Classification of image distortion type in task of noreference image quality assessment

Development of noreference image quality assessment is a very popular research task. However, most of the proposed noreference quality assessment algorithms are distortion dependent. Which means that they can predict quality of image with specific distortion type. Аim was to develop noreference image quality algorithm that can predict image quality without any information about the type of distortion introduced to the image. We found the following solution to this task.

Firstly, the type of image distortion is classified by the algorithm. The algorithm can choose among the distortion types known to it. After that system assessing quality of the image with the algorithm specially developed for the certain type of image distortion. This system can be expanded on the other types of distortion by adding quality assessment algorithms to it.

We propose the method of automatic image distortion type classification by means of machine learning methods in this article. Six types of possible image distortion were used: Gaussian noise, salt and pepper impulse noise, artifacts introduced by JPEG and JPEG2000 compression algorithms, image blur caused by median filtration and motion blur. Algorithm based on support vector machine was used as a training method.