Activist and developing countries alike have a small voice on the Internet as they are either too repressed or poor to vocalize their frustration with often unruly conditions or activity. Likewise, smart-phone users do not have the time to remove that "one pimple" that ruined Thanksgiving photos and as such often succumb to difficult circumstances associated with smart-phone photography.
As a solution, MiKKO offers activist and those in developing countries cheap and available software which would allow citizens to develop media in support of their causes. As for the common iPhone or Android carrier, applying a style to an image which has "perfect" lighting would allow for the beautification of any photo.
MiKKO was developed in Python using Tensorflow, Google's professional library for machine learning. The algorithm was inspired by Leon Gatys et al.'s "Image Style Transfer Using Convolutional Neural Networks."
The team's success was the byproduct of a rather clever implementation of the VGG network. Rather than training the underlying convolutional neural network, the algorithm's implementation was quick and consisted of many layers which contributed to a greater image quality.
Coming in with an AI-based solution as opposed to a hardware solution like last year, we learned that the finicky nature of deep learning models ultimately reduces a researcher's ability to implement solutions in a timely manner as is necessary in a Hackathon.
Improving the speed of the MiKKO algorithm would be a necessity if commercialized. Fortunately, developments in arbitrary style transfer allow for real-time style transfers, although the algorithms are convertible more computationally expensive to train.
Log in or sign up for Devpost to join the conversation.