Multi-class Classification of Insects: A Focus on the Asian Hornet

Project Overview

What is this Project About?

• Focus on Species Classification: Leveraging advanced deep learning techniques for accurate species identification.
• Relevance to Conservation Efforts: Highlights the ecological importance of species identification.
• Efforts to Collect Real-Time Images: Synthesizing and gathering diverse datasets for real-world applicability.
• Broader Scope with 8 Classes: Includes various insect species for comprehensive analysis.

Why This Project Held Interest for Me?

• Fine-Grained Differences in Species: Addressing challenges in identifying visually similar species.
• Scarcity of Comprehensive Studies: Bridging gaps in available research and data.
• Dataset Challenges and Innovation: Employing innovative approaches to create a balanced dataset.
• Application of State-of-the-Art Techniques: Incorporating cutting-edge deep learning methods.

---

Dataset Creation and Preparation

Sources of the Dataset

• Roboflow (Website)
• Video Frames
• Web Scraping
• Image Augmentation Techniques: Improving dataset diversity and robustness.

Dataset Balancing

• Key Benefits of Dataset Strategy:
  • Diversity
  • Class Balance
  • Generalisation

Selection Criteria for the Species

• Abundance and Impact in the UK: Focus on species like Asian Hornet, European Hornet, and Oriental Hornet.
• Similar Appearance and Potential for Misclassification: Addressing challenges with species like Hoverfly, Honey Bee, Carpenter Bee, Common Wasp, and Asian Giant Hornet.
• Importance for Ecological and Public Awareness: Enhancing awareness and research on insect biodiversity.

---

Selected Species

• Asian Hornet
• Asian Giant Hornet
• European Hornet
• Carpenter Bee
• Oriental Hornet
• Hoverfly
• Honey Bee
• Common Wasp

---

Objectives Achieved

• Dataset Development and Curation
• Implementation of Baseline Models
• Fine-tuning and Evaluation of Pretrained Models
• Analysis and Reporting

---

Deep Learning Models for Insect Classification

Custom CNN Models

• Development of Five Custom CNN Versions
• Hyperparameter Tuning:
  • Learning Rate
  • Batch Size
  • L2 Regularisation
  • Convolutional Layers
  • Augmentation Variations

Transfer Learning Models (Pretrained Models Used)

• VGG16
• MobileNetV2
• EfficientNetB0
• DenseNet121
• InceptionV3
• ResNet50

Parameter Tuning for Pretrained Models

Augmentation Parameters	First Version	Second Version
Rescale	1.0/255	1.0/255
Zoom Range	0.3	0.2
Width Shift Range	0.3	0.2
Height Shift Range	0.3	0.2
Rotation Range	30 degrees	20 degrees
Brightness Range	[0.7, 1.3]	[0.8, 1.2]
Shear Range	0.2	-
Horizontal Flip	TRUE	TRUE
Fill Mode	'nearest'	'nearest'

---

Impact of Augmentation Variations on CNN Performance

Metric	First Augmentation Variation	Second Augmentation Variation	Improvement
Training Accuracy	61.81%	73.97%	12.16%
Validation Accuracy	72.17%	80.27%	8.10%
Test Accuracy	74.15%	80.43%	6.28%

---

Improving Accuracy with Custom CNN Parameters

Parameters	Version 2	Version 3	Version 4	Version 5
Batch size	32	32	32	64
L2 Regularization	0.00005	0.00005	0.001	0.001
Learning Rate	0.001	0.0001	0.001	0.001
Augmentation Strategy	Second Version	Second Version	Second Version	Second Version
Training Accuracy	73.97%	70.44%	86.01%	85.83%
Validation Accuracy	80.27%	74.38%	90.70%	83.05%
Test Accuracy	80.43%	76.04%	86.32%	85.07%

---

Stratified K-Fold Validation Analysis on Custom CNN

Metric	Value
K-Fold	5
Fold Accuracies	85.44%, 87.20%, 87.57%, 85.19%, 85.06%
Mean Accuracy	86.09%
Standard Deviation	1.07
Variance	1.14
Coefficient of Variation (CV)	0.0124

---

Accuracies of Pre-trained Models with Optimum Parameters

Model	Convergence (Epochs)	Train Accuracy	Validation Accuracy	Test Accuracy
VGG16	78	94.44%	91.78%	88.08%
MobileNetV2	28	98.81%	96.14%	96.61%
EfficientNetB0	23	98.28%	97.79%	97.19%
DenseNet121	37	98.77%	98.29%	97.81%
InceptionV3	25	98.95%	98.92%	98.73%
ResNet50	24	98.82%	98.04%	97.42%

---

Reflections

• Learned complexities of model selection, training, and evaluation.
• Emphasized understanding performance factors over relying on advanced models.
• Highlighted critical role of data augmentation and hyperparameter tuning.
• Identified minimal impact of certain techniques like shear augmentation.
• Addressed practical challenges: computational limits and data constraints.

---

Simple API for Model Predictions

Overview

• Created a Python Flask-based API to provide predictions for insect classification using both custom CNN and pretrained models.
• Hosted in the Api_Dissertation folder.

Folder Structure

• DenseNet121.keras, InceptionV3.keras, ResNet50.keras, MobileNetV2.keras, EfficientNetB0.keras, Moderate_augmentation.keras: Pretrained and custom CNN models.
• templates/ folder containing the index.html file for the API interface.
• App2.py: The main Flask application for running the API.

Features

1. Pretrained Models:

   • DenseNet121
   • InceptionV3
   • ResNet50
   • MobileNetV2
   • EfficientNetB0
   • Custom CNN Model

2. Key Functionality:

   • Accepts uploaded image files.
   • Provides prediction for selected model.
   • Accessible through a web interface or API endpoint.

3. Usage Instructions:

   • Run App2.py to start the API.
   • Access the web interface at http://127.0.0.1:5000/.
   • Upload an image and select a model for predictions.

Example Prediction Flow

• Upload an Image: Users upload insect images via the web interface.
• Select a Model: Choose from DenseNet, Inception, ResNet, MobileNet, EfficientNet, or Custom CNN.
• Receive Predictions: The predicted class of the insect is displayed.

---

Conclusion

This project combined advanced deep learning techniques, robust dataset preparation, and an intuitive API to address the challenges in multi-class insect classification, with a specific focus on the Asian hornet. The implementation provides significant contributions to ecological research and awareness while showcasing the potential of machine learning in biodiversity conservation.