Adaptive total variation based image segmentation with semi-proximal alternating minimization
Introduction
Image segmentation has always been a hot topic, which aims to divide an image domain into mutually disjoint regions reasonably. This plays a significant role in many application fields, such as medical imaging [1], [2], vehicle license plate recognition [3], [4], etc. During the past decades, many excellent approaches have been proposed [5], [6], [7], [8], [9], [10]. In 1989, one of the most important segmentation models called MS model was proposed by Mumford and Shah [6], which was a landmark achievement in segmentation fields. Let be a bounded open connected set, and be a compact curve in the MS model can be formulated aswhere are positive parameters, is the degraded image and is the ideal image. Here, the length of can be written as i.e., the 1-dimensional Hausdorff measure in . Model (1) adopted the energy minimization equation to find an approximate solution, but it is difficult to solve due to its nonconvexity.
Some early attempts [11], [12] to solve (1) were done by approximating it using a sequence of simpler elliptic problems. Meanwhile, more works have focused on exploring the MS model such as active contour methods [13], [14], graph cut methods [15] and convex relaxation approaches [16], [17], etc. In 2013, Cai et al. proposed a two-stage image segmentation strategy using a convex variant of the MS model and thresholding [18], which got good segmentation results. The first stage is to find a smooth approximation solution by minimizing the energy functional:where is the problem related linear operator. Once is obtained, the segmentation is done by thresholding properly in the second stage. Since then, many novel image segmentation methods [19], [20], [21], [22], [23], [24] have been proposed based on the two-stage idea [18], and they show the subtle connection between image restoration and segmentation to some extent. The two-stage strategy and corresponding improved models have noticeable virtues in image segmentation, refer to [18], [19], [23] for details.
As we know, image restoration (including denoising, deblurring, inpainting, etc.) is to estimate a clean image from the degraded image. And the segmentation can be used as the preprocessing or postprocessing for restoration. In [19], Chan et al. used the two-stage method for segmenting blurry images with Poisson or multiplicative Gamma noise. Duan et al. [20] introduced Euler’s Elastica as the regularization in the MS model based on a two-stage segmentation strategy to capture the geometry of object shapes in noisy images with missing pixels. In [24], the discrete-MS model was proposed, and the proximal alternating linearized minimization algorithm was employed to deal with the objective function to restore a degraded image and extract its contours. In [23], authors explored a linkage between the PCMS model [6] and the ROF model [25], then derived a novel thresholded-ROF (T-ROF) segmentation method to illustrate the advantages of image segmentation through image restoration techniques. However, as the regularizers of the aforementioned methods are isotropic, non-ideal results can be generated, especially when images with complex geometry structures.
Many previous works [18], [25], [26], [27] have considered the total variation (TV) regularization and anisotropic filtering as standard methods for image restoration because of their capability to detect and maintain image edges. However, the anisotropic filtering produces artifacts and the TV regularization causes the staircase effect [25], [28]. In order to overcome these drawbacks, many approaches [28], [29], [30], [31], [32] were proposed by combining the TV regularization and anisotropic filtering, where the isotropic TV norm is replaced by an anisotropic term. For example, Pang et al. proposed an anisotropic TV-based restored model combining a weighted matrix with the gradient operator [30]. A model based on the adaptive weighted TV regularization in [31] was subsequently proposed for image denoising, where the rotation matrix and weight matrix are embedded in TV regularization.
Different from the reweighted algorithm [33], [34], the weighted algorithm [35], [36], TV [37], [38], [39] and the nuclear/-norm regularization [40], we often need that the target model diffuses along the tangent direction of the local features. Generally, the setting for and with the same weight cannot efficiently couple with the local features. To manifest the geometry structures of targets in images adequately, we give different penalties in the -axis and -axis directions according to the gradient information and . Based on the above facts, we propose an adaptive TV-based two-stage segmentation model with the following novelties:
- a.
