The research is devoted to studying the influence of data augmentation by means of wavelet filtration on recognition accuracy of noised images by the neural network detector. It is proposed adding noise to samples of the training dataset and then processing them by Wiener filter in the domain of discrete wavelet transform (DWT). Generating of input training images uses the model applying white additive Gaussian noise to distort them. In addition to the Wiener filtering, we considered some thresholding methods in the domain of DWT for images denoising. Based on the obtained experimental data, we compare the results of the detector recognition for the different values of noise standard deviation. Also, we considered the case, when the neural network was trained on noised-augmented images without subsequent filtering. It is shown that the proposed approach allows higher accuracy of recognition to be achieved, than the augmentation, which doesn’t use noise suppression.

Data augmentation, machine learning, neural networks, wavelet filtration

Alimagadov Kurban Alimagadovich,  ,  orcid.org/0009-0008-3420-3048,  National Research University of Electronic Technology

Umnyashkin Sergei Vladimirovich,  ,  orcid.org/0009-0000-1932-8709,  National Research University of Electronic Technology