Modeling of applicative noise in images using deep neural networks

Abstract

The process of modeling specific defects in digital images is of great practical importance in data augmentation for trainable algorithms for image recognition, classification and analysis. Applicative noise is one of the common types of digital image distortions in technical vision systems. The effect of applicative noise is manifested in the replacement of individual sections of a useful image with fragments with random amplitude, texture and shape. The article deals with the problem of modeling applicative noise on images using modern algorithms for statistical processing and machine learning. The main types of algorithms and methods for applicative noise modeling are briefly described. The description of the nonparametric methods for texture synthesis, methods of applicative noise generation as local occlusion areas of source images, procedures for texture synthesis using convolutional neural networks, and generative adversarial neural networks for applicative noise synthesis is given. A comparative analysis of the statistical algorithm for generating applicative noise as local areas of occlusion with a trainable algorithm for generating images, distorted by applicative noise, based on the use of GAN models was carried out. The TILDA Textile Texture Database data set was used to test the algorithms. A method based on the use of deep convolutional classifiers for an objective assessment of the realism of the resulting applicative errors is given. The results of the study of the identity of generated and natural applicative noises in the image are analyzed.