Testing a Cold War-Era AI on Satellite Image Datasets

I came across a cool model developed during the Cold War. I wanted to see how it would perform at image recognition, so I downloaded the UC Merced Land Use Dataset and wrote a script to add Gaussian noise to the photos and measure performance over a series of trials using Monte Carlo simulations. It is very efficient and appears well suited for FPGA implementation. It only uses about 50 MB of RAM. The satellite photos are converted to grayscale, downscaled to roughly 32×32, and converted into a fingerprint that is roughly 128 bytes in size. Therefore, the database of 800 TIFs is about 100 KB total. I’ll include the test and debug images so you can see how the process works. The model basically selects the stored pattern that best matches the noisy input based on what it has learned from the data.