🍊 Introducing Citrus-Net

An end-to-end image classification pipeline that combines 🧱 Custom CNN, 🔁 Transfer Learning, and 🔭 Vision Transformer models - enhanced with 🧩 Explainable AI (XAI) visualizations and deployed in a Streamlit app.

---

---

📌 Repository Description

This repository contains the complete project for building and explaining a full image classification pipeline. The pipeline moves from a hand-crafted CNN baseline to state-of-the-art deep learning models, enriched with Explainable AI (XAI) methods, and deployed as an interactive Streamlit web app.

🚀 Project Stages

• 🔎 Data Preparation – Exploratory analysis, class balance checks, preprocessing, and augmentations.
• 🧱 Custom CNN (CitrusNet) – Novel convolutional network designed and trained from scratch.
• 🔁 Transfer Learning – Fine-tuning four pretrained CNN backbones (ImageNet weights).
• 🔭 Vision Transformer (ViT) – Transformer-based classifier for advanced performance.
• 🧩 XAI (Explainability) – Applying Grad-CAM, Score-CAM, LIME, and SHAP to interpret predictions.
• 🌐 Streamlit App – User-friendly interface for image upload, prediction, and interactive XAI visualizations.

✅ Key Features

• 🧱 Custom Made CNN [CitrusNet] with ≥65% accuracy (baseline target).
• 🔁 Transfer learning with four unique CNN architectures.
• 🔭 Vision Transformer experiment.
• 📊 Comparative evaluation (accuracy, precision, recall, F1, confusion matrices).
• 🧩 XAI overlays on ≥10 test images with detailed interpretation.
• 🌐 Streamlit app for model selection, predictions, and side-by-side explanations.

Model	Training Type	Test Accuracy	Macro F1	Notes
EfficientNet-B0	Transfer Learning	Highest (reported best)	Highest	Selected for XAI & deployment
ResNet34	Transfer Learning	Slightly lower than EfficientNet	High	Very stable baseline
ConvNeXt-Tiny	Transfer Learning	Comparable to ResNet	High	Heavier, no clear gain
CitrusNet (Proposed)	From Scratch	~97%	~0.97	Excellent for custom CNN
DenseNet121	Transfer Learning	Lower than ResNet	Moderate	Less stable
ViT-B/16	Transformer	Lowest	Lowest	Data-limited underperformance

🛠️ Tech Stack

• Python (PyTorch, torchvision, scikit-learn, Captum, pytorch-grad-cam, LIME)
• Streamlit for deployment
• Matplotlib / Seaborn for plots
• Google Colab / Kaggle GPU for training

📂 Dataset

• I used the Citrus Fruit Dataset containing high-quality images of citrus fruits for training, validation, and testing.

📑 Deliverables (After Contact)

• 📄 Scientific Project Report (IEEE/ACM style).

• 📦 6 Runnable Jupiter Notebook Code.

• 💾 6 Trained Model Weights.

  • CitrusNet_[custom_cnn].pt
  • Efficientnet_B0.pt
  • Resnet34.pt
  • Densenet121.pt
  • Convnext_Tiny.pt
  • VIT_Best.pt

🚀 Quick Start

🌐 Use the Web App

1. Open the live application [No installation required]
   👉 Citrus Fruit Classification

2. From the left sidebar:

   • Select a model: (Best Model Used)
     • CitrusNet [Custom]
     • EfficientNet-B0 [Best]
     • ResNet34
     • or Upload your own .pt model
   • Upload one or more citrus fruit images (.jpg, .png)

3. View the results:

   • Predicted fruit class
   • Confidence score
   • Grad-CAM visual explanations
   • Optional LIME explanations

⚠️ Important Display Notice

• This app is styled for Dark mode. Light theme may hide some UI elements.
• Change via [Right Side ⋮ → Settings → App Theme → Dark].

💻 Run Locally (Optional)

git clone https://github.com/Shourav-Deb/Citrus-Net.git
cd Citrus-Net
pip install -r requirements.txt
streamlit run app.py

---

📧 Contact

If you need the trained notebook, models or report file for academic purposes, feel free to Contact Anytime.

⭐ Acknowledgements

• Built with PyTorch, torchvision, scikit-learn, pytorch-grad-cam, LIME, and Streamlit.
• Inspired by cutting-edge research in Computer Vision & Explainable AI (XAI).

---