When it's hard to fish for boulders, you need a neural net
You can get data about the ocean floor where humans cannot easily go, identifying where there are boulders or debris. But how easily and quickly can you interpret that data?
It's important to get accurate information for things like hydrographic and geophysical surveys, sub-sea cable and pipeline inspection, or identifying unexploded ordnance. Experts working at marine survey contractor Bibby Hydromap can make sense of raw data in side-scan sonar images by deciphering the noisy greyscale pictures.
But that manual process is expensive in both time and cost. Our IBM Research team developed and trained a method for object detection that Bibby can use. It's a convolutional neural network (CNN for short) that adopts a specialised kind of linear operation, and is inspired by biological processes.
Take a look for yourself in our demo. See how the team trained a deep learning model that accurately detects boulders in thousands of images in less than thirty minutes on a standard personal computer, and in only a few minutes on an IBM Power8 machine.