Single-image super-resolution reconstruction based on global non-zero gradient penalty and non-local Laplacian sparse coding
Introduction
High resolution (HR) images are desired in most electronic imaging applications such as biometrics identification, medical imaging, military surveillance and so on. Unfortunately, due to the physical limitation of relevant imaging devices, the images we observed are usually noisy, blurred and downsampled. To obtain the HR image, we can either reduce the pixel size by sensor manufacturing techniques or increase the chip size of charge-coupled device sensors, which are both severely limited in increasing the cost of digital imaging systems and reducing the processing efficiency of real-time environment [1]. Therefore, the signal processing methods are selected to reconstruct potential details and features hidden in the low resolution (LR) image.
Generally, the existing methods can be classified into three categories: interpolation-based methods [2], [3], [4], [5], regularization-based methods [6], [7], [8], [9], [10], [11], [12] and example-based methods [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32]. However, the interpolation-based methods are usually prone to yield overly smooth images with ringing and jagged artifacts when a larger magnification ratio (such as a factor of more than double) is performed. The regularization-based methods are limited in modeling the visual complexity of the real images and selecting correct regularization parameters. The focus of this paper is the example-based methods because the methods are of stronger capability of image super-resolution reconstruction as the magnification becomes larger.
In recent years, the example-based methods have been explored. This kind of methods presumes that the high-frequency details lost in the LR image can be predicted by learning the co-occurrence relationship between LR training patches and their corresponding HR patches. Freeman et al. [13] first proposed a relation model between the local regions of images and scenes by using the Markov network. However, this approach depends heavily on a large training data set. Chang et al. [14] introduced locally linear embedding from manifold learning to process the image super resolution task. Although this method has advantages over Freemanʼs, the problems of the number of neighbor and feature representation of LR and HR image patches remain unsettled. Others based on learning primal sketch prior are proposed [15], [16], [17]. However, due to the lack of priors of textures and details, they are weak in hallucinating both textures and details. Recently, Yang et al. [20] proposed a sparse coding to reconstruct HR images. In their works, HR image patches are sparsely coded under over-complete dictionary learned with coupled pattern. Considering that there are different types of image patches (such as smooth regions, texture regions and edge regions) in image, Jing et al. [21] proposed a multi-space sparse representation method, which first decomposes image into structural and textural components, and then the HR image is recovered by coding the structural and textural components respectively. And Yang et al. [22] also proposed a multiple-geometric-dictionaries-based clustered sparse coding scheme, which first trains the geometric dictionaries of geometric clusters, and then HR image patches are sparsely coded under different geometric dictionaries. Recently, the geometric structure information of image patches has been successfully used in various image processing applications [33], [34], [35]. Some research works have pointed out that the reconstruction quality greatly depends on geometrical structures of the data [31]. Hence, it is important to explore these potential geometrical structures to enhance existing sparse coding stability. By transferring the non-local information of images patches into the sparse coefficients, the non-local sparse coding methods [30], [31], [32] are widely proposed for image reconstruction. However, the methods lose the difference among the image patches. Moreover, they are not effective in reconstructing images which contain the patterns with strong edge and reconstruct incorrectly the edge structures (such as continuity and orientation) [23].
To resolve the above problem, we propose a new approach based on global non-zero gradient penalty and non-local Laplacian sparse coding. The overall framework of proposed approach is illustrated in Fig. 1. As shown in Fig. 1, firstly, by exploring the global non-zero gradient penalty (GGP) which can globally sharpen major edges and preserve their geometric structure by increasing the steepness of transition in a sparsity-control manner, the HR edge component can be reconstructed. Meanwhile, by exploring the non-local Laplacian sparse coding (NLSC) which can preserve the difference by exploring the one-to-one relationship of the image patches and the non-local prior simultaneously, the HR texture component can be reconstructed. Then, the global and local optimization (GLO) is applied on the initial image for removing the possible artifacts and making the final image more natural. Figs. 2(c)–(d) show the edge component and the texture component of “Butterfly” image. Fig. 3 shows the decomposition process. Due to the different image components reconstructed by different methods, it makes the obtaining of desired HR image possible. The performance of the approach is tested by various typical experiments in terms of visual evaluation, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Compared with the related single image super resolution (SISR) approaches, the proposed approach has the following characteristics:
- (1)
GGP is proposed for reconstructing edge component.
- (2)
NLSC is proposed for reconstructing texture component.
- (3)
GLO is applied on the initial HR image to further improve the imageʼs quality.
The rest of the paper is organized as follows. In Section 2, we present our SISR approach in detail. The experimental results together with relevant discussions are given in Section 3. Finally, conclusions are discussed in Section 4. In addition, the descriptions of the acronyms used in this paper are listed in Table 1.
Section snippets
Our proposed approach
In this section, we first show how to reconstruct edge component of the desired HR image by the GGP. And then develop the NLSC to reconstruct its texture component. Finally, we present the GLO to further improve the quality of reconstruction image.
Experimental results and analysis
In this section, we perform 3× magnification experiments on ten test images (refer to Fig. 5) to validate the effectiveness and robustness of the proposed SISR approach. In Fig. 5, Moth, Boats, Tiger, Koala and Parthenon come from http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping, and the rests come from [12], [20].
Conclusions
The state-of-the-art sparse coding methods have two common drawbacks. The first one is that they easily reconstruct incorrectly the edge structures of image. The second one is that they easily lose the difference among the patches. To resolve the two problems, we proposed a new approach based on global non-zero gradient penalty and non-local Laplacian sparse coding. Firstly, we developed the global non-zero gradient penalty to reconstruct correctly edge component and the non-local Laplacian
Acknowledgements
This work was supported by the Key Science and Technology Projects of CSTC (CSTC2012GG-YYJSB40001, CSTC2013-JCSF40009), the Fundamental Research Funds for the Central Universities under Grant No. CDJXS11122216, and the National Natural Science Foundation of China under Grant No. 61105093.
