Research article Special Issues

Spatial general autoregressive model-based image interpolation accommodates arbitrary scale factors

  • This paper proposed a novel image interpolation algorithm with an arbitrary upscaling factor based on the spatial general autoregressive model. First, to accommodate arbitrary scale factors, a non-integer mapping method was modulated into the spatial general autoregressive model, which was employed to model the piecewise stationary pattern with a higher description capacity than autoregressive models. A gradient angle guided extension method was utilized to extend the spatial general autoregressive model, and more pixels in the neighborhood were included to estimate the parameters of the spatial general autoregressive model. To realize the high-accuracy estimation of the model parameters, a regularization method via an elastic network was adopted to maintain the complexity of the object function in a reasonable state and address the overfitting problem. We also introduced an iterative curvature method to refine the interpolation result of those image blocks with large variances of gray intensities. Experiments on 25 images were conducted with integer and non-integer magnification factors to systematically verify the objective and subjective measures of the proposed method. The visual artifacts were effectively suppressed by the proposed method, and a flexible interpolation method for arbitrary scale factors was implemented.

    Citation: Yuntao Hu, Fei Hao, Chao Meng, Lili Sun, Dashuai Xu, Tianqi Zhang. Spatial general autoregressive model-based image interpolation accommodates arbitrary scale factors[J]. Mathematical Biosciences and Engineering, 2020, 17(6): 6573-6600. doi: 10.3934/mbe.2020343

    Related Papers:

    [1] Zhuang Zhang, Wenjie Luo . Hierarchical volumetric transformer with comprehensive attention for medical image segmentation. Mathematical Biosciences and Engineering, 2023, 20(2): 3177-3190. doi: 10.3934/mbe.2023149
    [2] Yingying Xu, Songsong Dai, Haifeng Song, Lei Du, Ying Chen . Multi-modal brain MRI images enhancement based on framelet and local weights super-resolution. Mathematical Biosciences and Engineering, 2023, 20(2): 4258-4273. doi: 10.3934/mbe.2023199
    [3] Yue Jiang, Xuehu Yan, Jia Chen, Jingwen Cheng, Jianguo Zhang . Meaningful secret image sharing for JPEG images with arbitrary quality factors. Mathematical Biosciences and Engineering, 2022, 19(11): 11544-11562. doi: 10.3934/mbe.2022538
    [4] Shuaiyu Bu, Yuanyuan Li, Wenting Ren, Guoqiang Liu . ARU-DGAN: A dual generative adversarial network based on attention residual U-Net for magneto-acousto-electrical image denoising. Mathematical Biosciences and Engineering, 2023, 20(11): 19661-19685. doi: 10.3934/mbe.2023871
    [5] Zhicheng Song, Jinbao Chen, Chuanzhi Chen . Application of discrete shape function in absolute nodal coordinate formulation. Mathematical Biosciences and Engineering, 2021, 18(4): 4603-4627. doi: 10.3934/mbe.2021234
    [6] Dan Yang, Shijun Li, Yuyu Zhao, Bin Xu, Wenxu Tian . An EIT image reconstruction method based on DenseNet with multi-scale convolution. Mathematical Biosciences and Engineering, 2023, 20(4): 7633-7660. doi: 10.3934/mbe.2023329
    [7] Xuehu Yan, Xuan Zhou, Yuliang Lu, Jingju Liu, Guozheng Yang . Image inpainting-based behavior image secret sharing. Mathematical Biosciences and Engineering, 2020, 17(4): 2950-2966. doi: 10.3934/mbe.2020166
    [8] Boyang Wang, Wenyu Zhang . MARnet: multi-scale adaptive residual neural network for chest X-ray images recognition of lung diseases. Mathematical Biosciences and Engineering, 2022, 19(1): 331-350. doi: 10.3934/mbe.2022017
    [9] Hua Wang, Weiwei Li, Wei Huang, Jiqiang Niu, Ke Nie . Research on land use classification of hyperspectral images based on multiscale superpixels. Mathematical Biosciences and Engineering, 2020, 17(5): 5099-5119. doi: 10.3934/mbe.2020275
    [10] Hong-an Li, Min Zhang, Zhenhua Yu, Zhanli Li, Na Li . An improved pix2pix model based on Gabor filter for robust color image rendering. Mathematical Biosciences and Engineering, 2022, 19(1): 86-101. doi: 10.3934/mbe.2022004
  • This paper proposed a novel image interpolation algorithm with an arbitrary upscaling factor based on the spatial general autoregressive model. First, to accommodate arbitrary scale factors, a non-integer mapping method was modulated into the spatial general autoregressive model, which was employed to model the piecewise stationary pattern with a higher description capacity than autoregressive models. A gradient angle guided extension method was utilized to extend the spatial general autoregressive model, and more pixels in the neighborhood were included to estimate the parameters of the spatial general autoregressive model. To realize the high-accuracy estimation of the model parameters, a regularization method via an elastic network was adopted to maintain the complexity of the object function in a reasonable state and address the overfitting problem. We also introduced an iterative curvature method to refine the interpolation result of those image blocks with large variances of gray intensities. Experiments on 25 images were conducted with integer and non-integer magnification factors to systematically verify the objective and subjective measures of the proposed method. The visual artifacts were effectively suppressed by the proposed method, and a flexible interpolation method for arbitrary scale factors was implemented.


