Building a Food Image Classification Web App

Business Scenario

This project demonstrates a complete machine learning workflow, from data preparation and model training to deployment and user interaction, making it a practical solution for users to upload food images and receive predictions about the type of food depicted. The app uses training data on the Food-101 Dataset, which contains 101,000 images of 101 food categories.

Techniques

• Data Preparation:
  • Pandas: For handling tabular data and creating dataframes.
  • NumPy: For numerical operations and preprocessing tasks.
  • scikit-learn: LabelEncoder for encoding class labels into numerical format, train_test_split for splitting the dataset into training and validation sets.
  • Pillow (PIL): For loading and processing image files.
  • torchvision.transforms: For applying transformations to images, including resizing, normalization, and converting to tensors.
• Model Building: PyTorch, a popular deep learning framework, to build and train the image classification model. The model is based on the ResNet18 architecture, leveraging PyTorch’s pre-trained models and transfer learning capabilities for efficient training and accurate predictions.
• Model Deployment: The app is deployed using Streamlit, a Python-based framework for building interactive web applications.

Workflow

1. Data Preparation

• Download and extract the Food-101 dataset.
• Images are resized to 128x128 pixels.
• Normalization is applied to standardize pixel values.
• Labels are encoded using LabelEncoder.

2. Model Training

• Split the dataset into training and testing sets using predefined metadata files (train.txt and test.txt).
• Load the ResNet18 model and modify the fully connected layer to match the number of food categories.
• Train the model using the training dataset.
• Early stopping to prevent overfitting.
• Validation split during training for performance monitoring.
• Adam optimizer for gradient-based optimization.

3. Model Deployment

• Load the trained model and label encoder in the Streamlit app.
• Predict the food category for uploaded images using the trained model.
• Define a preprocessing pipeline for uploaded images.

Functionalities

• Sidebar with dataset and model information

• Display predictions in a user-friendly interface. Users can upload food images in .jpg, .png, or .jpeg formats. The app predicts the food category using the trained model

Demo