Page 559

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

PSO-Cascadenet: An Intelligent Hybrid Deep Learning Model for

Medicinal Plant Classification

Prof. Ashwani Utture, Mr. Om Suryawanshi, Mr. Vikas Phatangare, Mr. Shubham parbhane

Department of Computer Engineering, Nutan Maharashtra Institute of Engineering and Technology,

Pune, India

DOI:

https://doi.org/10.51583/IJLTEMAS.2026.150400050

Received: 12 April 2026; Accepted: 17 April 2026; Published: 06 May 2026

ABSTRACT

In the modern technology-oriented world, identifying medicinal plants has become very important for healthcare,

biodiversity preservation, and the development of natural medicines. Traditional methods of plant identification

mainly depend on expert knowledge and manual inspection, which makes the process slow and sometimes

inaccurate. To address these challenges, Pso-CascadeNet presents an intelligent deep learning–based system that

can recognize medicinal plants using images of their leaves. The system uses Convolutional Neural Networks

(CNNs) to extract visual patterns from images, Particle Swarm Optimization (PSO) to automatically tune model

parameters, and Support Vector Machines (SVM) to improve the accuracy of classification. A simple and

interactive interface built with Streamlit enables users to upload leaf images and receive instant predictions,

while FastAPI supports smooth backend communication and deployment. Performance evaluation using metrics

such as accuracy, precision, recall, and F1-score demonstrates that the hybrid CNN–PSO–SVM model performs

better than traditional classification techniques. Overall, the proposed framework offers a dependable, scalable,

and user-friendly approach for digital identification of medicinal plants, benefiting research, learning, and

sustainable use of herbal resources.

Keywords-Herbal Plant Recognition, Convolutional Neural Networks (CNN), Particle Swarm Optimization

(PSO), Support Vector Machines (SVM), Deep Learning Techniques, ImageBased Classification, Visual Feature

Extraction, Streamlit-Based User Interface, FastAPI-Based System Deployment, AI-Powered Plant

Identification.

INTRODUCTION

For many centuries, medicinal plants have played an important role in healthcare due to their therapeutic

properties, and they continue to be essential in modern herbal medicine and pharmaceutical development.

Accurately identifying medicinal plant species, however, remains challenging because many plants share similar

characteristics such as leaf shape, color, and texture. This similarity makes manual identification difficult and

often requires specialized botanical expertise. In addition, environmental conditions including climate, soil type,

and the stage of plant growth can alter the appearance of leaves, further complicating accurate recognition.

Traditional methods of plant identification are generally slow, susceptible to human error, and not easily scalable

for research or practical use. Recent progress in Artificial Intelligence (AI) and deep learning has made

automated plant recognition more effective, particularly through image-based analysis. The Pso-CascadeNet

system introduces an AI-driven hybrid framework that integrates Convolutional Neural Networks (CNN) for

analyzing visual patterns in leaf images, Particle Swarm Optimization (PSO) for optimizing model parameters,

and Support Vector Machines (SVM) for accurate final classification. This combined approach improves

classification accuracy, minimizes overfitting, and enhances the model’s ability to generalize across different

datasets.

The system also includes a user-friendly interface developed with Streamlit, allowing users to upload leaf images

and obtain real-time identification results. FastAPI is used to manage backend communication, ensuring efficient

deployment and system scalability. By providing a quick, reliable, and easy-to-use solution for medicinal plant

recognition, Pso-CascadeNet supports herbal research, increases awareness of natural medicinal resources, and

Page 560

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

encourages sustainable healthcare practices.

LITERATURE SURVEY

Identifying medicinal plants accurately is still a significant challenge because many species share similar

morphological characteristics, and environmental conditions can alter the structure and appearance of leaves. To

overcome this issue, Islam and Rahman proposed the PSO-CascadeNet framework, which combines

Convolutional Neural Networks (CNN) with Particle Swarm Optimization (PSO). In this approach, PSO

dynamically adjusts the model parameters to ensure consistent and reliable feature extraction across different

image conditions. Their findings showed improved recognition accuracy, although the training process required

considerable computational resources.

In a related study, the HerbGuard system introduced an ensemble architecture that integrates EfficientNetV2-S

with Vision Transformers. This method demonstrated the advantages of combining hierarchical convolutional

features with attention-based global relationships for detailed plant recognition. While HerbGuard achieved

strong performance in fine-grained classification tasks, it relies heavily on powerful GPU resources, making it

less practical for lightweight or field-level applications.

