Page 58 - 2025S

UEC Int’l Mini-Conference No.54 51

Table 1: Comparison and Gap Analysis of Prior Work on Medicinal Plant Identification

Reference Model Pre-Processing Accuracy (%) Gap Analysis
[1] MobileNetV2 Resizing, Splitting 98.05 No detection system
[10] Faster R-CNN Resizing 67.34 Single approach
[17] DenseNet201 Smoothing, Sharpening 93.00 No application
[18] VGG16 RGB conversion, Filtering 98.70 No application
[21] Inception v3 Splitting, Resizing 95.00 No application
[23] VGG16 Resizing, Splitting 97.00 Not deployed
[29] DenseNet201 Histogram Equalization 98.58 No implementation
[31] DENN Scaling, Enhancement 98.50 No use-case
Our work DenseNet201 Optimization, Histogram Equalization 99.00 MPIcam application

In earlier studies, traditional classifiers such ture has been an active area in recent years. In-
as Support Vector Machines (SVM), K-Nearest vestigators can utilize pre-trained DL models,
Neighbors (KNN), and Random Forests were transfer learning approaches, or custom-built
commonly employed for plant identification models in their work. Several studies have ex-
based on leaf features like vein patterns and leaf plored the use of CNNs to detect and classify
margins. However, these methods often suffer various types of plants, leaves, crops, fruits, and
from limitations, including poor performance on vegetables based on images [9].
small or limited datasets. In contrast, Convo-
One notable study introduced AyurLeaf, a
lutional Neural Networks (CNNs) have demon-
deep learning-based CNN model designed to
strated superior capabilities in automatically ex- classify medicinal plants by analyzing leaf fea-
tracting complex and high-level features from tures such as shape, size, color, and texture [24].
images [5]. To enhance interpretability in deep This study utilized a secondary dataset, where
learning models, Explainable AI (XAI) frame- the SVM classifier achieved the highest accu-
works such as SHapley Additive exPlanations
racy of 96.76% [12]. Another approach employed
(SHAP) and Local Interpretable Model-agnostic the VGG16 architecture for the classification of
Explanations (LIME) have been applied to ana- herbal leaves from images [23]. Besides, a sepa-
lyze model predictions and risk factors [25], [6].
rate study used YOLOv5 to identify therapeu-
The remainder of this paper is structured as tic plants, achieving a classification accuracy of
follows: Section 2 reviews related work in the 83.00% on a Philippine dataset containing four
field. Section 3 details the proposed research common medicinal plant species. However, that
methodology. Section 4 presents the experimen-
work lacked the development of an automated
tal results along with a discussion. Finally, Sec- application, limiting its real-world usability [27].
tion 5 concludes the paper and outlines potential
directions for future research. To tackle this obstacle, [22] developed a
smartphone application which capable of iden-
tifying the medicinal benefits of a plant leaf
2 Related Work by analyzing its image. They used a dataset
of 30 Indian medicinal leaf species and ap-
For more than sixty years, researchers have plied pre-processing techniques such as resiz-
worked toward enabling machines to understand ing, scaling, and data splitting. A custom
and interpret visual information [22]. The use of CNN model was employed, achieving 94.00%
DL algorithms and image pre-processing tech- accuracy. The app interface was designed us-
niques for image classification, plant disease de- ing Android Studio [22]. Similarly, another
tection [19], [24], and identification in agricul- study implemented an automated mobile ap-