☕ Coffee Bean Roast Level Classification

Using Custom CNN and MobileNetV2 (Comparative Study)

---

Project Overview

Coffee roast level grading is traditionally performed by human experts, which introduces subjectivity and inconsistency in large-scale production. This project presents an automated computer vision system for coffee bean roast level classification using deep learning.

We classify coffee beans into four roast levels:

• 🟢 Green
• ☕ Light
• ☕☕ Medium
• ☕☕☕ Dark

A comparative study is conducted between:

• A Custom Convolutional Neural Network (CNN)
• A Transfer Learning model using MobileNetV2

Both models are evaluated on unseen test data under realistic conditions.

---

Objectives

• Automate coffee bean roast classification using images
• Reduce subjectivity in roast grading
• Evaluate robustness under real-world noise and lighting variations
• Compare custom CNN vs MobileNetV2
• Demonstrate real-time classification with confidence visualization

---

📁 Dataset Structure

/train
 ├── /Dark
 ├── /Green
 ├── /Light
 └── /Medium

/test
 ├── /Dark
 ├── /Green
 ├── /Light
 └── /Medium

Each folder contains labeled coffee bean images corresponding to roast levels.

---

Preprocessing & Segmentation

• Image resizing to 150 × 150 × 3

• Normalization to [0,1]

• ROI segmentation using:

  • Grayscale thresholding
  • Morphological operations

• Feature focus on color intensity and texture

• Data augmentation to simulate real-world conditions:

  • Brightness variation
  • Noise
  • Blurring

---

Models Used

1️⃣ Custom CNN Architecture

Layer	Purpose
Conv2D (32)	Edge & color gradient extraction
Conv2D (64)	Texture & crack patterns
Conv2D (128)	Roast intensity & surface features
Dense + Dropout	Non-linear classification & regularization
Softmax	4-class probability output

Training Details:

• Optimizer: Adam
• Loss: Categorical Crossentropy
• Epochs: 10
• Input Shape: (150, 150, 3)

---

2️⃣ MobileNetV2 (Transfer Learning)

• Pre-trained on ImageNet
• Lightweight & efficient
• Global Average Pooling + Dense layers
• Fine-tuned for 4-class classification

Why MobileNetV2?

• Faster inference
• Fewer parameters
• Suitable for mobile & embedded deployment

---

Evaluation Results

Custom CNN Performance

Class	Precision	Recall	F1-Score
Dark	0.98	0.95	0.96
Green	1.00	1.00	1.00
Light	1.00	0.99	0.99
Medium	0.94	0.98	0.96
Accuracy			0.98

---

MobileNetV2 Performance

Class	Precision	Recall	F1-Score
Dark	1.00	0.99	0.99
Green	0.99	1.00	1.00
Light	1.00	0.93	0.96
Medium	0.93	1.00	0.97
Accuracy			0.98

---

Comparative Analysis

Aspect	Custom CNN	MobileNetV2
Accuracy	98%	98%
Model Size	Larger	Lightweight
Training Time	Higher	Faster
Feature Learning	Task-specific	Pre-trained
Deployment	Server	Mobile / Edge

➡ Conclusion: Both models perform equally well, but MobileNetV2 is preferred for deployment, while Custom CNN provides better architectural interpretability.

---

Interactive Demo

A real-time demo:

• Randomly selects a test image
• Predicts roast level with confidence
• Displays real-world usage of the roast type

Implemented using Matplotlib + ipywidgets

---

Technologies Used

• Python
• TensorFlow / Keras
• OpenCV
• NumPy, Pandas
• Matplotlib, Seaborn
• Google Colab

---

Key Contributions

• Light-invariant coffee roast classification
• ROI-based segmentation
• Robust evaluation under noise
• Comparative deep learning study
• Deployment-ready architecture

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

---

Acknowledgments

• Open-source deep learning community
• TensorFlow & Keras documentation
• Academic references on computer vision and CNNs

---

Contact

For academic discussion or collaboration:

VIVEKA S Big Data Analytics Core Project

---