Other studies have explored more specialized methods for plant identification. For instance, Karnik and Nair

proposed a Multi-Scale Venation Pattern Analysis technique, which focuses on analyzing leaf vein structures to

improve recognition when plant species appear visually similar. However, the effectiveness of this approach

decreases when the venation patterns are not clearly visible. Additionally, a comparative study by Sibiya

examined the performance of DenseNet, CNN, and MobileNet models for medicinal plant classification. The

results indicated that DenseNet performs well with smaller datasets, while MobileNet supports faster real-time

predictions but with a slight reduction in accuracy.

Overall, previous research indicates that hybrid deep learning models combining CNN, PSO, and SVM

techniques can offer better adaptability and efficient deployment for plant recognition systems. These findings

directly influence the design of the Pso-CascadeNet framework, which aims to provide accurate, scalable, and

real-time identification of medicinal plants.

Dataset Use

This section explains the datasets used in earlier medicinal plant classification studies as well as those adopted

for the proposed Pso-CascadeNet framework. Since the system is designed to classify medicinal plants using

leaf images, the dataset selection emphasizes high-quality labeled images that include a variety of leaf shapes,

venation structures, and environmental conditions. These datasets support the training, validation, and testing

stages of the hybrid CNN–PSO–SVM model, enabling the system to evaluate its performance under different

scenarios.

The datasets referenced in previous studies include the following:

1. Indian Medicinal Leaves Image Dataset (Mendeley Data): This dataset contains more than 9,000 images

representing over 30 medicinal plant species commonly found in India. The images are captured under

different lighting conditions and include both the front and back sides of leaves. Each plant category contains

approximately 250–300 images. Due to its rich representation of leaf texture and venation structures, this

dataset is widely used in medicinal plant recognition research.

2. MED117 Dataset: The MED117 dataset consists of 117 medicinal plant species, with several leaf images

taken from multiple viewing angles to represent real-world variability. Because of its diverse images and the

similarity between certain plant species, this dataset is frequently used in modern deep learning studies for

evaluating classification performance.

3. Flavia Leaf Dataset: The Flavia dataset includes 1,907 leaf images belonging to 32 botanical species. It is

commonly used as a benchmark dataset for leaf shape recognition and contour-based plant classification

Page 561

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

methods. This dataset is especially useful for training early CNN feature extraction layers and comparing

model accuracy.

4. LeafSnap Dataset: The LeafSnap dataset contains more than 30,000 leaf images collected using mobile

devices in natural environments. The dataset includes both scanned leaves and images captured directly in

outdoor conditions. This diversity helps evaluate system robustness when dealing with noise, shadows, and

complex backgrounds.

Overall, the Indian Medicinal Leaf Dataset, MED117, Flavia Leaf Dataset, and LeafSnap serve as the primary

benchmark datasets for training and validating the Pso-CascadeNet medicinal plant classification model.

Together, they represent a wide range of leaf shapes, textures, and venation patterns captured under different

environmental and lighting conditions. These datasets enable the system to learn both detailed structural features

and broader botanical variations, which are essential for accurate plant classification.

In addition, the integration of a custom leaf dataset strengthens the evaluation process by providing real-world

samples collected through mobile cameras in natural environments. This ensures that the model performs

effectively outside controlled laboratory conditions. While the Indian Medicinal Leaf and MED117 datasets help

the model learn clearly labeled medicinal species, the Flavia and LeafSnap datasets improve robustness testing

under complex backgrounds and noisy conditions. The custom dataset further evaluates real-time system

performance, responsiveness, and deployment feasibility in field applications.

Together, these datasets create a strong foundation for training, benchmarking, and validating the Pso-

CascadeNet CNN– PSO–SVM framework, ensuring high classification accuracy, strong generalization

capability, and practical applicability for real-world medicinal plant identification.

Objectives

Based on the analysis of existing research, it is clear that medicinal plant identification systems still have

considerable potential for improvement, especially in terms of accuracy, adaptability to different environmental

conditions, and real-time usability.

The main goal of this study is twofold. First, it aims to design a hybrid deep learning framework capable of

accurately identifying medicinal plants using leaf images. Second, it focuses on ensuring that the system can be

deployed in an efficient, scalable, and user-friendly manner suitable for applications in education, agriculture,

and healthcare.

