This project involves the training and evaluation of Convolutional Neural Network (CNN) models for classifying fashion items using the Fashion MNIST dataset. The main goal is to achieve high accuracy in recognizing different fashion categories.
The dataset used is the Fashion MNIST dataset, a collection of 28x28 grayscale images of 10 fashion categories, such as T-shirts, trousers, dresses, etc.
The CNN model is designed with three blocks, each consisting of convolutional layers, batch normalization, max-pooling, and dropout for effective feature extraction and classification.
The models were trained using a subset of the data for a specified number of epochs. Additionally, data augmentation was implemented to enhance the model's ability to generalize.
Evaluation metrics include accuracy, loss, and a detailed classification report showcasing precision, recall, and F1-score for each category.
- Base Model Accuracy: 92.22%
- Alternative Model Accuracy: 92.18%
- Final Model Accuracy: 93%
- Plots illustrating training and validation loss and accuracy over epochs.
- Confusion matrix for a detailed understanding of classification performance.
- Correct and incorrect predictions showcased through image plots.
The project demonstrates the successful training and evaluation of CNN models for fashion classification. The models show promising accuracy, and visualizations provide insights into their performance.
- Monitor and assess model performance over time.
- Consider fine-tuning or experimenting with different architectures.
- Deployment for predictions on new fashion images.
Feel free to explore the Jupyter notebooks and code for detailed insights into the project.