Leaf Identification and Classification Using CNN

Leaf Identification and Classification Using CNN

Deeksha U S ,

Student , St Joseph Engineerigng College , Mangaluru

Nishmitha. J

Assistant Professor , St Joseph Engineering College , Mangaluru

Abstract – With the growth of e commerce platforms, reviews submitted online about businesses have become a vital factor impacting consumer decisions and corporate practices. Nevertheless, the increase in fraudulent or deceptive reviews aimed at influencing public opinion and obtaining unfair advantages has raised significant issues regarding integrity and reliability. In this article, we present a system designed to detect false reviews, employing natural language processing (NLP) techniques to tackle this problem. After tokenization, eliminating biases, and refining lemmatization, the system converts the text into a vector using TF IDF (term frequency- inverse document frequency). Systems utilizing artificial neural networks (ANNs) analyze these vectors to determine if a review is genuine or fraudulent. This application operates online, featuring a strong backend that employs MySQL for managing user data and presenting information, complemented by Bacund and TensorFlow Flask for its automated learning functions. Users can register for the application, access the platform, and receive live notifications related to classification. Training for the detection systems uses real-life instances, including Yelp reviews. Experimental results show that combining linguistic and behavioral elements improves the precision of identifying fraudulent reviews. This approach aids businesses in maintaining a strong reputation and boosting consumer confidence by automating the identification of manipulated content. Consequently, it elevates the overall quality and credibility of user generated content, providing e commerce platforms with an effective and adaptive solution to the challenges posed by reviews.

Keywords: cosine similarity, deep learning, model comparison, feature extraction, plant classification, Leaf identification, MobileNetV2, EfficientNetB0, ResNet50, Vision Transformer.

1. INTRODUCTION

   Automated plant identification systems are becoming more important because there is a growing need for accurate, scalable, and efficient tools in environmental monitoring and precision farming, biodiversity research, and plant disease management. Traditional methods, which primarily rely on expert observation and human classification of morphological features, are usually time consuming, subjective, and unsuitable for broad application.As computer vision As deep learning has progressed, there's increasing interest in the idea of automatically identifying plant species by analyzing images of their leaves. Because leaves offer a consistent and identifiable set of characteristics, such as shape, texture, venation, and color patterns, they are ideal for classification tasks.

   However, variability caused by intra-species variety, inter species similarity, and environmental factors significantly

   hinders the automatic recognition process. To overcome these obstacles, robust models that can learn and To perform well in image classification, CNNs have shown great effectiveness by processing a variety of complex visual data. Architectures like ResNet50, which is well-known for its residual learning techniques, are capable of efficiently capturing deep hierarchical features. By incorporating self attention mechanisms that capture global contextual information, new transformer based models like the VisionTransformer (ViT) have surfaced that provide an alternative to traditional convolutional processes. Additionally, lightweight models like MobileNetV2 and EfficientNetB0 provide an alluring trade off between accuracy and computational economy, particularly

   in resource constrained contexts.

   FIG 1:ARCHITECTURE DIAGRAM

   This study looks at how three different deep learning techniques perform when it comes to identifying leaves: (1) Vision Transformer (ViT), which recognizes long-range dependencies in leaf structure; (2) ResNet50, which learns robust residual features; and (3) a hybrid method that combines MobileNetV2 and EfficientNetB0 to achieve accurate and efficient classification. . Each model is evaluated on a custom dataset that includes five distinct categories: tree, herb, shrub, climber, and creeper. Cosine similarity is used to predict the final class. By comparing model accuracy, computational cost, This study focuses on practical use and aims to provide a useful framework for automatically identifying plants. It sets the stage for future research into scalable plant classification methods that can work in different environments and with

   various types of plant data.

