Page 61 - 2025S

54 UEC Int’l Mini-Conference No.54

Table 2: Comparison of CNN Model Perfor- mizer function, which is an improvement on the
mances stochastic gradient descent optimization tech-
nique. This function adjusts learning rates for
CNN Models Accuracy (%) each parameter, corrects estimator bias in the
VGG16 93.17 gradient’s first and second moments, and im-
Xception 95.58 proves model stability and performance during
training.
MobileNetV2 97.99
ResNet50 98.59
DenseNet201 99.00

3.3 Hyperparameter setting

To train our DCNN models, this study uses cus-
tom data. 80% of the dataset was used in the
training procedure. In order to track model per-
formance and avoid overfitting, an additional
10% of the data was allocated for validation
at the beginning of each epoch. Additionally,
a completely separate dataset of 10% was used
for testing to provide an objective assessment of
the model’s performance. Besides, new and real
data is used in the test to check if the model
works well with different stuff it hasn’t seen be-
fore. For accurate gradient estimations, a batch
size of 32 was utilized throughout the training
phase, even though it required more memory. Figure 4: The process of classification approach
The ”softmax” function is also used in the out-
put layer to map the final layer’s output and
enhance the predictability of the model. To 3.4 Leaf Features Extraction Analysis
make sure the models always give the same re-
sults, we start the randomness from a specific Feature extraction plays an important role in de-
number, which is 64. In order to ensure con- tecting features in digital images such as edges,
sistent training and evaluation for accurate per- shapes, or motion in DL research. All medicinal
formance comparison, the model’s parameters plant leaves are not the same age. There are
are modified through training based on input mixed-age leaves, like pre-mature, semi-mature,
data and the optimization process. By utiliz- and mature. Selecting these three conditions
ing ImageDataGenerator’s ”horizontal flip” fea- for each species of medicinal plant allows for an
ture, more training data will be available to the in-depth study because the leaves appear differ-
model, helping in its capacity to identify fea- ently at various stages of growth. The types
tures independent of horizontal movements and of medicinal plants for this research are neem,
enhancing the handling of test data with com- moringa, Malabar nut, holy basil, arjun, and
parable movements. For multiclass classifica- green chiretta. Utilizing the features that a
tion, the ”categorical” class mode was employed, neural network was trained on before is quick
which allowed models to understand connections and efficient when performed via feature extrac-
between classes and predict multiple classifica- tion. This research’s main objective is to de-
tions. Images were utilized with 0.2-20% zoom velop an automated system for identifying the
in and out, and the images were randomly ro- accurate medicinal plants that are deployed by
tated with 20% shifting along the X and Y transfer learning models. Additionally, trans-
axes. However, we applied the ”Adam” opti- fer learning serves as a feature extraction and