Salient edges combined with image structures for image deblurring
Introduction
Blind deconvolution is a common issue in digital image processing. The purpose of blind deconvolution is to recover a blur kernel K and a desirable clear image I from the input blurry image B. If the blur is uniform and spatially invariant, a standard deblurring model is usually formulated as where the stands for the convolution operator, N represents the addition Gaussian noise. Model (1) is named blind deconvolution when the blur kernel is unknown; Otherwise, it is non-blind deblurring [1], [2]. More recently, Dong et al. [3], [4] creatively proposed two simple and effective deep non-blind methods, one of which is learning a spatially variant MAP model for non-blind image deblurring [3], and the other is to perform an explicit deconvolution process in feature space by combining a classical wiener deconvolution framework with learned deep features [4]. Both methods can progressively restore fine-scale structures and details, along with robustness to noise and saturation. As it is hard to find reasonably pairs (I, K) from the blurred image B, several methods utilize extra knowledge for this problem.
In recent years, deblurring methods using image prior information have been leveraged in the maximum a posterior (MAP) framework [5]. For instances, sparse priors of the blur kernels or latent images [6], channel-based priors [7], [8], gradient-based priors [9], [10], patch-based priors [11], [12], [13], low-rank priors [14], [15]. Such statistical priors focus on finding certain features of the image that are sufficient to distinguish between blurred and clear images and prefer clear ones in the MAP framework. But the prior-based approach is time-consuming due to the extensive computation involved.
Contrary to investigated blind image deblurring works such as those mentioned above, the methods that focus on the edge-based are widely used to explicitly or implicitly achieve salient edges to facilitate kernel estimation [16], [17], [18], [19], [20], [21], [22], [23]. Commonly, salient edge methods mainly use filtering techniques [24], [25], [26], [27], [28] for edge extraction to provide salient edge information for kernel estimation, and this simple operation leads to a time-saving implementation. However, most of the previously published literature is very limited because they do not consider the use of image structure while studying the salient edge selection. As a result, the recovery result was not ideal.
We develop a novel structure-based deconvolution term that enforces the estimated latent image to contain finer scales and builds a more expressive model with image structure and salient edges through MAP. In addition, an constraint on the gradient is also applied for edge-preserving and detail removal. Fig. 1 is a visual example showing the effectiveness and robustness of the proposed image deblurring model. The contributions are highlighted as follows.
(1) We develop a novel structure-based deconvolution term that uses mutually guided image filtering (muGIF) to provide guidance information for image structures.
(2) We enhance the robustness of the salient edge selection by building a more expressive model with image structure and salient edges through MAP framework.
(3) A new salient edge selection model is proposed and further the -norm of the gradient is adopted for edge-preserving and detail removal.
(4) Extensive experiments have shown that our approach achieves better results on both benchmark datasets and real scenarios. Quantitative and qualitative evaluations show that our method is comparable to state-of-the-art methods.
Section snippets
Related work
The problem of blind image deblurring is a longstanding issue in low-level vision. Various methods exist to facilitate restoration results to yield reasonably clear images. We will discuss the related work on deblurring methods in this section.
Prior-based methods. In [29], it has been shown that the normalized sparsity prior is beneficial to estimate blur kernels. A method called enhanced low-rank prior was presented in [14] by using low-rank characteristics of similar patches. However,
Proposed method
Our goal is to develop an effective blind image deblurring approach to achieve high-quality results with fine details. In view of the fact that existing methods only focus on the term involving salient edges, our approach is to take advantage of salient edge combined with image structure in the MAP framework. We first briefly review the relevant filtering technology of image structure extraction in this work, followed by modeling as well as problem-solving.
Experimental results
We compare our method with state-of-the-art approaches on synthetic datasets [10], [43], [44], [45] and real-world images. The main evaluation metrics are as follows: Error-Ratio [10], peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and the sum of squared differences error (SSDE) [10].
In all experiments, the four main tunable parameters are set to be , , , . For a fair comparison, after obtaining the blur kernels of the proposed method and all the
Effectiveness of our model
We analyze the effectiveness of our model (6) on the dataset [10]. Fig. 16(a) compares the proposed model with different settings, only salient edge term, without gradient term, without salient edge term, without image structure term, which clearly shows that our model generates the highest success rate. Then, the model of this paper can produce the smallest SSDE [10] as shown in Fig. 16(b). The above analysis demonstrates that our model is better than applying either one alone. Moreover, we
Conclusions
In this paper, we explore a new approach to blind image deblurring. We developed a new structure-based term to guide fine structure for latent images and established an adaptive deblurring model based on the combination of image structure and significant edges, which is more effective than either one alone. In addition, the -norm constraint of the gradient is added to the model to preserve edges and remove useless details. Extensive experiments on synthetic datasets and real-world scenes
CRediT authorship contribution statement
Dandan Hu: Conceptualization, Methodology, Software, Writing – original draft. Jieqing Tan: Conceptualization, Methodology, Data curation, Writing – original draft. Li Zhang: Supervision, Writing – review & editing. Xianyu Ge: Visualization, Software, Investigation. Jing Liu: Writing – review & editing, Validation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
We would like to thank the reviewers for their helpful suggestions which greatly improve the clarity of the paper.
This work is supported by the National Natural Science Foundation of China (62172135); National Key Research and Development Program, China (2018YFB2100301).
References (53)
- et al.
Blind deblurring with sparse representation via external patch priors
Digit. Signal Process.
(2018) - et al.
Blur kernel estimation via salient edges and low rank prior for blind image deblurring
Signal Process., Image Commun.
(2017) - et al.
Kernel estimation from salient structure for robust motion deblurring
Signal Process., Image Commun.
(2013) - et al.
Nonlinear total variation based noise removal algorithms
Physica D
(1992) - et al.
Fast image deconvolution using hyper-Laplacian priors
Adv. Neural Inf. Process. Syst.
(2009) - et al.
Handling outliers in non-blind image deconvolution
- et al.
Learning spatially-variant MAP models for non-blind image deblurring
- et al.
Deep wiener deconvolution: wiener meets deep learning for image deblurring
- et al.
High-quality motion deblurring from a single image
ACM Trans. Graph.
(2008) - et al.
Blind motion deblurring from a single image using sparse approximation
Blind image deblurring using dark channel prior
Image deblurring via extreme channels prior
Blind image deblurring with local maximum gradient prior
Efficient marginal likelihood optimization in blind deconvolution
Edge-based blur kernel estimation using patch priors
Blind deblurring using internal patch recurrence
Image deblurring via enhanced low rank prior
IEEE Trans. Image Process.
PSF Estimation using sharp edge prediction
Blur kernel estimation using the radon transform
Fast motion deblurring
ACM Trans. Graph.
Two-phase kernel estimation for robust motion deblurring
Blur kernel estimation using normalized color-line priors
Surface-aware blind image deblurring
IEEE Trans. Pattern Anal. Mach. Intell.
Feature-oriented image enhancement using shock filters
SIAM J. Numer. Anal.
Signal and image restoration using shock filters and anisotropic diffusion
SIAM J. Numer. Anal.
Image sharpening by morphological filtering
Pattern Recognit.
Cited by (3)
An extended sparse model for blind image deblurring
2024, Signal, Image and Video ProcessingA multiscale network for deblurring coal slurry foam images
2024, Signal, Image and Video ProcessingAn Extended Sparse Model for Blind Image Deburring
2023, Research Square