Test Images for Training Convolutional Neural Networks for Image Contrast Assessment

In the paper the construction of test images for the training of convolutional neural networks for image contrast automatic assessment is considered. The test image database is formed by using a random number generator at a given coefficient of filling the image with pixels with a given brightness (from 0 to 255), the remaining pixels have a brightness value of 255. This method of construction makes it possible to estimate the global absolute contrast of test images for training the convolutional neural network to assess the contrast of images. A convolutional neural network is constructed to train it on the constructed test images and to verify its functioning on real test images of the database TID2013. The structure and parameters of the neural network are presented, as well as the results of training and test image accuracy.