Image Analysis Algorithms

Image analysis algorithms are ubiquitous in our modern day lives. Image analysis algorithms have changed dramatically over the last decade with the introduction algorithms using deep convolutional neural networks. These algorithms are heuristic by design, and they perform significantly better than the best algorithms manually created by humans. However, these algorithms operate as so-called black boxes of functionality and they require an enormous amount of data to train. We have shown that automatic programming is capable of improving manually created image analysis algorithms by inferring heuristic program functionality using a relatively small amount of training data. This method can therefore provide a transparent alternative of inferring heuristic image analysis algorithms. We are currently working on improving the Canny edge detector and other related algorithms, and we are planning to move on to more complex algorithm types such as object classification and identification algorithms in the near future.

--Lars Vidar Magnusson

Publisert 11. mai 2018 13:28 - Sist endret 11. mai 2018 13:28