1. Development of Hybrid Classification Model To design and implement a CNN–PSO–SVM hybrid model

that can effectively extract distinctive leaf features, enhance classification accuracy, and minimize the

risk of overfitting.

2. Hyperparameter Optimization using PSO To utilize Particle Swarm Optimization (PSO) for automatic

tuning of model hyperparameters, aiming to achieve a medicinal plant recognition accuracy of at least

95%.

3. Real-Time User Interface Implementation To develop a Streamlit-based interactive interface that allows

users to upload leaf images and receive plant identification results in real time, with an inference response

time of less than 2 seconds.

4. Backend Integration and Deployment To integrate FastAPI for efficient backend communication,

enabling lightweight deployment and easy accessibility of the model across multiple platforms, including

mobile devices and edge computing environments. These objectives collectively aim to create a reliable

and practical system for accurate medicinal plant identification while ensuring usability and scalability

in real-world scenarios.

Page 562

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

METHODOLOGY

Pso-CascadeNet operates as an intelligent plant recognition system designed to identify various medicinal plant

species through the analysis of leaf images. The framework employs a hybrid artificial intelligence architecture

in which Convolutional Neural Networks (CNNs) learn important leaf characteristics, Particle Swarm

Optimization (PSO) adjusts model parameters to improve performance, and Support Vector Machines (SVM)

perform the final classification of plant species.

A Streamlit-based web interface enables users to upload leaf images and receive real-time predictions, making

the platform convenient for students, farmers, researchers, and herbal medicine practitioners. In addition,

FastAPI is used for backend model deployment and communication, allowing efficient operation in both local

and cloud environments. By combining deep learning capabilities with optimization techniques and an easy-to-

use interface, Pso-CascadeNet offers a reliable and practical solution for medicinal plant identification.

System Architecture

The complete workflow of the Pso-CascadeNet system is organized into several layers:

Image Input Layer:

Users submit leaf images through a simple and interactive Streamlit interface. Before processing, the images

undergo standard preprocessing steps such as resizing, normalization, and noise reduction to ensure consistent

input quality for the model.

Feature Extraction (CNN):

A Convolutional Neural Network extracts significant visual features from the leaf images, including shape,

texture, edges, and venation patterns. These features are transformed into informative feature maps that represent

the leaf characteristics.

Optimization Layer (PSO):

Particle Swarm Optimization is applied to adjust CNN hyperparameters—such as learning rate, filter size, and

network configuration—to enhance model performance and minimize overfitting.

Classification Layer (SVM):

The optimized feature vectors generated by the CNN are then provided to a Support Vector Machine classifier,

which determines the final plant species label with high accuracy, especially when dealing with visually similar

plants.

Backend Processing (FastAPI):

FastAPI manages the communication between the model and the user interface by handling inference requests

and responses. This ensures quick processing, low latency, and compatibility with mobile devices and cloud-

based deployment.

User Interface (Streamlit):

The prediction results are displayed through the Streamlit interface, showing the identified plant name along

with its confidence score, allowing users to interact with the system easily in real time.

Figure 1 presents the architectural workflow of the Pso-CascadeNet system developed for efficient medicinal

plant identification using leaf images. The framework follows a structured pipeline consisting of image

acquisition, preprocessing, deep feature extraction, parameter optimization, and final classification. Each

component—input interface, CNN-based feature extraction, PSO-based optimization, and SVM classification—

Page 563

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

works together to ensure reliable learning and accurate predictions across diverse leaf samples.

The architecture is designed to support real-time plant identification through a user-friendly interface powered

by Streamlit, while FastAPI enables smooth backend communication. By integrating deep learning techniques

with optimization strategies and lightweight deployment, the Pso-CascadeNet system offers a scalable and

practical solution for botanical research, healthcare awareness, and field-level plant recognition.

Figure 1. System Architecture of the Proposed PSO-cascadeNet Framework

Performance Metrix

To assess the effectiveness, reliability, and efficiency of the proposed Pso-CascadeNet medicinal plant

classification system, multiple evaluation metrics are employed. These metrics are divided into two main

categories: (1) Classification Performance Metrics, which evaluate how accurately the system identifies plant

species from leaf images, and (2) Computational Efficiency Metrics, which measure the model’s speed,

optimization performance, and suitability for real-time deployment. Together, these evaluation measures provide

a balanced assessment of both predictive accuracy and practical usability of the system.

