🍔 Food & Beverage Classification using CNN

📌 Project Overview

This project classifies food & beverage images using a Convolutional Neural Network (CNN). The dataset consists of 9323 training images and 484 test images across 61 classes (e.g., water, pizza-margherita-baked, broccoli, salad, egg, etc.).

🔹 Objective

• Train a CNN model to classify food items from images.
• Improve generalization using data augmentation.
• Monitor training progress with validation curves.
• Provide model interpretability using Grad-CAM visualizations.

---

📂 Dataset Details

📍 Directories

data/
├── training_set_128/    # 9323 images (train + validation)
├── test_set_128/        # 484 images (unlabeled test data)
loss_accuracy_curves/    # images for some of the models tested
├── accuracy/
├── loss/
saved_models/            # best model trained
images/                  # for README

📍 Example Classes & Image Distribution

Class	Number of Images
Water	863
Bread-White	595
Salad-Leaf	535
Pickle	28

📍 Example Training Images

Below are some sample images from the training set (some of them are difficult to recognize even for human eyes - hard-cheese hard ineed!):

---

🔹 Data Preprocessing & Augmentation

To improve generalization, applied:

• Rotation (±30°)
• Zooming (20%)
• Shifting (20%)
• Horizontal Flipping
• Rescaling (0-255 → 0-1)

---

🏗️ Model Architecture

The CNN consists of:

• 4 Convolutional Blocks with ReLU activation
• Softmax Activation for multi-class classification

model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(128, 128, 3)),
    MaxPooling2D((2, 2)),

    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D((2, 2)),

    Conv2D(128, (3, 3), activation='relu'),
    MaxPooling2D((2, 2)),
    
    Conv2D(256, (3, 3), activation='relu'),
    MaxPooling2D((2, 2)),

    Flatten(),
    Dense(128, activation='relu'),
    Dense(61, activation='softmax')
])

---

📊 Model Training & Evaluation

• Optimizer: Adam (learning_rate=0.001)
• Loss Function: Categorical Crossentropy
• Metrics: Accuracy, Precision, Recall

🖥️ Training Curves

Loss & Accuracy over epochs:

🔍 Confusion Matrix

Visualizing misclassified images:

---

🎨 Model Interpretability

We use Grad-CAM to visualize important regions in an image that influenced predictions.

Example Grad-CAM Visualization:

---

📌 Future Improvements

• Use Transfer Learning (e.g., MobileNet, ResNet) for better accuracy
• Optimize hyperparameters using KerasTuner

---