Convolutional Neural Network Assessment of Image Quality Based on the TID2013 Database

The problem of automatic neural network assessment of image quality based on the widely used TID2013 image database is considered in the paper. Image quality assessment without the use of basic reference images, as one of the preliminary stages of their processing, is of great importance from the point of view of the information capacity of the images, the necessity and possibilities of enhancement of the image quality when solving the next tasks of segmentation, classification, recognition, etc. A preliminary analysis of the used image quality assessment metrics and, as an alternative, the use of neural networks, primarily deep neural networks, was conducted. The structures of used neural networks and variants of input data for training neural networks in the problem of image quality assessment are presented. The first variant is a regression problem, and the other one is a classification task. Keras software using TensorFlow for tensor operations is used to build and study neural networks.