Page 59 - 2025S
P. 59
52 UEC Int’l Mini-Conference No.54
Figure 2: Architecture of the Proposed Model: DenseNet201
plication for medicinal flora recognition, ex- curacy offering faster performance compared to
perimenting with various DL models, includ- MobileNet [8].
ing CNN, SVM, PNN, and FNN. Among these, Moreover, [30] introduced the SDAMPI algo-
the CNN model achieved the highest accuracy
of 99.70%. Flutter was used to design the rithm for identifying immunity-boosting medic-
inal plants, attaining 96.00% accuracy. Pre-
application’s graphical user interface [13]. In trained CNN models such as VGG16 were em-
[14], the PB3C (Big Bang–Big Crunch) algo- ployed by [28] and [33], while [26] explored trans-
rithm achieved an accuracy of 93.20%, showcas- fer learning with ResNet50, VGG16, and Mo-
ing the effectiveness of CNNs in plant species
bileNet2, with ResNet50 achieving the best per-
identification. In the same way, [32] proposed
formance at 91.00%. However, most of these
a custom ANN model that achieved 98.30% studies lacked a deployable application. By
accuracy, despite challenges posed by features contrast, [1] developed a fully functional, user-
like leaf morpho-colorimetry and visible/near- friendly application for herbal plant recognition.
infrared spectroscopy. On the flip side, [20] pre- Correspondingly, [2] presented a mobile app that
sented an IoT-based approach with a DCNN,
classifies herbs based on geometric features. [15]
which outperformed standard CNNs by achiev-
applied MobileNetV2 for classifying diseases,
ing 98.45% accuracy. In addition, another study medicinal plants, and fruits, demonstrating im-
argued that conventional CNNs remain effective, proved learning rate optimization and multi-
with a custom MNN model achieving 85.15% ac-
functional detection. Furthermore, feature se-
