Repository for Image Classification Projects
This repository contains a deep learning project focused on classifying fashion images into one of three categories: Casual, Formal, and Smart Casual. The project explores both a baseline CNN model and an enhanced version incorporating data augmentation and dropout regularization.
In the notebook, a pipeline was built to:
- Load and preprocess a fashion dataset (
styles.csv+ image folder) - Train a baseline CNN model using unaugmented image data
- Develop an enhanced CNN model using techniques such as:
- Data augmentation
- Batch normalization
- Dropout layers
- Evaluate and compare both models based on accuracy and loss
The models were trained and tested using TensorFlow and Keras with clear separation between training, validation, and test sets.
The dataset consists of:
This project uses the Fashion Product Images (Small) dataset from Kaggle.
styles.csv: Metadata and label information for each imageimages/: A folder of fashion item images
Only three target classes were used: Casual, Formal, and Smart Casual. These classes were filtered and relabeled before model training.
| Model | Validation Accuracy | Test Accuracy |
|---|---|---|
| Baseline CNN | ~85% | ~84% |
| Enhanced CNN | ~96% | ~96% |
The enhanced model significantly outperforms the baseline, demonstrating the impact of regularization and data augmentation techniques.