Classification of Noisy MNIST Images Using Deep Learning

This project aims to classify noisy images of handwritten digits from the MNIST dataset using various deep learning models. The project explores different architectures, including Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN), to determine the most accurate model for this task.

Project Structure

The project is implemented in a Jupyter notebook:

• Projet_MNIST.ipynb - Main notebook containing the data preprocessing, model development, training, and evaluation steps.

Objective

The primary objective of this project is to accurately classify noisy MNIST images using deep learning models. The noise added to the images simulates a more challenging real-world environment where input data might be noisy or distorted.

Approach

We experimented with two main types of neural network architectures:

• MLP (Multi-Layer Perceptron): A fully connected network that takes the flattened pixel data as input.
• CNN (Convolutional Neural Network): A convolutional architecture that processes the spatial structure of the image for feature extraction.

Key Models:

1. MLP (Fully Connected Networks): Various architectures with hidden layers and dropout for regularization.
2. CNN (Convolutional Neural Networks): Different CNN architectures with convolutional layers followed by max-pooling, flattening, and dense layers.

Data Preprocessing

• The MNIST dataset is used, with additional noise added to simulate real-world scenarios.
• Normalization: The pixel values are normalized to a range of [0, 1] for faster convergence during training.
• One-hot encoding: The labels are one-hot encoded for categorical classification.

Model Performance

After testing multiple architectures, the CNN model showed the best performance on both the training set and during cross-validation.

• Best Model: CNN with the following performance:
  • Accuracy: 96.46%
  • Loss: 0.1569
  • Cross-validation: 96.62% (+/- 0.25%)

Model Architecture

CNN (Best Performing Model):

• Input Layer: Image input with dimensions 28x28.
• Convolutional Layers: Two or more Conv2D layers with ReLU activation and MaxPooling.
• Fully Connected Layers: Dense layers with ReLU activation and dropout for regularization.
• Output Layer: Softmax for multi-class classification (digits 0-9).

The CNN architecture outperformed the MLP models due to its ability to capture spatial hierarchies in the image data, which are crucial for digit classification.

Evaluation

The models were evaluated using cross-validation to ensure robust performance across different subsets of the data.

• Cross-validation accuracy: 96.62% with a standard deviation of +/- 0.25%.
• Accuracy and Loss were recorded for each epoch to monitor training progress.

Installation & Usage

To run this project, follow these steps:

1. Clone the repository to your local machine:

   git clone https://github.com/theo-serralta/MNIST_classifier.git

2. Navigate to the project directory:

   cd MNIST_classifier

3. Open the Jupyter notebook:

   jupyter notebook Projet_MNIST.ipynb

Requirements

Make sure the following dependencies are installed:

• Python 3.x
• TensorFlow
• Keras
• NumPy
• Pandas
• Matplotlib
• Scikit-learn

Conclusion

This project demonstrates the power of CNNs for image classification tasks, especially when dealing with noisy data. The CNN model achieved a high accuracy of 96.46% on noisy MNIST images, outperforming the MLP models. The use of convolutional layers allowed the model to capture important features, leading to better generalization and robustness in the presence of noise.

Future Work

Potential improvements and future directions include:

• Experimenting with deeper CNN architectures such as ResNet or DenseNet.
• Using advanced data augmentation techniques to further improve robustness.
• Applying the model to other noisy datasets to test generalizability.

---

Author

Theo Serralta