Single MR-image super-resolution based on convolutional sparse representation
- PDF / 1,109,504 Bytes
- 9 Pages / 595.276 x 790.866 pts Page_size
- 19 Downloads / 273 Views
ORIGINAL PAPER
Single MR-image super-resolution based on convolutional sparse representation Shima Kasiri1 · Mehdi Ezoji1 Received: 29 September 2019 / Revised: 3 April 2020 / Accepted: 23 April 2020 © Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract In this paper, a method is proposed to achieve a high-resolution image from a low-resolution image. Because of the illposedness of the super-resolution problem, sparsity constraint is used as a prior, in this work. On the one hand, we use convolutional sparse representation on the whole image different from the patch-based method. On the other hand, we apply fewer filters even in smaller sizes for reconstructing the high-resolution image. Therefore, despite the reduced processing time, the reconstructed image quality is improved compared to the reference methods. In this work, the training images are different in terms of content from the testing images. Experimental results on a variety of MR images indicate improvement in the quality of the high-resolution MR image, in terms of qualitative and quantitative criteria. Keywords Super-resolution · Sparse representation · MRI · Convolutional sparse representation · Dictionary learning
1 Introduction Recently, high-resolution (HR) images play an important role in many image applications, such as medical imaging, satellite images analysis, remote sensing, astronomy, and surveillance cameras. In medical imaging, e.g., magnetic resonance imaging (MRI), the image resolution is a critical parameter during an investigation and for early and accurate diagnosis of disease, and also in the treatment process [1, 2]. However, acquiring HR MRI is a timeconsuming procedure and needs more complex/expensive imaging devices. This procedure has emotional cost for patients and increases maintenance costs and power consumption. Therefore, software-based techniques to increase the resolution of MRI data have attracted interest from researchers. The process of generating HR images by one or more low-resolution (LR) images is called image super-resolution (SR) [1]. SR methods can be categorized into two groups [3]: single-image SR (SISR) methods [4–8] and multi-image SR (MISR) methods [9–13]. SISR method can achieve HR image by only one LR image. In contrast, in the MISR
B 1
Mehdi Ezoji [email protected] Department of Electrical and Computer Engineering, Babol Noshirvani University of Technology, Babol, Iran
method, several LR images are combined to reconstruct the HR image [14]. This combination is carried out in frequency domain or spatial domain [11]. In the frequency domain methods, to eliminate the spectrum aliasing, and also to recover the high-frequency information and details, different transformations such as Fourier transform [15], discrete cosine transform (DCT) [16], and wavelet transform [17] are used. Despite computational efficiency, it is difficult to model the image degradation and use image prior knowledge in these methods [1]. Due to these limitations, many spatial domain methods such as the non-unifor
Data Loading...
