🌾 HuskNet: Deep Learning Based Multi-class Classification of Husk Species in Bangladesh

This repository contains the official code and resources for the research paper,
"HuskNet: A Deep Learning Model for Classifying Husk Species."

---

📖 Project Overview

In an agriculture-dependent country like Bangladesh, cattle play a vital role in farming and meeting nutritional demands.
Husks, the protective coverings of plants, are a key and easily available component of cattle feed.
Understanding the nutritional content of different husk species is crucial for formulating balanced and cost-effective diets for livestock.

This project introduces HuskNet, a deep learning model designed to accurately classify eight common husk species found in Bangladesh.
By leveraging computer vision, HuskNet aims to provide a data-driven approach to enhance cattle feed formulation, ultimately supporting more sustainable agricultural practices and improving livestock productivity.

---

🌾 The BDHusk Dataset

The model was trained and evaluated on the BDHusk dataset, which includes:

• Content: A comprehensive collection of images from eight different husk species commonly found in Bangladesh.
• Classes (8 total):
  • Rice Husk
  • Corn Husk
  • Wheat Husk
  • Chickpea Husk
  • Lentil Husk
  • Soybean Husk
  • Grass Pea Husk
  • Field Pea Husk
• Size: 2400 total images (300 per class).
• Origin: Images were captured in the Sirajganj district of Bangladesh using mobile camera devices.

---

🧠 Model Architecture: HuskNet

HuskNet is a deep learning model specifically tailored for this multi-class classification task.

• Base Architecture: Built on the powerful ResNet50 architecture, utilizing transfer learning.
• Modifications:
  • Output layer with 8 neurons (one for each husk class).
  • Softmax activation function for multi-class predictions.

---

📊 Performance & Results

HuskNet demonstrated exceptional performance across all evaluation metrics:

• Average Accuracy:
  • Achieved 97% average accuracy across all eight classes.
• ROC Curve Analysis:
  • Achieved perfect micro-average AUC score of 100%.
• Class-Specific Metrics:
  • High precision, recall, and F1-scores for each class.
  • F1-scores reached up to 99% for Corn, Lentil, and Field Pea husks.

---

💻 Tech Stack & Environment

• Software & Libraries:

  • Python
  • TensorFlow
  • Keras
  • Scikit-learn
  • NumPy
  • Seaborn

• Hardware:

  • Intel Core i7 CPU
  • 32GB RAM
  • Nvidia RTX 3060Ti GPU