A CNN Based Single Image Super-Resolution Using Residual Networks With Non-Local Multiple Image Phases

Image-super resolution is a fast-growing research field that has been enhanced by the introduction of deep learning methods that achieved greater results and performance. Convolutional deep networks that utilize residual architecture have shown the best performance. One of the best models that use this architecture is the EDSR model [16], which showed a great result in extracting and reconstructing images but the features that it extracts are not always optimal due to the locality of the convolutional window which might result in missing some general key features about the image from the areas outside the convolution kernel of the image. In this project, we propose an improvement on the convolutional network architecture that uses residual blocks to detect the local features of the image by adding a phased version of the input that will add the missing nonlocal features and improve the quality of the feature space that will result in a better reconstruction. We achieved our objective by introducing a new phasor block to the model which will create different perspectives of the image which we trained using a smaller version of EDSR for each phase then concatenated the results of the original and the phases into one big feature space containing deep and shallow features which enhanced the reconstruction of the image. Keywords: Convolutional Networks, Residual Networks, EDSR, Phasor