Neural Network Image Recognition Robustness with Different Augmentation Methods

Using numerical simulation, we obtained the dependence of image sets recognition accuracy with different Gaussian kernel size values on the training set augmentation methods using neural networks. We show that the training dataset augmentation method significantly affects the recognition accuracy and the dependence of recognition accuracy on the Gaussian blur kernel size in the recognized images. This research showed that the augmentation method consisting in distortion of half of images from the original training set can provide a stable neural network performance for distorted images recognition and significantly increase the recognition accuracy. The proposed augmentation method is efficient to use in various computer vision systems.