CNN-based single image super-resolution network and biomedical image applications
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
In this thesis, we propose a convolutional neural network (CNN) based single image super-resolution network model with sparse representation by combining three existing state-of-the-art methods SC \cite{sr-sc}, SRCNN \cite{srcnn} and SCN\cite{scn} models with a modified pre-processing step. Firstly, standard gaussian box filter is applied to test image, which is a low-resolution image (LR), to remove low-frequency noises. After that, the given low-resolution image is up-scaled by bicubic interpolation method to the same size with desired output high-resolution image (HR). Secondly, a convolutional neural network based dictionary learning method is employed to train input low-resolution image to obtain LR image patches. Also, a rectified linear unit (ReLU) thresholds the output of the CNN to get a better LR image dictionary. Thirdly, to get optimal sparse parameters, we adopted Learned Iterative Shrinkage and Thresholding Algorithm (LISTA)\cite{lista15} \cite{lista16} network to train LR image patches. LISTA is a sparse-based network that generates sparse coefficients from each LR image patches. Finally, in the reconstruction step, corresponding high-resolution image patches are obtained by multiplying low-resolution image patches with optimal sparse coefficients. Then corresponding high-resolution image patches are combined to get final HR image. The experimental results show that our proposed method demonstrates outstanding performance compare to other state-of-the-art. The proposed method generates clear and better-detailed output high-resolution images since it is important in real life applications. The advantage of the proposed method is to combine convolutional neural network based dictionary learning and sparse-based network training with better pre-processing to create efficient and flexible single-image-super-resolution network.
Description
Keywords
Applied sciences, Health and environmental sciences, Deep-network, Image processing, Machine learning, Medical imaging, Super-resolution