---

RiceVision Lab is a curated workspace exploring automated rice variety classification using modern computer-vision techniques. This project was created by me and my team, which analyzes two well-known rice datasets and compares six powerful CNN architectures to understand how models behave on both single-grain and multi-grain images.

---

🎯 Project Goals

• Explore how different CNN architectures perform on single-grain vs multi-grain images
• Understand the impact of augmentation on model generalization
• Visualize attention patterns using Grad-CAM
• Provide a clean and reusable workflow for other ML learners

📁 Datasets Used

1. Aruzz Dataset (20 varieties - single grain images)

A large-scale rice dataset featuring high-quality single-grain images.
👉 Mendeley - Aruzz

2. BD Rice Dataset (8 categories - multiple grains per image)

Images captured from mixed environments with visible clusters of rice grains.
👉 Mendeley - BD Rice

Both datasets were used in original and augmented forms for performance comparison.

🧠 Models Used (6 CNN Architectures)

This project evaluates six modern and diverse architectures:

• ResNeXt50
• DenseNet201
• GhostNet
• EfficientNetV2-S
• NASNet-A-Large
• Xception

Each model was trained using a unified PyTorch pipeline with identical hyperparameters for a fair comparison.

📓 Repository Contents

This repository is intentionally kept clean and lightweight. To help learners focus on the core workflow understanding, only demonstration is included:

✔ Website Showcase

• Open the 🌾Rice Vision web application in your browser.

• Select the Grain Type from the sidebar:

  • Single Grain (Aruzz) → for single rice grain images
  • Multi Grain (BDRice) → for multi-grain images

• Choose a model architecture:

  • EfficientNet-V2-S
  • DenseNet-201
  • ResNeXt50-32×4d

• Upload one or more rice images using the Upload image(s) option.

  • Supported formats: .jpg, .png, .jpeg

• For each uploaded image, the application shows:

  • Image name and preview
  • Predicted rice variety
  • Confidence score
  • Top 3 predictions displayed as confidence bars

• A low-confidence warning is displayed if the prediction confidence is below the threshold.

• Enable Grad-CAM (optional) to visualize image regions influencing the prediction.

• Multiple images are processed sequentially for batch evaluation.

✔ Visualizations Included

The repository includes all essential visual outputs generated during the experiments, providing a clear understanding of model behavior:

• Training & Validation Accuracy/Loss Curves
• Confusion Matrices
• ROC-AUC & Precision–Recall Curves
• Grad-CAM Heatmaps (Aruzz & BDRice)
• Both Dataset Sample Image's

These visuals highlight learning patterns, model stability, and the specific grain features each model focuses on - making the entire workflow transparent and easy to interpret.

---

⚠️ 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].

---

🏆 Highlight Findings (From the Experiments)

• EfficientNetV2-S achieved the highest performance across both datasets
• Reached 99.95% accuracy after augmentation
• DenseNet201 and NASNet-A-Large also performed extremely well
• Grad-CAM confirmed meaningful feature focus (grain edges, texture patterns)
• Augmentation significantly improved generalization, especially for BD Rice

Dataset ↓ / Model →	ResNeXt50	DenseNet201	EfficientNet-V2-S	NASNet-Large	GhostNet	Xception
Aruzz Original	🟡 Good	🟡 Good	🟢 Best	🟠 Average	🔴 Weak	🟡 Good
Aruzz Augmented	🟢 Best	🟢 Best	🟢 Best	🟡 Good	🟠 Average	🟢 Best
BDRice Original	🟡 Good	🟡 Good	🟢 Best	🟠 Average	🔴 Weak	🟡 Good
BDRice Augmented	🟢 Best	🟢 Best	🟢 Best	🟡 Good	🟠 Average	🟢 Best

---

📦 Other Resources

To ensure the repository remains lightweight and easy to demonstrate, training notebooks, trained model checkpoints (.pth), and the full PDF report are intentionally omitted.

📬 You can request them at: contact me

---