    With the rapid development of multimedia technologies, the collection and transmission of multimedia data have become greatly convenient and easy. Meanwhile, multimedia data quality of imaging equipment limits the wider application of machine vision inspection system. The method to improve the image resolution through algorithms—namely, image interpolation, has become an active research field in machine vision. To a certain extent, image interpolation can effectively increase the image resolution to meet various applications, such as motion tracking and pose estimation. In addition, image interpolation has a wide range of practical applications and commercial prospects in remote sensing, cloud-of-thing system [1], surveillance [2,3], medical imaging [4], multimedia processing [5,6], and consumer electronic [7].

    In the registration and fusion of non-homologous images, image size adjustment is needed. Especially in the visual measurement process of large machined parts, due to the angle of view, the images captured by different cameras have different dimensions [8]. This will lead to dimensional distortion. In order to reduce visual artifacts and adapt to the requirements of arbitrary magnification, a more flexible image interpolation method is needed. Although the visual quality of simple linear interpolation methods is not sufficient, these methods still are widely used due to their low computational complexity and interpolation flexibility. Linear interpolation methods include nearest-neighbor interpolation [9,10], bilinear interpolation [11,12], bicubic interpolation [11,12,13,14], cubic spine interpolation [15,16] and iterative linear interpolation [17,18], which predict the gray intensity of an unknown pixel according to the distance between the unknown pixel and the reference pixels. The pixels at the boundary, edge or texture of a low-resolution (LR) image have a greater impact on image interpolation than those perpendicular to the edge. Therefore, an isotropic low-pass filter is employed to weaken the high-frequency components of the edges, and the edges are consequently smoothed, which will lead to artifacts such as blurring, ringing, checkerboard effects, edge discontinuities and jagging.

    In related works, contributions have been made in reducing the visual artifacts. However, those methods are mainly suitable for integer magnification of images. Some edge-guided interpolation algorithms have been designed to correct the artifacts stemming from isotropic filters. The explicit methods [19,20] were developed to consider local structural information such as edges or isophotes. However, for images containing many complex structures, an edge map estimated with these methods tends to be unpredictable, which increases the intensity variation of the interpolation. Implicit adaptive methods have been proposed, which embed the local structures into an objective function that can be solved using linear or nonlinear optimization methods. These methods can model the image’s local patterns and estimate the unknown pixels around the edges in texture-rich areas. Li and Orchard [21] proposed an edge adaptive interpolation method according to the geometric duality of edges, named new edge directed interpolation (NEDI). Geometric duality between the low-resolution covariance and the high-resolution covariance, couple the pair of pixels along the same orientation, enables NEDI to estimate the high-resolution covariance from its low-resolution counterpart with a qualitative model characterizing the relationship between the covariance and the resolution. Because the inaccurate estimation of covariance in texture-rich regions, jagging and color infidelity are inescapable on some RGB samples. Analogous to NEDI, other studies have been conducted with geometrical similarity measurement [22,23,24,25,26] to improve interpolation accuracy. Chang and Kevin [27] used collaborative representation and exploiting non-local self-similarity of natural images and introduced the external HR information into the interpolation process. Zhang and Wu [22] proposed soft-decision adaptive interpolation (SAI), which learnt and adapted varying pixel structures by using a piecewise autoregressive (PAR) model in the local rectangular window of the image. To implicitly use AR models for better solving nonlinear problems at the edges in image interpolation, irregular windows were employed by Guo et al. [23,24,25], which were extended adaptively from the root windows according to the geometric features, such as isophotes and curvatures. The similarities between pixels were calculated using patch-geodesic distance. Cheng et al. [28] used Fast Fourier Transformation (FFT) multichannel interpolation to reconstruct or approximate the continuous signals from a series of discrete points. A spatial interpolation can predict the values of the unknown points by processing the surrounding variables with meaningful values within the same region. In our previous work [26], a spatial general autoregressive model (SGAR), which is a uniform expression for both linear and nonlinear AR models, was employed to implement a noise-insensitive and edge-preserving interpolator. Although it still uses regular windows, its nonlinear description ability was improved by introducing nonlinear terms into the model. The similarities between pixels in the image windows were implicitly exploited by the robust parameter estimation method named generalized M-estimator.

