This project shows a set of Jupyter Notebooks demonstrating a variety of Convolutional Neural Networks models I built and tuned to classify images for the Intel Image dataset. The dataset has been split into train and test folders, each of which contains a series of 150x150 images that range from 6 classes.
- 1-Convolution Layer CNN: Trained a simple CNN classifier with 1 convolution layer, 1 max-pooling layer, and 2 dense layers, which achieved 94% accuracy for train and 72% for test. Tuned with data augmentation and achieved lower accuracy scores (~70%) yet more stable learning curves.
- 3-Convolution Layer CNN: Trained a deeper CNN classifier with 3 convolution layers, 3 max-pooling layers, 1 dropout layer, and 2 dense layers, which achieved 91% accuracy for train and 80% for test. Tuned with data augmentation and achieved higher accuracy scores (~80%) compared to 1-Conv CNN and more stable learning curves.
- VGG16: Used a pre-trained model, VGG16, to extract image features, then trained a neural network classifier with 3 dense layers and 2 dropout layers, which achieved 92% accuracy for train and much higher score, 86%, for test.
- pandas
- numpy
- matplotlib
- seaborn
- sklearn
- tensorflow