Fashion MNIST CNN
Image Classification with Convolutional Neural Networks
Client
Academic Project
Year
2024
Role
ML Engineer
Stack
Python · TensorFlow · Keras
Status
Complete
Context
Teaching machines to see clothes.
A standard clothing classification dataset with 60,000 grayscale images across 10 categories. This project builds a CNN to classify these images, demonstrating fundamental computer vision and deep learning concepts.
The Vision
CNN fundamentals, production mindset.
The goal was to build a robust CNN pipeline: convolutional and pooling layers for feature extraction, dense layers for classification, with careful evaluation beyond just accuracy.
The Product
CNN classification pipeline.
A convolutional neural network with feature extraction layers (Conv2D + MaxPooling), dropout for regularization, and dense classification layers. Achieved 90.12% accuracy with an F1-score of 89.93%.
Architecture
Network design.
Input
60,000 training images, 10,000 test images across 10 clothing categories.
Feature Extraction
Alternating convolutional and pooling layers for hierarchical feature learning.
Classification
Fully connected layers with dropout regularization.
Output
10 clothing categories with softmax activation for probability distribution.
Outcome
Performance.
Challenge
From notebook to production-ready CNN.
The challenge was building a CNN pipeline that works reliably across all 10 clothing categories without overfitting. The solution: alternating Conv2D + MaxPooling layers for hierarchical feature extraction, paired with dropout for regularization. Achieved 90.12% accuracy with only 60k training samples.
Performance
Performance
90.12%
Accuracy
89.93%
F1-Score
10
Categories
60K images
Training Set
Dataset
Dataset
Fashion MNIST
Train/Test
60K / 10K
Image Size
28x28 grayscale
Classes
10 clothing types
Next case