Journal of Intelligent Systems: Theory and Applications

Abstract

The single image super resolution problem has been extensively studied in the literature using various deep learning-based techniques. Super resolution based on deep convolutional networks has become a rapidly growing area of interest with many practical applications. However, the first studies based on deep learning were based on convolutional neural networks and peak signal-to-noise ratio (PSNR) oriented. Thanks to the models developed based on generative adversarial networks (GAN) in recent years, it has been determined as the main objective to increase the visual quality; however, this is not seen when the image quality metrics are examined. In this study, both mean square error and perceptual loss values were used for the network loss used during the training of the network. Also, the combination of three different training datasets was used as a new training dataset. As a result of these factors, both the visual quality has been increased and a significant increase has been achieved in the image quality metric values. In addition, batch normalization layers are not included in the network architecture and the training speed of the deep network architecture is increased by using the skip connection technique. The success performance of the proposed model was compared with the state-of-the-art models in the literature. Here, the peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) values were calculated and evaluated separately for three different test data sets commonly used in the literature. When the obtained results are evaluated, it is seen that the proposed model is more successful than other models and generates higher quality images. When all the findings are evaluated, it is seen that the proposed model is a more efficient model in terms of both success and training speed compared to state-of-the-art models.