Nike Shoe Model Image Classification Using TensorFlow and Transfer Learning

This project aimed to develop an automated system to classify images of Nike shoes into predefined categories. The dataset included multiple shoe models such as Air Jordan 1, Air Max Plus, Mercurial, and React. Transfer learning was applied to leverage pre-trained models, significantly reducing training time while maintaining high accuracy.

Training a deep convolutional neural network (CNN) from scratch requires a large labeled dataset and extensive computational resources. The available dataset was relatively small, making it challenging to achieve high accuracy without overfitting. In addition, the pre-trained model MobileNetV2, while efficient for generic image classification, struggled to differentiate between Nike shoe models. It often classified them into broader generic categories such as running shoe, failing to capture model-specific distinctions.

A transfer learning approach was applied using MobileNetV2 from TensorFlow Hub as a pre-trained convolutional base to extract image features. All images were resized to 224×224 pixels and augmented through rotation, flipping, and zoom in order to improve the model's ability to generalize. The extracted features were passed through a fully connected classifier trained specifically on the target Nike shoe categories. This approach allowed faster convergence and better performance compared to training a convolutional neural network from scratch.