Super-resolution Augmentation of Large Images with Architecture Treatment of Deeply Recursive Convolutional Network

Modifications in the architecture of deeply recursive convolutional networks (DRCN) are implemented and modified DRCN (MDRCN) is introduced to attain super-resolution (SR) in relatively large images (256×256) and RGB space. Identical components with the expressed baseline (including the embedding, inference, and reconstruction steps) are utilized in this research. The reconstruction stage is divided into two parts: a) Concat. A and b) Concat. B. A particular number of the Div2k and Yang T91 images (891) are utilized in the training phase. High-resolution (HR) and low-resolution (LR) images have been created with bicubic interpolation in the training and test phases. Datasets of Set5, Set14, B100, and Urban100 are used for the testing phase. The criteria of PSNR and SSIM are utilized to examine the attained results in the testing phase that have improved in MDRCN relative to the investigated algorithms in 2× scaling and all datasets. The amounts of these criteria significantly were improved in MDRCN by augmenting the scaling factor to 4×, especially in the Urban100 dataset. According to the observed outcomes, it can be concluded that MDRCN has proper efficiency in the low-scaling compared to the intended models.