Neural Network Method for Detecting Textural Anomalies in a Digital Image

Abstract

Modern computer vision systems often use neural networks to process images. But to use neural networks, you need to create databases to train these neural networks. In some cases, creating a training database takes the vast majority of the project's financial and human resources. Therefore, the actual task of finding methods to improve the quality of learning neural networks on small data is considered in this article. The ability to process data, which nature was not present in the original training database is relevant, also. To solve the problem of improving the quality of image segmentation by textural anomalies, this research is proposed to use as input to the neural network not only the image but also its local statistic data. It can increase the information content of the input information for the neural network. Therefore, neural networks do not need to learn to choose statistical features but simply use them. This investigation classifies the requirements for image segmentation systems to indicate atypical texture anomalies. The literature analysis revealed various methods and algorithms for solving such problems. As a result, in this science work, the process of finding features in the photo is summarized in stages. The division into stages of search for features allowed to choose arguments for methods and algorithms that can perform the task. At each stage, requirements were formed for methods, that allowed separate the transformation of image fragments into a vector of features by using an artificial neural network (trained on a separate image of the autoencoder). Statistical features supplement by the vector of features of the image fragment. Numerous experiments have shown that the generated feature vectors improve the classification result for an artificial Kohonen neural network, which is able to detect atypical image fragments.