    In contrast to the above interpolation strategies, the curvature-based interpolation algorithms introduced high-order derivative information of image intensity to achieve high-accuracy interpolation from the rough results. Giachetti and Asuni [29] proposed the iterative curvature based interpolation (ICBI) method, which generates high-resolution (HR) pixels within the grid along the directions with the lowest second-order derivative and subsequently updates the values of the HR pixels by minimizing the local variations of curvature of the image intensity. However, the hole-filling strategy limits the application with non-integer upscaling factors. Kim and Cha [30] proposed the curvature interpolation method (CIM) based on a partial differential equation (PDE). Because it is known to tend to converge to a piecewise constant image [31,32]. These methods share the characteristic that the final interpolation results are modified by an iterative procedure upon the initial values provided by other methods. It was emphasized that better interpolation accuracy was achieved, which was largely due to the first-step interpolation method.

    Therefore, an arbitrary scale factor image interpolation based on the SGAR model was used to implement image interpolation from three aspects: image window adaptive extension by gradient angle, model regularization by the elastic network, and refinement of the result accuracy by the curvature constraints. The rest of the paper is structured as follows. Section 3 introduces the SGAR model-based image interpolation method and its implementation. Experimental results and a comparison study with some existing popular image interpolation techniques are presented in Section 4. Section 5 contains our conclusions.

    AR models are effective tools for image modeling [22,23,24,25,26,33,34,35]. The linear AR model in vectorization form is shown as follows.

    ˆy=hθ(x)=θTx (1)

    where ˆy is the prediction value, θ=[θ0,θ1,,θn] is the model’s parameter vector, containing the intercept term θ0 and the feature weights θ1 to θn, x is the sample’s feature vector, and hθ is the hypothesis function, using the model parameters θ.

    However, the data distributions of a digital image are more complex. To better fit the image model, the product of the pixels can be involved in the polynomial regression, which is adopted as a new feature, such as the SGAR model that was initially explored based on Weierstrass theory for digital image adaptive filtering in our previous work [36]. The pixels in the image follow a certain regular pattern that indicates the direction of gray scale change. In other words, every unknown interpolated pixel in a piecewise image can be estimated by its known adjacent neighbor with certain weights. The original images are broken up into small fragments. Based on the stationarity assumption in piecewise images [22], we model the fragments as a locally stationary Gaussian process. Inevitably, some fragments with abrupt and unnatural gray scale changes are brought into the modeling process and have negative effect on describing the pattern of the image [36]. An adaptive filter was implemented based on the SGAR model. The new filter removed these artifacts while effectively conserving detailed image information. This was because the SGAR model fuses both linear and nonlinear AR models into a uniform expression[26]. However, the high-order polynomial regression model will probably overfit the training data. Moreover, simple linear regression has the problem of underfitting. For an image window, it could be modeled as

    x(p)=ri=1{s1AsiAϕ(s1,s2,,si)iΠk=1x(sk)}+a(p) (2)

    where x(p) is the forecasted value of a pixel, x(sk) is the reference value of a pixel in the LR image, a(p) is the modeling residuals. s is a two-dimensional vector representing the location of the specific pixel. A is a set of vectors representing the location of adjacent reference pixels. The composition of set A is discussed in section 2.3. ϕ(s1, s2, …, si) is the model parameters. r is the order of the model, which indicates the dimension of vector ϕ. When r is 1, the SGAR model is degenerated into AR model as is shown in Eq (3). The dimension of the vector ϕ is only 8 and consists of the eight adjacent pixels around the anchor pixel.

    x(p)=siAϕ(si)x(si)+a(p) (3)

    When r is 2, the SGAR model can also be rewritten as Eq (4). The dimension of vector ϕ is 44 and consists of 8 adjacent pixels around the anchor pixel, 36 self-multiplication and product of two different pixels are taken into consideration. We will explain how to obtain the parameters of the SGAR model in section 2.4.

    x(p)=siAϕ(si)x(si)+saAsbAϕ(sa,sb)x(sa)x(sb)+a(p) (4)

    The relationship between the coordinates of the pixels of the HR image and those of the corresponding LR image must first be clarified.

