Using Deep Learning to Remember Pearl Harbor
I was inspired today by this article from Newsweek titled "25 Striking Photos of the Hawaii Attack." The first cover photo was striking indeed! It had been colorized somehow, which really brought it to life!
All the rest were in black and white, but the content was amazing! I wondered what they would look like in color as well. That's where Artificial Intelligence and Deep Learning come in. I'd seen a method for using Deep Learning to perform Neural Style Transfer, like this one by Adrian Rosebrock on PyImageSearch.
Basically, you're training the network to take the style of one image and transfer it to another. Pretty cool! I was also pretty sure I'd seen someone using Deep Learning to colorize video, so I looked it up.
Sure enough, I found this great paper by Richard Zang titled, Real-Time User-Guided Image Colorization with Learned Deep Priors. No need to go into the details, but he has a github repo with a handy jupyter notebook you can follow to try it out yourself.
The hardest part was getting Caffe installed since I usually work with TensorFlow. It took several hours to get everything configured correctly, but the result was these stunning images shown below. Well worth the effort!
The network was pre-trained, but initial guesses weren't quite creating the dramatic pop that the cover photo did, so I used the colorized cover photo as a global reference, and that refined the color predictions on the other images. I think they turned out pretty nice!
I'm grateful to those who served and fought in WWII. I wish I'd gotten to know more veterans to understand their stories better, but I'm glad we have some pictures to help remind us of their sacrifices.
Credit for photo captions goes to Newsweek and their excellent article. Be sure to check it out for some historical content behind the images.