We present a robust and efficient anisotropic segmentation model, where the regularization combines the adaptive weighted matrix (see details in Section 2.2) and gradient operator ( ) to describe geometry structure details and more edge information of segmentation objects in images. The adaptive weighted matrix has good properties in capturing elongated fine structures, which are useful for fine structure segmentation.
- b.
To deal with the adaptive TV-based convex objective function, an elegant and effective numerical method is designed through the sPADMM algorithm with guaranteed convergence [41], [42]. Furthermore, extensive experiments demonstrate that the proposed algorithm and model outperform advanced methods on kinds of image segmentation problems under various degraded situations such as noise, blur and missing pixels, etc.
The outline of this paper is as follows. In Section 2, we introduce our anisotropic segmentation model and present the properties of the model. In Section 3, the sPADMM algorithm is carefully applied to solve the convex optimization model and the convergence analysis of the algorithm is shown. We list some preparation works of numerical experiments in Section 4. In Section 5, numerical examples are presented to compare different methods including a celebrated U-Net approach. Finally, the conclusions are drawn in Section 6.
Section snippets
The proposed model
This section mainly introduces our proposed method. And we show that our model has a unique solution under mild conditions. Recall that is a bounded open connected set. Let be the function from to . Then are images living on . Throughout the paper, we do not distinguish the linear operator and the corresponding discrete matrix.
Algorithm
This section focuses on our algorithm, where we exhibit the solving process of model (3) and the convergence analysis of the algorithm. In the discrete settings, we assume that the gray image is an matrix, denote the Euclidean space by and set .
Preparation for numerical experiments
To verify the effectiveness and robustness of our method, we compare our method with state-of-the-art approaches (CV [13], SaT [18], BCEEMS [20], ITCV [58], ICTM-LIF [10], T-ROF [23], WBHMS [32], U-Net [1], SLAT [59]) on some natural images. Except for the operation of the U-Net [1] experiment, all other experiments are performed via using MATLAB R2020a and Windows 10(x64) on a PC with Intel Core (TM) i7 10750H CPU 2.60GHz and 16.0GB memory. The U-Net is conducted under PyTorch and Windows
Image segmentation with different levels of noise and blur
In this example, we test our method on images with different noise levels and blur to show the stability and robustness of our model. In Fig. 8(b) and (c), we add Gaussian noise with mean 0 and =0.3, =0.5, respectively. In Fig. 8(d)--(f), we add =fspecial(‘disk’, 5), =fspecial(‘motion’, 15, 45), =fspecial(‘gaussian’, 12, 12) respectively, and then Gaussian noise with mean value 0 and variance =0.01 on all of them, where and means disk blur, motion blur and Gaussian blur.
Conclusions
In this paper, we developed an adaptive TV-based two-stage segmentation model. We introduced an effective adaptive TV regularization, which can smooth out unimportant details in intrinsic regions of targets and preserve characteristic structures such as edges, sharp corners, etc. A suitable weighted matrix can make our method more robust by coupling with the local structure information of the object. The convex minimization problem can be solved by the efficient sPADMM algorithm with guaranteed
CRediT authorship contribution statement
Tingting Wu: Conceptualization, Methodology, Writing - original draft. Xiaoyu Gu: Data curation, Software. Youguo Wang: Formal analysis, Validation. Tieyong Zeng: Project administration, Supervision.
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.
Acknowledgment
The research was supported by the Natural Science Foundation of China (No. 61971234, 11501301, 11671002, 62071248), CUHK start-up and CUHK DAG 4053296, 4053342, 4053405, RGC 14300219, RGC 14302920, NSFC/RGC N_CUHK 415/19, Hunan Provincial Key Laboratory of Mathematical Modeling and Analysis in Engineering (2018MMAEYB03), Postgraduate Research & Practice Innovation Program of Jiangsu Province (Grant No. SJCX20_0229), the “1311 Talent Plan” of NUPT, and the QingLan Project for Colleges and
References (63)
- et al.
A novel character segmentation-reconstruction approach for license plate recognition
Expert Syst. Appl.
(2019) - et al.
Weighted variational model for selective image segmentation with application to medical images
Pattern Recognit.
