Page 62 - 2025S

UEC Int’l Mini-Conference No.54 55

shape extractor, increasing data efficiency, en-
hancing generalizations, and enabling domain
adaptation from input data. It offers two key
benefits: (i) higher model performance (ii) bet-
ter data and resource efficiency. It reduces the
training time and increases sustainable compu-
tational resources. The most popular pretrained
models are InceptionV3, DenseNet201, VGG16,
VGG19, ResNet50, ResNet121, Xception, Mo-
bileNetV2, and DenseNet121 [3], [11]. Table. 3
shows the DenseNet201 classification report.

4 Experimental result and dis-

cussion

Advanced, automated, and interpretable diag-
nostic tools are necessary for clinical decision-
making due to the rapid growth of medical imag-
ing technologies [16]. A total of 9,150 medicinal
plant leaf images were used in this study. The
dataset was divided into training sets (80%),
validation sets (10%), and testing sets (10%).
The images were captured using an OPPO A76
smartphone. The original image resolution of
1380×780 pixels was resized to 224×224 pix-
els to align with the input requirements of the
DenseNet201 model. The DenseNet201 model
was trained using the Adam optimizer with a
Figure 5: User interface: Input image selection
learning rate of 0.001 over 50 epochs. The train- step
ing process employed a batch size of 32 and uti-
lized the categorical cross-entropy loss function.
Data augmentation techniques, including rota- saved in in.h5 format. This application allows
tion, zoom, and horizontal flipping, were applied users to upload or capture leaf images and re-
to enhance the model’s generalization capabili- ceive immediate identification results, facilitat-
ties. Experiments were conducted on a Win- ing accessible and efficient plant recognition.
dows 10 system equipped with an 11th Gen In- Finally, this study introduced an automated
tel(R) Core(TM) i5-11400 CPU @ 2.60 GHz, application for detecting accurate Bangladeshi
32 GB RAM, and Intel(R) UHD Graphics 730 medicinal plants that overcomes the previous re-
GPU. Additionally, Google Colab was utilized search gap and it has already been mentioned in
for data analysis and model training, leveraging Table.
its GPU acceleration capabilities. The trained
DenseNet201 model achieved a test accuracy 4.1 Discussion
of 99.00%, outscored other evaluated models
such as ResNet50, MobileNetV2, Xception, and In this work, the author examined five transfer
VGG16. This high accuracy underscores the ro- learning models for detecting medicinal plants:
bustness and suitability of the model for real- DenseNet201, MobileNetV2, ResNet50, VGG16,
time medicinal plant identification. A web- and Xception. Based on prior research, these
based application was developed using Stream- models are selected for training. In comparing
lit, integrating the trained DenseNet201 model the training of models, DenseNet201 achieved