. The model's overall training accuracy is 99.8%, and its validation accuracy is 98.7%. The model loss is 0.001% and 0.05%, respectively. Furthermore, the model's extremely low hamming loss indicates that it is working very well at identifying class labels. Slightly superior results are achieved with our model inception V3.