Basic AI

A small scale Astronomical CNN

What is this project about?

This project tests whether small datasets can be used to train CNN and AI models effectively. I use image augmentation to see which level of image variation affects the final result. To show that this can be done on my own principles, the project uses only open data from NASA and ESO, so no one's work is used without permission.

The problem

AI models are often trained on large datasets to improve performance and generalization. This practice leads to work by human artists and writers being stolen by corporations. This project tries to address this issue in two ways.
1. I only use open, publicly available NASA and ESO images.
2. I use augmentation (changing the existing images to create more variety) to make a small dataset generate a larger amount of useful data, then test the results against benchmarks to see how the final result is affected.

Idea

Image augmentation uses pre-existing data to create new data samples that can improve model optimization and generalizability.

Try it out!

Click on one of the image classes then click on one of the factors of Augmentation!



Model Architecture

CNN model architecture Four blocks of Conv2D and MaxPooling2D with 16, 32, 64 and 64 filters, then global average pooling, a 64-unit dense layer and a softmax output with one unit per class. Conv2D(16, (3,3), activation='relu'), MaxPooling2D(), Conv2D(32, (3,3), activation='relu'), MaxPooling2D(), Conv2D(64, (3,3), activation='relu'), MaxPooling2D(), Conv2D(64, (3,3), activation='relu'), MaxPooling2D(), GlobalAveragePooling2D(), Dense(64, activation='relu'), Dense(len(class_names), activation='softmax') Inline SVG is not supported in your browser. Please use a modern browser to view this chart.

Results

Inline SVG is not supported in your browser. Please use a modern browser to view this chart.

Check out the GitHub!

You can run the model by yourself! Use any dataset of your choosing then run augment.py to augment the data then main.py to make the model and train and test it! Check the ReadMe for more details!