    {u=i×1hv=j×1h (5)

    where (u, v) represents a pixel in the LR image, (i, j) represents a pixel in the HR image, and h is an arbitrary scaling factor.

    The reference pixels in the LR image of a reconstructed pixel in the HR image is determined with the following mathematical expression.

    {u1=φ(u)v1=φ(v),{u2=ρ(u)v2=ρ(v),{u3=u1+1v3=v1+1 (6)

    where φ(·) is a function that tends toward negative infinity, and ρ(·) is a function that rounds to the nearest decimal or integer.

    When r is 1, the product of the LR pixels and SGAR model parameter vector ϕ is regard as the value of interpolated pixel. For the convenience of understanding, taking 1.25× interpolation as an example, the coordinates of HR pixels are (63, 22), (63, 23), (64, 22), (64, 23), the LR reference pixels coordinate at (52, 19), (52, 20), (53, 19) and (53, 20) are used to interpolate the HR pixel at (63, 22), other HR pixels interpolation schemes are shown in Figure 1:

    Figure 1.  Pixel arrangement in interpolation process.

    When r is 2, the parameters of SGAR model are estimated by modified samples. For each sample, it consist of an anchor pixel at the center of 3 × 3 image block, 8 LR reference pixels and 36 parameters generated by these pixels. Then the LR pixels used in the interpolation process share the same arrangement with the samples used in parameter estimation process.

    The gradient guided method [20,37] was employed to adaptively extend the parameters estimation window to increase the training dataset and ultimately improve the interpolation accuracy. The gradient angles were calculated by the Scharr operator, which is formulated as follows.

    Scharrx=[30310010303],Scharry=(Scharrx)=[31030003103] (7)

    where Gx and Gy are the gradients of the x and y directions, respectively, according to the Scharr operator.

    {Gx=[3f(x1,y+1)+10f(x,y+1)+3f(x+1,y+1)][3f(x1,y1)+10f(x,y1)+3f(x+1,y1)]Gy=[3f(x1,y1)+10f(x1,y)+3f(x1,y+1)][3f(x+1,y1)+10f(x+1,y)+3f(x+1,y+1)] (8)

    Therefore, the gradient angle is

    θ=arctan(GyGx)180π (9)

    The gradient angle space was divided into eight regions according to an interval of 45°, and the eight regions were placed into four groups.

    {[22.5,22.5)[157.5,202.5)[22.5,67.5)[202.5,247.5)[67.5,112.5)[247.5,292.5)[112.5,157.5)[292.5,337.5) (10)

    According to the gradient direction of each pixel in the LR image, the image windows involved in modeling the image with the SGAR model would be extended in four directions as shown in Figure 2.

    Figure 2.  Adaptive window extension method based on gradient guidance. (a) Image windows extended horizontally. (b) Image windows extended in the direction of 45°. (c) Image windows extended vertically. (d) Image windows extended in the direction of 135°.

    The gradient directions of the n pixels were grouped according to Eq (10), and the frequency of each group was counted. The extension direction was subsequently determined according to the maximum of the four frequencies. The number of fitting samples is determined by the gradient based extension method. In fact, 15 samples were used in the horizontal and vertical directions, and 19 samples are used in the other two directions.

    The earlier works suggested that the patterns of the AR models containing only a linear relationship between pixels are sufficient. The least-square (LS) method [18], which is highly sensitive to outliers, was used to solve the parameters of the interpolation model with cross-direction constraints. To address the noise-sensitive problem of the ordinary LS, some methods were proposed [38]. Among them, regularization is a common method. The l1-norm and l2-norm regularization terms were added to the objective function by Liu et al. to enhance the stability of the LS solution. The l2-norm penalty term [27] (i.e., ridge regression) was used to regularize the objective function. Weighted ridge regression (WRR) [19] was adopted to restrain the expansion of variance and thus achieve more reliable estimations by modulating weights into the regression to evaluate the reliability of each sample. In contrast to the above methods, the products of the pixels were employed to increase the description capacity of the SGAR model to suit various digital images, and the overfitting problem was noted.

    The SGAR model adopted the products of pixels as the new features, so the descriptive power of the model was improved compared with some interpolation methods that consider only simple linear relations [15]. Therefore, the model was especially effective for rebuilding the local nonlinear relationship of image windows [22,23,26,35]. However, it may cause overfitting problems because of the higher degree of freedom [26]. To address this problem, the regularization method could be used to reduce the freedom of models and keep the complexity of the object function in a reasonable state. Ridge regression [23,39], Lasso regression [40] and Elastic Network [41] are the three most commonly used methods.

