Efficient super-resolution and MR image reconstruction networks

Abstract

The process of generating a high-resolution image from a low-resolution image is known as image super-resolution. From a low-resolution image, many high-resolution images can be produced. Therefore, super-resolution is a difficult problem with ill-posed nature. In recent years, many deep neural networks were suggested to retrieve missing high-frequency features. Models started to get deeper to enhance performance. But, using these models in devices with limited resources is challenging due to their high computational complexity and memory requirement. Therefore, in this thesis, different knowledge distillation schemes are suggested to compress super-resolution networks. Offline- and self-distillation frameworks are used to decrease the number of repeating residual blocks in SRResNet and EDSR networks. Test results are obtained on benchmark datasets for scale factor four. After many experiments, results show that previous layers learn to produce similar outputs to following layers. However, when redundant layers are removed, performance of the compressed networks are less than their vanilla trained versions. Therefore, further study on this subject is required to prevent performance decrease. Magnetic resonance imaging is a valuable tool in medicine to identify diseases. To obtain a high quality image, enough $k$-space data is needed. This increases the necessary scan time. As in other image processing fields, deep neural networks are also used in magnetic resonance image reconstruction task from undersampled data. Since super-resolution also aims to restore missing information, using some concepts from super-resolution can help improve reconstruction performance. In this study, Iterative Up and Down Network is proposed to solve this problem. Network benefits from iterative up and down sampling framework and multiple scale factors. Training of the network is done using fastMRI dataset. Test results are obtained on two datasets which are fastMRI and IXI. Proposed network has processing units and test results show that increasing the number of these units improve the performance of the network. Also, using multiple scale factors further increased the performance. Quantitative results show that suggested approach is superior than some well-known state-of-the-art networks. When qualitatively compared to other methods, suggested model performs favorably.