Classification Performance Metrics

Precision (Confidence) Precision evaluates the accuracy of the model’s positive predictions, indicating how

many of the predicted plant species are actually correct.

Precision = TP / (TP + FP). Where: TP = True Positives FP = False Positives

Recall (Sensitivity) Recall measures the system’s ability to correctly identify all samples belonging to a specific

Page 564

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

plant species.

Recall = TP / (TP + FN) Where: FN = False Negatives

False Positive Rate (FPR) The False Positive Rate indicates how frequently the model incorrectly assigns a leaf

sample to the wrong species category.

FPR = FP / (FP + TN) Where: TN = True Negatives.

Specificity measures the model’s ability to correctly identify samples that do not belong to a particular class.

Specificity = TN / (TN + FP).

F1-Score The F1-score represents the harmonic mean of Precision and Recall and is especially useful when

dealing with imbalanced datasets.

F1 = 2 * (Precision * Recall) / (Precision + Recall)

AUC (Area Under the Curve) The AUC value reflects the model’s capability to distinguish between different

plant categories. A higher AUC score indicates better classification performance.

AUC = Area under ROC curve

Geometric Mean (G-Mean) The Geometric Mean measures balanced classification performance, particularly in

binary classification tasks.

G-Mean = sqrt(Sensitivity * Specificity) Sensitivity = Recall = TP / (TP + FN).

Computational Efficiency Metrics

Inference Time (IT)

IT = T_output - T_input

Inference Time measures the time required by the system to generate a prediction after a leaf image is uploaded.

Lower inference time indicates better real-time performance.

PSO Optimization Overhead (PO)

PO = T_optimized - T_baseline

This metric measures the additional computation time introduced by the PSO algorithm during CNN parameter

optimization.

Model Size (MS)

MS = Total_Parameters * Precision_bits

Model Size represents the storage requirements of the trained model and determines its suitability for deployment

on mobile or edge devices.

Memory Utilization Efficiency (MUE)

MUE = (Memory_used / Memory_available) * 100

This metric evaluates how efficiently memory resources are used during feature extraction and classification.

Page 565

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue IV, April 2026

Throughput (TH)

TH = N_images / T_total

Throughput measures the number of images the system can classify per second during continuous operation.

These performance and efficiency metrics collectively provide a comprehensive evaluation framework for the

Pso-CascadeNet CNN–PSO–SVM model, ensuring both high classification accuracy and practical real-time

deployment capability for medicinal plant identification.

ACKNOWLEDGMENT

The Pso-CascadeNet system introduces a reliable, efficient, and easy-to-use approach for identifying medicinal

plants through the analysis of leaf images. By integrating Convolutional Neural Network (CNN)–based feature

extraction, Particle Swarm Optimization (PSO) for parameter tuning, and Support Vector Machine (SVM) for

classification, the framework achieves high prediction accuracy and stable performance across multiple datasets.

Experimental findings show that this hybrid approach performs better than conventional CNN and CNN+SVM

models, delivering improved accuracy, well-balanced F1-scores, and strong generalization ability when

evaluated on both benchmark datasets and real-world samples. The training and validation curves indicate stable

model convergence without signs of overfitting, while the confusion matrix demonstrates effective

differentiation between plant species, even when their leaf structures appear visually similar.

Additionally, the integration of a Streamlit-based user interface along with a FastAPI backend allows smooth

realtime interaction, making the system practical for students, researchers, farmers, and herbal practitioners. The

framework effectively handles common challenges such as lighting variations, similarities in leaf texture, and

dataset imbalance through optimized feature extraction and well-defined classification boundaries.

Overall, the Pso-CascadeNet framework demonstrates that combining optimization strategies with deep learning

and advanced classification techniques can significantly improve the accuracy of medicinal plant recognition.

This system supports the promotion of herbal knowledge, encourages sustainable healthcare practices, and

provides accessible plant identification tools suitable for real-world applications.

CONCLUSION

The Pso-CascadeNet system introduces a reliable, efficient, and easy-to-use approach for identifying medicinal

plants through the analysis of leaf images. By integrating Convolutional Neural Network (CNN)–based feature

extraction, Particle Swarm Optimization (PSO) for parameter tuning, and Support Vector Machine (SVM) for

classification, the framework achieves high prediction accuracy and stable performance across multiple datasets.

Experimental findings show that this hybrid approach performs better than conventional CNN and CNN+SVM

models, delivering improved accuracy, well-balanced F1- scores, and strong generalization ability when