Fashion-MNIST Image Classifier

Deep Learning with TensorFlow & Keras

Overview

This project explores image classification on the popular Fashion-MNIST dataset (60,000 training and 10,000 test grayscale images of clothing items). The model classifies images into 10 fashion categories including shirts, trousers, sneakers, and bags.

Implemented with Keras Sequential API, the network learns dense feature representations from 28x28 pixel images, achieving strong performance while remaining compact and interpretable.

Model Architecture

Flatten layer converts 28x28 image to 784 vector input.
Dense layer with 300 hidden neurons, ReLU activation.
Dense layer with 100 hidden neurons, ReLU activation.
Output layer: 10 neurons, Softmax activation (one per class).
Loss: Sparse categorical cross-entropy, Optimizer: SGD.

Training for 30 epochs showed rapid improvement in classification accuracy on both training and validation sets.

Results

The trained model successfully recognized clothing types with good accuracy. Example predictions are shown below (model’s predicted label vs. true label):

Future Improvements

Introduce Convolutional Neural Networks (CNNs) to capture spatial patterns better.
Experiment with data augmentation to improve model generalization.
Use advanced optimizers like Adam for faster convergence.
Deploy the model with a simple web interface for real-time classification.

GitHub link

Explore the Code

Full training script and saved Keras model available on GitHub.

View on GitHub