Program In Computing

PIC Spotlight

Mathematical Methods for Image Deblurring

Anirudh Kannan (Electrical Engineering + Applied Mathematics)

The same photo of a man behind a camera on a tripod, shown twice: motion blurred (left) and deblurred (right).

What happens when an image becomes blurry, and can mathematics help us recover what was lost? Image deblurring is a fundamental problem in computational imaging, with applications ranging from medical imaging and scientific microscopy to everyday photography. In this talk, I'll introduce how we can model image blur mathematically and use techniques from linear algebra, optimization, and numerical analysis to reconstruct sharper images. Through visual examples and simple computational demonstrations, I'll explore why image restoration is more challenging than it might initially seem, how mathematical assumptions help us recover useful information, and what these methods reveal about the relationship between mathematics and the images we see.

Pokemon Type Classifier

Two example predictions from the classifier. Left: a Pokémon whose true type is dragon was predicted to be ice and normal, which is wrong; the model's highest confidences were normal 9.2%, ice 7.8%, bug 7.2% and rock 6.9%. Right: a Pokémon whose true types are dragon and grass was predicted to be grass, which is partially correct; the highest confidences were grass 96.7%, dragon 50.6%, dark 14.2% and rock 6.3%.

Pokémon is a Japanese media franchise, created and owned by the company Game Freak, based on cartoonish creatures by the titular name. Each Pokemon in the franchise has a type of its own, chosen as a singular or a combination of 18 different types, which give it special abilities and advantages in combat. However, while Game Freak claims to design these creatures to match the type they represent, a lot of designs are often not actually very telling to the human mind. So, in PIC 16B, my team and I went in with an idea to test how good an ML model, already trained on real life creatures, could do in learning what design choices go into different type combinations and use that to guess types for new unseen Pokemon. After a quarter of work, what we found resulted in interesting conclusions about not only Pokemon designs, but of the many issues and considerations one must consider in classification models in general.