The authors would like to thank the editors and reviewers for their valuable comments and suggestions.
Jinming Li received his B.S. degree in computer science from Qufu Normal University, China, in 2008 and the M.S. degree from Jiangnan University, China, in 2011. Currently, he is a Ph.D. candidate in the College of Opto-Electronic Engineering, Chongqing University. His research interests are in information acquiring and image processing.
References (43)
- et al.
Super-resolution image reconstruction using fast inpainting algorithms
Appl. Comput. Harmon. Anal.
(2007) - et al.
A multi-frame image super resolution method
Signal Process.
(2010) - et al.
Region-based weighted-norm with adaptive regularization for resolution enhancement
Digit. Signal Process.
(2011) - et al.
Bayesian combination of sparse and non-sparse priors in image super resolution
Digit. Signal Process.
(2013) - et al.
Example-based image super-resolution with class-specific predictors
J. Vis. Commun. Image Represent.
(2009) - et al.
Novel super resolution restoration of remote sensing images based on compressive sensing and example patches-aided dictionary learning
- et al.
Multitask dictionary learning and sparse representation based single-image super-resolution reconstruction
Neurocomputing
(2011) - et al.
Image super-resolution by curve fitting in the threshold decomposition domain
J. Vis. Commun. Image Represent.
(2012) - et al.
Semisupervised dual-geometric subspace projection for dimensionality reduction of hyperspectral image data
IEEE Trans. Geosci. Remote Sens. (99)
(2013) - et al.
Geometric optimum experimental design for collaborative image retrieval
IEEE Trans. Circuits Syst. Video Technol. (99)
(2013)
Super resolution image reconstruction: A technical overview
IEEE Signal Process. Mag.
Cubic splines for image interpolation and digital filtering
IEEE Trans. Acoust. Speech Signal Process.
An edge-guided image interpolation algorithm via directional filtering and data fusion
IEEE Trans. Image Process.
Markov random field model-based edge-directed image interpolation
IEEE Trans. Image Process.
Image interpolation by adaptive 2D autoregressive modeling and soft-decision estimation
IEEE Trans. Image Process.
Generalizing the non-local-means to super-resolution reconstruction
IEEE Trans. Image Process.
Super-resolution without dense flow
IEEE Trans. Image Process.
Single image super-resolution with non-local means and steering kernel regression
IEEE Trans. Image Process.
Example-based super resolution
IEEE Comput. Graph. Appl.
Super-resolution through neighbor embedding
Generic image hallucination with primal sketch prior
Cited by (8)
Multi-baseline SAR 3D reconstruction of vehicle from very sparse aspects: A generative adversarial network based approach
2023, ISPRS Journal of Photogrammetry and Remote SensingChannel rearrangement multi-branch network for image super-resolution
2022, Digital Signal Processing: A Review JournalCitation Excerpt :Li et al. [5] proposed locally adaptive bilateral total variation and consistency of gradients to keep edges and flat regions in LR images. In order to reconstruct correctly the edge structure and retain the difference among the image patches to be reconstructed, Li et al. [6] proposed a new approach based on global non-zero gradient penalty and non-local Laplacian sparse coding. Liu et al. [9] used non-local self-similarity, cross-resolution similarity, and global structure sparsity to achieve high-resolution images.
Parameter estimation in sparse representation based face hallucination
2014, Digital Signal Processing: A Review JournalCitation Excerpt :The optimum setting enables an appropriate tradeoff between reconstruction accuracy and robustness against noise. Otherwise, too large values will lead to excessive smoothness (or low-precision prediction) while too small values will result in poor robustness against noise on the contrary [20]. In practice, previous face hallucination methods [4–9,20] empirically specify λ by keeping manual tuning until the best ultimate results are achieved.
Research on the single image super-resolution method based on sparse Bayesian estimation
2019, Cluster ComputingAn improved image super-resolution reconstruction algorithm based on centralised sparse representation
2017, Information Science and Electronic Engineering - Proceedings of the 3rd International Conference of Electronic Engineering and Information Science, ICEEIS 2016Noisy image magnification with total variation regularization and order-changed dictionary learning
2015, Eurasip Journal on Advances in Signal Processing
Jinming Li received his B.S. degree in computer science from Qufu Normal University, China, in 2008 and the M.S. degree from Jiangnan University, China, in 2011. Currently, he is a Ph.D. candidate in the College of Opto-Electronic Engineering, Chongqing University. His research interests are in information acquiring and image processing.
Weiguo Gong received his doctoral degree in computer science from the Tokyo Institute of Technology of Japan in March 1996 as a scholarship gainer of Japan Government. From April 1996 to March 2002, he served as a researcher or senior researcher in NEC Labs of Japan. Now he is a professor of Chongqing University, China. He has published over 100 research papers in international journals and conferences and two books as an author or co-author. His current interests are in the areas of pattern recognition and image processing.
Weihong Li received her doctoral degree from Chongqing University in 2006. Now she is an associate professor in Chongqing University. Her current research interests are in the areas of pattern recognition and image processing.
Feiyu Pan received his B.S. degree in measuring and control technology and instrumentations from North China University of Water Resources and Electric Power, China, in 2011. Currently, he is an M.S. candidate in the college of Opto-Electronic Engineering, Chongqing University. His research interests are in information acquiring and image processing.