(2018) - et al.
Weighted variational model for selective image segmentation with application to medical images
Pattern Recognit.
(2018) - et al.
Nonlinear total variation based noise removal algorithms
Physica D
(1992) - et al.
Image denoising via a new anisotropic total-variation-based model
Signal Process.
(2019) - et al.
Image denoising based on the adaptive weighted TV regularization
Signal Process.
(2020) - et al.
A weighted bounded Hessian variational model for image labeling and segmentation
Signal Process.
(2020) - et al.
Variable exponent functionals in image restoration
Appl. Math. Comput.
(2010) - et al.
Image restoration via the adaptive TV regularization
Comput. Math. Appl.
(2020) - et al.
FCM: the fuzzy c-means clustering algorithm
Comput. Geosci.
(1984)
An efficient iterative thresholding method for image segmentation
J. Comput. Phys.
U-Net: Convolutional networks for biomedical image segmentation
International Conference on Medical Image Computing and Computer-Assisted Intervention
A high-order scheme for image segmentation via a modified level-set method
SIAM J. Imaging Sci.
Vehicle license plate character segmentation
Int. J. Autom. Comput.
Snakes: active contour models
Int. J. Comput. Vis.
Optimal approximations by piecewise smooth functions and associated variational problems
Commun. Pure Appl. Math.
A continuous max-flow approach to Potts model
European Conference on Computer Vision
Frame based segmentation for medical images
Commun. Math. Sci.
The iterative convolution-thresholding method (ICTM) for image segmentation
Comput. Vis. Pattern Recognit.
Approximation of functionals depending on jumps by elliptic functionals via -convergence
Commun. Pure Appl. Math.
On the approximation of free discontinuity problems
Bollettino Della Unione Matematica Italiana B
Active contours without edges
IEEE Trans. Image Process.
A multiphase level set framework for image segmentation using the Mumford and Shah model
Int. J. Comput. Vis.
Reformulating and optimizing the Mumford-Shah functional on a graph–a faster, lower energy solution
European Conference on Computer Vision
A convex relaxation approach for computing minimal partitions
2009 IEEE Conference on Computer Vision and Pattern Recognition
A convex approach to minimal partitions
SIAM J. Imaging Sci.
A two-stage image segmentation method using a convex variant of the Mumford–Shah model and thresholding
SIAM J. Imaging Sci.
A two-stage image segmentation method for blurry images with poisson or multiplicative gamma noise
SIAM J. Imaging Sci.
A two-stage image segmentation method using Euler’s elastica regularized Mumford-Shah model
2014 22nd International Conference on Pattern Recognition
Image segmentation via mean curvature regularized Mumford-Shah model and thresholding
Neural Process. Lett.
Linkage between piecewise constant Mumford–Shah model and Rudin–Osher–Fatemi model and its virtue in image segmentation
SIAM J. Sci. Comput.
Cited by (35)
A fractional-order image segmentation model with application to low-contrast and piecewise smooth images[Formula presented]
2024, Computers and Mathematics with ApplicationsImage Segmentation Based on the Hybrid Bias Field Correction
2023, Applied Mathematics and ComputationMulti-phase image segmentation by the Allen–Cahn Chan–Vese model
2023, Computers and Mathematics with ApplicationsGeneral nonconvex total variation and low-rank regularizations: Model, algorithm and applications
2022, Pattern RecognitionLearning multi-level structural information for small organ segmentation
2022, Signal ProcessingAn ℓ<inf>0</inf>-overlapping group sparse total variation for impulse noise image restoration
2022, Signal Processing: Image CommunicationCitation Excerpt :One of the regularization functions proposed was the total variation (TV) norm. It was first introduced by Rudin et al. in [7] which has successfully used to solve denoising [8], deblurring [9], segmentation [10], and superresolution [11], due to its ability to allows sharp recovery that preserves edges in the restored image [12]. However, in the presence of noise in piecewise affine signal regions, TV restoration suffers the staircase artifacts [13,14].