- Survey made about Super resolution state of the art architecture. (To read survey, you can download file directly. Sometimes pdf is not loading on screen)
- Please check : SRCNN/Deep_learning_super_resolution_survey.pdf
- SRCNN architecture :
Super resolution is to recover Low resolution images to Highresolution images, but there are many ways to recover indifferent way. It is called Regular Inverse Problem or ill-PosedProblem that there are no unique answers as shown in Figure 1.SRCNN(Super-Resolution Convolutional Network) is thefirst convolutional neural network proposed by Dong to dealwith Single Image Super Resolution(SISR) in 2014. At thetime when CNN was just introduced, so only 3 convolutionallayers were used without stacking hundreds of layers, andit showed higher performance values compared to methodswithout deep learning. (Figure 2). This study shows thepossibility that deep learning can be applied to the super-resolution field as well. When constructing an architecture, themeaning of each convolutional layer is interpreted in termsof traditional super-resolution, and each layer is in chargeof patch extraction, non-linear mapping, and reconstruction.From a given low resolution image, first convolution layerextracts feature maps. The Second layers map non linearly tohigh resolution patch problems. The last layer combines theprediction to produce high-resolution output.
Matlab DEMO file for SRCNN Architecture based on paper : C. Dong, C. C. Loy, K. He, and X. Tang, “Learning a deep convolutional network for image super-resolution,” in Proceedings of the European Conference on Computer Vision, 2014.