Department of Emerging Technologies BTech CSE(DS) IV Year I Semester

Technical Seminar, July 2024

BY: PRC Members

Akati Satvika Reddy Dr. Y. Alekhya Rani
21B81A6737 Dr. B. Ramakrishna
Dr. Yasmeen
Mrs. A. Srichandana
The Analysis of Plants Image
Recognition Based on Deep
Learning and Artificial Neural
Network
ABSTRACT
This paper focuses on using image analysis and deep learning
to classify and identify plant species based on their leaves.
Leaf images are segmented and processed to extract shape and
texture features. These features are then used to classify the
leaves using different algorithms, including KNN, Kohonen
networks, and SVM. Tests on a database of 50 plant species
and a comparison among seven plant types revealed that
ginkgo leaves were the easiest to identify. The proposed
method achieved good recognition accuracy and efficiency,
especially for complex backgrounds, with the deep learning
model showing the highest correct recognition rate and the
shortest recognition time.
 INTRODUCTION
Problem Domain Knowledge: Plants are essential
ecological resources, crucial for food production, air
purification, and environmental regulation.
Relevant Contemporary Issues: Traditional plant
classification methods are subjective, labor-intensive,
and require extensive taxonomic knowledge.
Problem Definition: Accurate identification of plant
species using leaf images is challenging due to the
diversity and complexity of leaf shapes and textures
and is done by ANNs.
Task Identification: Utilize deep learning techniques
to extract and analyze leaf shape and texture features
for efficient and accurate plant species classification.
 PROPOSED METHOD
A.IMAGE RECOGNITION
Image recognition in your proposed method can be broken down into three main
steps: image processing, image feature extraction, and image classification. Each
of these steps involves multiple techniques and methods, which work together to
accurately recognize the target features in the image.
1. Image Processing
•Preprocessing
• Image Restoration: Remove noise and enhance useful information.
• Re-encoding and Compression: Optimize images for real-time
processing.
•Segmentation Image pre-processing stage
• Edge-based Segmentation: Detect sudden changes in pixel values using
operators like Roberts and Prewitt.
• Threshold-based Segmentation: Segment based on differences in gray
levels.
• Region-based Segmentation: Group similar pixels or split and merge
regions.
2. Image Feature Extraction
•Global Features
• Color Features: Represent color distribution using color histograms.
• Texture Features: Describe surface patterns with spatial autocorrelation.
• Shape Features: Identify object contours and regions using methods like
the Generalized Hough Transform.
•Local Features
• Corner Detection: Identify key points where edges intersect (e.g., Harris
corner detection).
• Spot Detection: Identify regions differing from their surroundings.
3. Image Classification
•K-Nearest Neighbor (KNN): Classify based on the most common label among
nearest neighbors.
•Support Vector Machine (SVM): Separate classes with a hyperplane
maximizing margin.
•Decision Tree: Use a tree structure to make decisions based on feature values.
•Artificial Neural Network (ANN): Simulate brain-like networks to recognize
patterns.
B. Artificial Neural Network (ANN)
1. Structure of ANN
•Components:
• Neurons: Basic units processing input
signals.
• Connections: Weighted links between
neurons.
• Activation Function: Non-linear function
(e.g., sigmoid, tanh).

•Types of Networks:
• Feedforward Network: Layers of
neurons with unidirectional connections.
• Feedback Network: Neurons receive
both external and feedback inputs.
• Hybrid Network: Combines features of
both feedforward and feedback networks.
2. Neuron Model
•Input Signals (x): Signals from other neurons.
•Weights (w): Connection strengths.
•Bias (b): Threshold value.
•Output (y): Result after applying activation function to the sum of
inputs and biases.
3. Learning in ANN
•Supervised Learning: Uses labeled data to adjust weights.
•Unsupervised Learning: Identifies patterns in data without labels.
•Delta Rule: Minimizes the error between actual and target outputs
using gradient descent.
• Error Signal (e):

• Objective Function (J): Measures mean square error.

4. Common Activation Functions

•Sigmoid Function:

•Tanh Function: FIGURE : Sigmoid Function

C. Commonly Used Neural Network Models for Image Recognition
1. Hopfield Network
•Structure: Single-layer feedback neural network.
•Neurons: Each neuron receives inputs, processes them, and outputs a signal.
•Equation:

•Excitation Function:

2. Self Organizing Map (SOM)

It mainly includes 5 steps:
 Initialization: Initialize the weight vectors for all neurons randomly.
 Input Selection: Choose an input vector from the dataset.
 Best Matching Unit (BMU): Identify the neuron whose weight vector is
closest to the input vector.
 Weight Update: Update the weight vectors of the BMU and its neighbors to
move closer to the input vector.
 Repeat: Repeat steps 2-4 for a specified number of iterations or until
convergence.
It also mainly follows 3 processes: competition, cooperation and synaptic
adaptation
ARCHITECTURE DIAGRAM
Classification
KNN output
SVM
Decision Tree
Neural network

