Evaluation of graphic data corruptions impact on artificial intelligence applications
O Molokovich, Andrey S. Morozov, Nafisa Islamovna Yusupova, Klaus Janschek · IOP Conference Series Materials Science and Engineering · 2021
Abstract The article is devoted to the problem of the reliability of applications based on artificial intelligence. The authors made an attempt to evaluate the impact of the graphic data distortion at the input of a convolutional neural network on the result of image classification. The experiment is based on the fault injection method. A series of independent tests were carried out for such distortions as Gaussian noise, salt and pepper, speckle and Poisson noise, as well as median blur, motion blur, scene brightness changing, rotation, rain, and snow. The results showed that Gaussian noise was the least critical distortion; environmental conditions (rain, snow, brightness) and image rotation up to 20 degrees are less critical than focus losing and motion blur, while the most critical distortion is speckle noise. It was verified that preprocessing the input data of the neural network improves the accuracy of image recognition.