Developed CNN that reads images of shoes and predicts their brand (70% accuracy)
A dataset from Kaggle contains images of shoes of 3 different brands: Converse, Nike, and Adidas. The goal of this project was to make a deep learning model that predicts the brand of shoe given the image. The dataset contains only a few hundred photos, so transfer learning using ResNet50V2 was employed. Image augmentation was also applied to broaden the dataset.
The base model was ResNet50V2, and a Dense layer of 1024 nodes was added on top of this. The below image shows the overall architecture of the neural network. The top 3 layers are the last layers of the ResNet model. Global average pooling replaces the top layer of the model. Then an additional dense layer is added. Then, a dropout layer is added to prevent overfitting. Lastly, the output layer is a dense layer with 3 nodes, with each node representing one of the 3 brand options.
The model ended after epoch 9 due to early stopping (min_delta = 0.1, patience = 3). The test accuracy was approximately 70%, which is over twice that of random chance (33%).
