Experts Successfully Trained An AI To Fix Distorted Underwater Pictures
Aadhya Khatri - Jan 02, 2020
The AI has the ability to de-haze images with green tones without ruining the original structure of the pictures
- Internet Users Are Using Google Street View To Find Deceased Relatives
- You Can Now Use Your iPhone As A Security Key, Google Announced
- Google AI Can Accurately Predict The Weather In Just A Few Minutes
As a result of back-scattering and light attenuation, pictures taken underwater tend to be distorted and blurry. To save photos affected by these phenomena, experts from the Harbin Engineering University proposed a solution, which comes in the form of two AI systems, one can produce realistic underwater pictures, the other learns from what the first one makes and then fix the distorted images.
The team pointed out that the majority of other picture-enhancing systems work without taking the physical imaging models into consideration, so they do not exactly do the task well.
The first AI is called GAN and it is trained with a corpus of 3,733 scenes with labels and their depth maps, mostly consist of sea cucumbers, scallops, sea urchins, and other underwater organisms living in marine farms. They also used some other open datasets, like NY Depth, a library of thousands of photos taken underwater.
After some tests, the researchers said that their approach has the following advantages over others. The first one is the uniform in color restoration; the second is the ability to de-haze images with green tones without ruining the structure of the pictures. The last benefit is that the system is able to maintain the color contrast and brightness over the process.
Experts from Harbin Engineering University were not the first ones to try restoring distorted pictures. We have seen DeepRay, Cambridge Consultants' Digital Greenhouse internal incubator. It was trained with 100,000 still pictures distorted by an opaque pane of glass. Another effort results in DeOldify, a system that can restore and colorize old footage and images.
Microsoft also has its fair share of the work in the form of automatic video colorization. Google has its own algorithm for grayscale colorization without the supervision of humans.