One of the most basic uses of convolutional neural networks is image classification, where the image is fed into the network along with some categories, and the network predicts which class the image belongs to. In this particular implementation, we use a satellite imagery dataset of ships from Kaggle. The dataset is well made, with multiple images of not just ships, but also similar pictures like bridges, pictures of obscured boats, the open sea and terrain.
The model in the notebook was made with the TensorFlow and Keras frameworks. After training the model, a validation accuracy of around 96% was achieved with the Adam optimizer.
In this directory, there is an interactive python notebook which walks through the steps of building a neural style transfer script with TensorFlow. The code was based off the code found at the TensorFlow documentation for neural style transfer. The style model works quite well with the oil painting style and tends to preserve colour in the water color style. The transition of the style output is shown in the gif below, with very minute changes which you can spot if you look closely.
Also given in the module is reusable model from the TensorFlow hub, which produces some rather different images with way more emphasis on the style's colors, as shown below.
Old Version
.png)
New Version
.png)
The directory also has some style images and some other images to test the program with, though you are free to try your own.
In this we again take the image of the BITS Pilani Rotunda and pass it through the VGG19 architechture and plot out the feature maps which it generates while running the image through the model. The feature map from the last convolutional layer in the first block is here.
.png)