2. LITERATURE SURVEY

   The area of automated plant identification has changed a lot, shifting from older image processing methods to more advanced deep learning approaches. In the beginning, systems mainly used basic features like shape descriptions, color information, and texture details. These features were often paired with traditional machine learning tools such as Support Vector Machines (SVMs) or k- Nearest Neighbors. Wu et al. achieved over 90% classification accuracy on the Flavia dataset by using geometric shape-based features and Principal Component Analysis (PCA) for dimensionality reduction, marking the first successful application of hand-crafted feature engineering. But these traditional methods often struggled to capture complex and abstract visual patterns, especially when parts of the image were hidden or lighting conditions changed. The arrival of Convolutional Neural Networks (CNNs) changed things a lot in the area of plant identification. CNNs can automatically learn different levels of features from raw image data, so there's no need for people to manually extract features. Because of its deep residual structure, which helps train deeper networks and solves the vanishing gradient issue, ResNet50 became especially popular. This approach has been widely used in many plant classification studies, showing how accurate and reliable it can be. In addition to CNNs, transformer-based architectures, like the Vision Transformer (ViT), have shown promise in a range of computer vision tasks. ViT differs from convolutional methods by using self-attention mechanisms, which let the model recognize long-range spatial relationships and overall patterns in an image. This is especially useful for plant species that show small changes across large leaf structures. Recent research has also focused on efficient and lightweight models that work well for use in embedded or mobile devices. While MobileNetV2 reduces computational load by using depthwise separable convolutions, EfficientNetB0 balances network depth, width, and resolution through a compound scaling approach. These models are suitable for edge-based and real- time plant identification systems., the hybridization of these models has demonstrated promise in maintaining classification accuracy while optimizing resource utilization. Despite these advancements, challenges remain. Intraclass variability, interclass similarity, ambient noise, and dataset limitations continue to hinder model generalization. As a result, many studies have used transfer learning, combined different types of data, or improved datasets. Some have looked into hybrid models that mix CNNs with traditional classifiers or ensemble methods to boost classification results. Overall, these studies show that while CNNs and ViTs are strong tools for plant image classification, there is still room for improvement in model design, dataset variety, and how efficiently they use resources. To find the best balance between classification accuracy and resource use for real-world use, this study builds on earlier work by comparing three advanced deep learning approaches.

3. PROPOSED METHOD

   The proposed system for classifying plant leaves is built on deep learning techniques that work well for image recognition. The method includes five main parts: classification, using advanced models to extract features, preparing the dataset, preprocessing images, and integrating the system architecture. The next sections explain each of these steps in detail.

   • Dataset Preparation:

     A custom dataset was made in order to gather a broad range of

     leaf photos from five distinct plant categories: trees, shrubs, climbers, creepers, and herbs. Each category contains several species, ensuring diversity in leaf form, color, texture, and structure. Various lighting conditions were used to take the photos.conditions and environments to increase model resilience and accurately depict real world situations.

   • Image Preprocessing:

     Preprocessing is needed to make sure the models get uniform and good quality input. Every image is resized to a set size of 224 x 224 pixels so it fits the requirements of deep learning models. The pixel values are adjusted to a standard range between 0 and 1 to help the model learn faster during training. Techniques such as random rotation, flipping, cropping, and changing brightness are also used to make the model work better on different types of data and avoid being too focused on the training data.

   • Feature Extraction Models:

   Three different deep learning approaches are used for extracting features and performing classification.

   • ResNet50

     ResNet50, a 50 layer deep CNN, can learn complex features thanks to its residual connections.layers without vanishing gradients. Its design is perfect for capturing the fine-grained features of leaves, such as edge contours and venation patterns.

   • VisionTransformer, or ViT

     Instead of employing convolutional processes, ViT uses a self attention mechanism to comprehend spatial relationships across the entire image. By dividing the input into patches, encoding them into embeddings, and processing them through transformer layers, it successfully captures global dependencies and leaf structure.

   • MobileNetV2 + EfficientNetB0 (Hybrid) This hybrid model combines the advantages of two lightweight architectures. While MobileNetV2 uses depthwise separable convolutions to simplify the model, EfficientNetB0 uses compound scaling to optimize accuracy and efficiency. Together, they offer a comprehensive solution for environments with limited processing capacity.

