🍇 Blackberry & Lime Visual Classifier

---

📌 Project Overview

We built an end-to-end image classification system to:

• 🥝 Identify whether the input fruit is a blackberry or a lime
• 🧺 Determine how the fruit is presented: e.g. Halved, In a Container, Whole, etc.

The app uses transfer learning via pre-trained ResNet-18, and is deployed as a live web app using Gradio on Hugging Face Spaces.

---

📂 Dataset

• ✅ Collected manually
• 🎯 2 produce categories:
  • blackberry
  • lime
• 🎯 6 variation categories:
  • Halved
  • In Context
  • In a Container
  • Single Berry
  • Small Group
  • Whole

Total images: ~6000
Split: 70% Train, 20% Validation, 10% Test

The original dataset has been published in kaggle, can be refered through the link.

Kaggle Link

---

🧠 Model

• Base Model: ResNet-18 (pretrained on ImageNet)
• Two-stage classifiers:
  • produce_classifier.pth – blackberry vs lime
  • variation_classifier.pth – 6-class variation
• Frozen all layers except the final classification head
• Optimizer: Adam
• Loss: CrossEntropyLoss
• Epochs: 5

---

📊 Performance

🍇 Produce Classifier

• Accuracy: 100%
• Precision, Recall, F1-score: 1.00 for both classes

🧺 Variation Classifier

• Accuracy: 99%
• Macro & Weighted F1-score: 0.99

---

🌐 Web App Demo

🧪 Try the live web app:

👉 Hugging Face Space:
https://huggingface.co/spaces/brs13/blackberry-lime-classifier

🔍 How to Use the App

1. Click the link above to launch the app.
2. Upload a fruit image (blackberry or lime).
3. The app will:
   • Predict the produce type (blackberry or lime).
   • Predict its visual presentation category (e.g., Whole, Halved).
4. You will also see example images and model outputs.

---

📄 How to View the Project Report (index.html)

This project includes a detailed and self-contained HTML report summarizing everything we did.

🧾 Steps to View:

1. Clone or download this GitHub repository to your computer.
2. Locate the file named index.html in the root directory.
3. Double-click the file to open it in any modern web browser (e.g., Chrome, Firefox, Safari).
4. The report contains:
   • ✅ Full project description
   • 📊 Confusion matrices and performance results
   • 🧠 Model details and training setup
   • 🚀 Summary, future work, and app links

---