ОЦІНКА ЯКОСТІ ЗОБРАЖЕНЬ ЗГОРТКОВОЮ НЕЙРОННОЮ МЕРЕЖЕЮ ПРИ ВИКОРИСТАННІ БАЗИ TID2013

Abstract

The article is devoted to the problem of automatic image quality assessment by a convolutional neural network when using the common TID2013 image database for training the neural network. The TID2013 database was chosen for the reason that it contains 25 base real-world images, which were distorted from these images using 24 different distortion methods and with 5 distortion levels, creating a sufficiently large database of 3000 images for training the neural network. For each image, an average expert assessment of its quality is given. All input images for the neural network are divided into two groups - the training set and the validation set. We consider two options for solving the problem of building and training a neural network and the corresponding data structure. The first option is reduced to the task of numerical regression using the numerical values of expert assessments. The second option is the task of classifying the training and validation images into 5 classes according to their quality corresponding to distortion levels. Keras and TensorFlow software tools are used to build and study the neural network. The neural network structures and relevant parameters for training each layer are presented, as well as graphs of accuracy changes for training and validation images during training.