     • Classification Strategy:

       After extracting features, a cosine similarity measure is used to compare the features from a test image with those in the labeled dataset for classification. The class that has the highest similarity score is considered the correct label. This method is especially helpful when there are not many examples in each class and eliminates the need for a fully connected classification layer in the network.

     • System Architecture:

   The system follows a modular architecture:

   Input Module: Allows users to upload leaf images.

   Preprocessing Module: Standardizes and augments input images.

   Feature Extraction Module: Uses the chosen deep learning model to extract high dimensional features. Classification Module: Compares features using cosine similarity to assign a class label.

4. RESULTS

   Several quantitative measures were used to check how well the proposed deep learning model for leaf classification worked. A specially made dataset with five types of leavescreeper, shrub, herb, tree, and climberwas used to test the MobileNetV2 and EfficientNetB0 models. Vision Transformer (ViT), and ResNet50. The results validate the system's precision, robustness, and usefulness.

   PERFORMANCE ON IMAGE DATASETS

   Subgroups of the dataset were made for training (80%), validation (10%), and testing (10%). The same test group was used to check all three models to make sure the results were consistent. The ResNet50 models did better than the others in classification accuracy and were more consistent across different categories.

   Performance Metrics for ResNet50:

   Accuracy: 86.6%

   Precision: 87.2%

   Recall: 86.1%

   F1-Score: 86.6%

   AUC (Area Under Curve): 0.91

   These results show that ResNet50 is very good at pulling out useful features from different leaf shapes and textures.particularly in challenging categories such as shrubs and climbers. Its residual learning mechanism contributed to robust learning with minimal overfitting.

   PERFORMANCE OF OTHER MODELS The Vision

   Transformer (ViT) and MobileNetV2 EfficientNetB0 hybrid models were tested under the same conditions. Despite having a slightly lower accuracy than ResNet50, both models generated results that were competitive. Vision Transformer (ViT):

   Accuracy: 80.0%

   Precision: 80.5%

   Recall: 80.0%

   F1-Score: 80.2%

   AUC: 0.88 %

   Despite being slightly more vulnerable to variations in lighting and background clutter, ViT performedadmirably in gathering global patterns. It performs exceptionally well in settings with clear visuals.

   MobileNetV2 + EfficientNetB0:

   Accuracy: 80.0%

   Precision: 79.1%

   Recall: 80.3%

   F1-Score: 79.7%

   AUC: 0.87%

   This hybrid model did very well when it came to using less computation and being fast at making predictions, which makes it great for use in real time, even though it had a little lower accuracy compared to ResNet50.

   ANALYTICAL PERSPECTIVES

   A comparative evaluation offers several key insights: ResNet50 is most effective for high accuracy applications where resource availability is not a constraint. It performed consistently across all five plant categories and handled visual variability well. ViT showed potential in simulating how species with complex leaf structures depend on each other globally, but better input conditions might be needed to get the best results..

   The MobileNetV2 EfficientNetB0 hybrid achieved competitive accuracy with significantly reduced model size and inference time, making it suitable for mobile and embedded environments. When combined, these findings show how well the suggested deep learning classification system works in identifying different plant species by analyzing images of their leaves. Each model's unique capabilities allow for selection based on use-case requirements, such as speed, hardware limitations, or maximum accuracy. Because it achieved competitive accuracy with a significant reduction in model size and inference time, it is suitable for embedded and mobile environments.

   Fig 2: Accuracy Comparison

   Fig 3:Classification

   Fig 4: Bar Chart for Model Accuracy Comparison

5. DISCUSSION

   The results of the experiment indicate that all three deep learning models tested in this study can effectively classify leaves. The fact that ResNet50 performed better than the other models suggests that its residual connections are helpful in identifying detailed features that are important for telling similar leaf structures apart.The Vision Transformer showed promise in recognizing global patterns and spatial relationships, despite its effectiveness being somewhat limited by its sensitivity to visual fluctuations such as illumination and background noise. Even though it can handle more comple datasets, it may require additional fine tuning or larger datasets in order to achieve CNN consistency. A well rounded alternative was offered by the hybrid MobileNetV2 EfficientNetB0. Although it was quicker and consumed fewer resources than ResNet50, its accuracy was not significantly higher. As a result, it is an excellent choice for real time applications or the deployment of mobile devices, where computational efficiency is crucial.The study's overall findings indicate that accuracy and efficiency must be compromised. While deeper models provide more accuracy, lighter models are more useful in constrained environments. The requirements of the application, such as high accuracy in controlled settings or real-time performance in the field, should thus guide the model selection.

