Astronomical Object Detection
Space, it is a marvel. There is just so much unknown, which is what makes it so interesting. For astronomers, studying certain astronomical objects is hard without enough data. Take for example Infrared Dark Clouds (IRDCs), large dense molecular clouds which are thought to be the early stage of large mass star formation. Due to their significance, detecting as many of these IRDCs are vital. Finding more IRDCs means that scientists have more data to support conclusions. For almost all astronomical objects, manual search is the only method to finding them.
Making an algorithm that could churn large datasets and spit out coordinates of IRDCs would prove to be highly beneficial. Which is why I set out on the quest to make one with my mentor Dr. Joesph L. Hora, using Deep Learning. Look at this picture which contains 2 IRDCs:
You might look at it and see 3 dark areas, but only the bottom 2 are actual IRDCs. This might puzzle you since, the one on the top right also looks like a dark cloud. Since they come in all shapes and sizes, finding hard coded rules for detection didn’t seem like the right route. This is what lead me to utilize the power of Deep Learning. Instead of me trying to figure out lots of rules, I let a Convolutional Neural Network (CNN) do that for me. You might have heard of those, for example when you open up Google Photos on your phone, you might wonder how it automatically sorts the pictures based on the content of the pictures itself. It also uses a CNN to classify the images.
The power of the CNN is it figures out what to look for. But I won’t go into CNNs, just know that they learn to classify based on the training images you feed them. I used Tensorflow‘s Inception model for my CNN. But I can’t just slap on a CNN and call it a day, in an image there might be multiple IRDCs. A CNN can only make the final decision, is the figure inside this bounding box an IRDC or not. Though, how do I figure out those bounding boxes? After the image goes through some filtering processes we are left with this:
This is the exact image from earlier, but using a sliding window and Gaussian Blurring I was able to simplify it. Now we can clearly see the clusters, each cluster is one of the black figures. The best way to find each cluster is through K Means Clustering. I won’t explain how this works either, but just know it is able to find each cluster. Once the clusters have been found, we have our bound boxes!
Each red box represents a bounding box. So all that is left is to send each section of the image, into the trained CNN and we should be done. This algorithm proved to be robust, with an accuracy of 96%. Which means it is on par with humans, but where it went beyond was when it discovered new IRDCs on the same section of the dataset that has already been extensively search by scientists.
Artificial Intelligence and its subset Machine Learning have been the limelight in so many fields where we are able to find patterns in data and automate process that were thought to be impossible. There hasn’t been much activity in the Machine Learning’s application in astronomy, and I hope this is one more step to bridge that gap.
About the Author
Jyo Pari is an Artificial Intelligence researcher and
student at Acton BR in Massachussetts. His main focus
is on applying Deep Learning in different fields such as Astronomy.