Basic Block Diagram of CNN

 PLANT LEAF RECOGNITION BASED ON DEEP LEARNING

A. Factors Affecting Plant Recognition in Deep Learning

•Image Recognition Technology: A vital field in •Parameter Selection: Crucial for training models,
artificial intelligence, relying on image features to involving choices like learning rates and iteration
perform tasks such as perception, planning, positioning, numbers. Multi-fold cross-validation is often used to
and control in robotics. It is particularly effective in optimize these parameters for better model performance.
complex environments. •Data Collection: A large amount of plant image data is
•Leaf Image Segmentation: Involves classifying pixels necessary for both training and testing stages. The data
using machine vision and image processing to separate includes original images and their labels, which helps in
plant leaves from backgrounds. Challenges include non- training deep learning models to learn complex
uniform dimensions, redundant information, and the functions and improve classification accuracy.
need for dimensionality and standardization. •Segmentation and Classification: Deep learning
•Feature Extraction and Preprocessing: Essential for models can extract high-dimensional features from
improving image recognition, focusing on removing images and use these to identify and classify plant
noise and enhancing important features for subsequent leaves. The process includes convolution, pooling, and
processing. This involves multiple layers in deep non-linear mapping to create feature vectors for
learning models to extract both global features and classification.
contextual information.
 B. Hierarchical Convolutional Deep Learning
1. Network Initialization and Training: The network parameters are
initialized using pre-trained networks, and then adjusted with
calibrated plant images through backpropagation until the training is
complete.
2. Output Layer: Utilizes a fully connected layer to fuse local features
into global ones. The loss function calculates prediction errors and
propagates them back for optimization.
3. Image Processing: Converts images into grayscale to simplify
processing while maintaining classification accuracy. The goal is to
reduce complexity and improve robustness and generalization.
4. Segmentation Challenges: While effective in controlled
environments, segmenting leaves in complex backgrounds remains
challenging due to background noise and varying environmental
conditions.
5. Deep Learning Advantages: These models can generalize from large
datasets, allowing for better feature extraction and classification. The
use of convolutional layers enables the detection of complex patterns
and features in images, making the models highly adaptable and
accurate.
6. Error Propagation: During training, errors are propagated back
through the network to refine and improve feature extraction and
classification accuracy. Techniques like cross-validation help in fine-
tuning parameters to prevent overfitting and ensure robust model
performance.
 EXPERIMENTS
Hardware and Software Environment:
•Hardware: Intel Core CPU at 3.40 GHz, 16 GB RAM.
•Operating System: 64-bit Windows 7.
•Software: MATLAB R2014b.

Data Enhancement:
•Methods Used: Translation and rotation transformations to improve
training data quality and reduce overfitting.
•Result: Enhanced blade images are shown in Figure 12.

Leaf Collection and Sample Preparation:

•Source: Local botanical garden.
•Total Leaves: Nearly 50.
•Samples Per Leaf: 200, with 180 for training and 20 for testing.
Image Segmentation:
•Issue: Some objects exceed image boundaries, making initial
curve placement difficult.
•Optimization: Wavelet transform and filter optimization
improved the topological structure, reducing nodes and
enhancing monitoring area.
•Performance Comparison: Shown in Figure 11 and Table 1,
demonstrating improvements before and after optimization.

Experiments Conducted:
1.Feature Validity: Compared different features for leaf
recognition.
2.Algorithm Performance: Compared recognition time and
rate across different algorithms.
3.BP Algorithm: Used to extract and compare seven different
plant leaf samples.
 DISCUSSION
A. Recognition Ratio of Different Features
•Findings: Plant leaf texture features have a lower recognition rate
compared to other features.
•Effective Approach: Combining shape features and texture features
enhances plant leaf recognition accuracy.
B. Comparison of Classification Methods
•Methods Tested: BP Neural Network, KNN, Kohonen Network, and
SVM.
•Results:
• BP Neural Network: Longest training time but highest
recognition rate and shortest recognition time.
• KNN: No training required; higher recognition time and lower
accuracy compared to BP Neural Network.
• SVM: Moderate training time, longest recognition time, and lower
recognition rate.
• Kohonen Network: Short training time but higher error rate and
lower effectiveness as a classifier.
C. Artificial Neural Network (BP Algorithm)
•Performance: High recognition rate (92.48%) for various plant leaves.
•Advantages:
• Wide applicability with minimal experimental setup.
• No need for manual feature setting or background stripping.
• High automation and strong model availability.
LITERATURE REVIEW OR RELATED WORKS
 T. Mclellan's study, "Geographic variation and plasticity of leaf shape and size in Begonia dregei and
B. homonyma (Begoniaceae)," explores the variations in leaf morphology and its adaptability to
different geographic regions. The research highlights the importance of understanding leaf shape and
size variations for taxonomic purposes and environmental adaptability.
 The work by J. S. Cope et al., "Plant texture classification using Gabor co-occurrences," demonstrates
the use of Gabor filters for classifying plant textures. The authors showed that combining Gabor filter
responses with co-occurrence matrices can significantly improve the accuracy of texture classification
in plants.
 E. Aptoula and B. Yanikoglu, in their paper "Morphological features for leaf based plant recognition,"
proposed using morphological features to recognize plant species based on leaf characteristics. Their
method focused on extracting contour and vein features, demonstrating high accuracy in plant
identification.
 The accuracy of leaf identification using deep learning is improved in the study by X. Wang et al.,
"Deep learning for plant identification in natural environment," where a CNN model is employed to
classify plant species based on images taken in natural settings. The model showed high accuracy and
robustness against varying environmental conditions.
 Y. LeCun et al. in "Gradient-based learning applied to document recognition," describe the
foundational principles of convolutional neural networks (CNNs), which have been adapted for
various image recognition tasks, including plant identification. The paper lays the groundwork for
using CNNs in complex image analysis.
 CONCLUSION
This study explores plant leaf recognition through digital image
analysis, demonstrating that Backpropagation Neural Networks (BP)
significantly improve accuracy. The BP method effectively identifies
plant leaves, although challenges remain, such as the need for
extensive training data, handling complex backgrounds, and
establishing standardized evaluation criteria. Future work should focus
on expanding leaf databases, refining recognition algorithms, and
developing unified standards for evaluation and plant species
identification. Despite its promise, the BP method requires ongoing
refinement and standardization to advance plant recognition
technology.
REFERENCES
THANK YOU!!