6. CONCLUSION

This study looked at five different types of plantstrees, shrubs, climbers, creepers, and herbsand how deep learning models can be used to automatically classify plant leaves. A special image dataset was used to test three models: ResNet50, Vision Transformer (ViT), and a mix of MobileNetV2 and EfficientNetB0. These models were carefully checked based on how accurate they were, how efficient

they were, and how useful they were in real-world situations. ResNet50 came out as the most accurate, with an accuracy of 86. 6% on the test data. which help improve performance It could effectively retrieve detailed information from various types of leaf structures.The Vision Transformer demonstrated promising results with an accuracy of 80%, particularly when global spatial patterns were important. Achieving 80% accuracy while significantly reducing computing overhead, the MobileNetV2 EfficientNetB0 hybrid model provided a feasible method for deployment on lightweight or mobile platforms. All things considered, the results demonstrate that deep learning models can provide reliable and scalable solutions for tasks involving plant identification. The hybrid model is used for efficiency, ResNet50 is used for high precision classification, and ViT is used for contextual image understanding. Using cosine similarity as a classification method helped the models better match feature vectors, particularly when there wasn't much labeled data available. pp. 11525. DOI. org (Crossref), This research sets the foundation for creating intelligent systems in areas such as agriculture and biodiversity conservation. and ecological monitoring. Future plans call for adding more species to the dataset, improving class balance, and integrating real time classification features for mobile and web based platforms. Multi-modal features, like environmental data or plant metadata, may also increase the model's resilience and level of classification.

VI.REFERENCES

1. Labed, Said, et al. Plant Recognition Using Data Augmentation and Convolutional Neural Network. Modelling and Implementation of Complex Systems, edited by Salim Chikhi et al., vol. 593, Springer International Publishing, 2023, pp. 192204. DOI.org (Crossref), https://doi.org/10.1007/978-3-031-18516- 8_14

2. Hangkaya, Febrian Lourenso, et al. Effects of Solvent to Feed Ratio and Microwave Power on Extraction of Essential Oil from Red Guava Leaves (Psidium Guajava l.) Using Microwave Hydro Distillation Method. International Journal of Eco-Innovation in Science and Engineering, vol. 1, no. 01, May 2020, pp. 4247. DOI.org (Crossref), https://doi.org/10.33005/ijeise.v1i01.15

3. Ye, Dapeng, et al. A Synthetic Review of Various Dimensions of Non- Destructive Plant Stress Phenotyping. Plants, vol. 12, no. 8, Apr. 2023, p. 1698.

   DOI.org(Crossref)https://doi.org/10.3390/plants12081698

4. Wong, W. K., et al. Intrusion Detection System Model: A White-Box Decision Tree with Feature Selection Optimization. Neural Computing and Applications, vol. 37, no. 7, Mar. 2025,

   pp. 565570. DOI.org (Crossref), https://doi.org/10.1007/s00521-024- 10942-4.

5. Mahmudul Hassan, S. K., and Arnab Kumar Maji. Identification of Plant Species Using Deep Learning. Proceedings of International Conference on Frontiers on Computing and Systems, edited by Debotosh Bhattacharjee et al., vol. 1255, Springer Singapore, 2021, pp.11525. DOI.org (Crossref), https://doi.org/10.1007/978-981-15- 7834-2_11.

6. Ouamane, Abdelmalik, et al. Enhancing Plant Disease Detection: A Novel CNN-Based Approach with Tensor Subspace Learning and HOWSVD-MDA. Neural Computing and Applications, vol. 36, no. 36, Dec. 2024, pp. 2295781. DOI.org (Crossref), https://doi.org/10.1007/s00521-024- 10454-1.