Computer Vision News - September 2016

That really goes back to the core of image processing . An image is defined by positional data which are the indexes in the image or, in other words, counting pixels. The image itself is defined by many sinusoidal contrasts. Put lots of sinusoidal contrasts together, and you get your complex image of clouds and mountains. That comes from the theory books saying that the sinusoidal contrasts in the frequency domain become an index position. What people always forget is that these theory books are showing you magnitude. With magnitude, you’ve lost the information to go back to the spatial domain. What actually is happening is the opposite of the spatial domain. You have the positional data and the movement data in the frequency domain defined as sinusoidal contrasts. You need those two things. You need the positional data with the sinusoidal contrasts. You also need your indexes or the image itself to go back again to the spatial domain. If you have the ability to recognize positional data like shapes and movements in the frequency domain, you have something which is much faster and more robust. It manages to model much more complex things because you’re not counting pixels anymore . You’re really looking at changes in information. Think of it from the point of view of gaze tracking . Imagine having a gaze tracking solution in which you are not counting pixels around the irises and around the edges of the eyes. You’re simply looking at the change of information as the eyes move. Using the change of information in the movement data in the frequency domain, you have something much more solid and robust which needs less data. Plus you have something which is a model. That is what is able to capture much more difficult processes in a model. Therefore, Cogisen ’s vision is to improve the underlying data for industry, allowing all of the deep learning and machine learning methods to reach a better performance. Not only that, if you work like this in the frequency domain, there is much less data required to do learning and training. You won’t have this incredible need for GPU acceleration. In this way, you also have the solution for the Internet of Things . 10 Computer Vision News Application Application