Translating a vision skill, such as text recognition, that was learned from land-based data to sea-based imagery involves domain adaptation—a kind of machine learning. To save time and resources, the Navy is looking to work with a large set of unannotated imagery, which would classify the project as unsupervised domain adaptation.

Optical character recognition (OCR), the technology used to automatically read letters and numbers, is a well-established technique, but its use at sea presents unique challenges. First, there is the question of when exactly to aim the cameras at the lettering to obtain effective results. “Ships are huge and the letters on them are small in comparison, so you can tell there’s a ship long before you have a chance to read anything on it,” Eaton says. At long ranges, additional problems such as atmospheric turbulence tend to blur and distort the shape of faraway lettering. Fortunately, another tool designed by Charles River helps correct blur and improve image quality in the preprocessing stage to make OCR more accurate.

A second challenge is that text recognition systems are not trained on structures unique to ships. To the untrained eye, a series of portholes might look like a cluster of the letter “O” strung together. In addition, the curves of a ship’s hull and the ridges on a cargo container can distort the shape of letters printed on them. Understanding what a letter looks like under ideal conditions is not enough for the algorithm to do its job at sea.