engleski

Application of deep convolutional neural network for image scaling

This paper proposes a model of deep recursive neural network that takes into account the specific features of the input images, performs self-tuning with the definition of optimal scaling parameters and continues training. A unified deep learning neural network DJSR (Deep joint super resolution) was used to improve the quality of the output image. At the beginning the network is trained using an external image database, the tuning is done using the input image samples. The network can then be expanded into several specialized sub-networks, and sample learning is performed. The main idea is to exploit the surplus part-examples between different neighboring image scales. The proposed network contains a degradation model, a degradation discriminator, and a reconstruction model. The degradation model is used to produce realistic lowresolution images. The degradation discriminator ensures the similarity of the example parts of the obtained low-resolution image with the real images. The reconstruction model is used to directly create final high-resolution images by using the resulting images for training.

neural network ; image scaling

The conference was canceled due the low number of presenters registered for the conference, but the Book of abstracts has been published.