    Therefore, the mean square error (MSE) was calculated to represent the intensity differences between the reference pixels and the predicted pixels. The parameters of the SGAR model will obtained by minimizing the loss function. The loss function of Elastic Network is shown as:

    L(X,θ,α,t)=1mmi=1(θTX(i)y(i))2+tαnj=1|θj|+1t2αnj=1θ2j (11)

    where m is the number of the samples in the dataset, n is the number of fit parameters, X(i) is the vector combined with all features of sample i, θ is the vector of the model parameter, y(i) is the anchor pixel of the region, and is the original value of LR pixel at the center of 3 × 3 image block, α is the hyper-parameter that dominates the regularization level of the model, and t is the hyper-parameter that dominates the mixing ratio. In the experiment, the value of t was manually set to 0.5.

    The SGAR model learns the structural relationship between pixels and their neighbors in the LR image to recover as much missing information as possible during down-sampling [42]. However, we know that it is very difficult to recover HR images in high-frequency areas where the gray level of the image changes rapidly, such as textures and gradients. Only a relatively small part of the missing information can be restored. Compared with low-frequency areas, modeling of image windows with the SGAR model in these areas is more likely to encounter the problem of insufficient generalization ability, which makes interpolation stability difficult to control. Therefore, in these regions, a strategy that updates the interpolated pixels by minimizing the local variation of the second derivative of the image was adopted. By transforming the two-dimensional image plane into three-dimensional space, the surface optimization methods [43] were employed to solve the problem of artifacts.


     | Show Table
    DownLoad: CSV

    An analogous filtering-based method proposed by Gong et al. [43] was explored to reduce the regularization part of the variational energy while guaranteeing non-increasing total energy. That is, the iterative curvature method was modulated into SGAR mode. The second-order derivatives of p (i, j) on the HR image were calculated in eight different directions, as shown in Figure 3, and were used as an approximation of the curvatures. The eight second-order derivatives were thresholded, and a set consisting of the thresholded second-order derivatives was obtained. P (i, j) was updated according to the absolute minimum of the elements of the set. The second-order derivatives of the digital image window were calculated repeatedly until the curvature of p (i, j) in all directions was below the threshold or the maximum iterations were reached. The curvatures around p (i, j) gradually adapted to the directions that shared a similar curvature change tendency. Thus, the accuracy of the interpolation results obtained in the first step was improved with the following algorithm process in List 1.

    Figure 3.  Curvature descriptions of 3 × 3 region in 8 different directions.

    Experiments were carried out on a dataset provided by ICBI [29] with different magnification factors to systematically verify the performance of the proposed method in terms of objective measurements, subjective visual effect and computational cost. For thoroughness and fairness, the chosen dataset contains images with different resolutions and various objects—animals, flowers and buildings—with a wide range of color and natural textures, which include typical and unique digital images for interpolation research.

    Comparisons of non-integer factors are studied with bicubic interpolation, bilinear interpolation, nearest neighbor interpolation and AREA interpolation. These methods are well-established in Open Source Computer Vision Library (OpenCV). The mean square error (MSE), peak signal to noise ratio (PSNR), structural similarity index (SSIM), feather similarity index (FSIM) and Correlation coefficient (CC) were used to evaluate the result of the interpolation algorithm. The experiments are conducted in non-integer interpolation factors to substantiate the performance of proposed method in arbitrary scaling. In 1.7× and 2.2× enlargement, images with 512 × 512 pixels are resized into 435 × 435 pixels and 563 × 563 pixels with bicubic method, and used as the reference for computing the MSE, PSNR and SSIM index. The images with 256 × 256 pixels are enlarged to the same size by proposed method and other four conventional non-integer interpolation methods. The experimental results are tabulated in Tables 1 and 2.

    Table 1.  Comparison of 1.7× magnification interpolation results with four non-integer methods.
    Image Scale Criterion AREA Nearest Bicubic Bilinear Proposed
    Zebra 1.7× MSE 81.207 228.335 75.005 77.776 59.349
    PSNR 29.035 24.545 29.380 29.222 30.397
    SSIM 0.924 0.864 0.929 0.920 0.926
    FSIM 0.9817 0.9512 0.9791 0.9772 0.9877
    CC 0.9935 0.9810 0.9939 0.9939 0.9952
    Bench 1.7× MSE 104.015 231.486 91.697 107.383 79.586
    PSNR 27.960 24.486 28.507 27.821 29.122
    SSIM 0.920 0.864 0.928 0.913 0.925
    FSIM 0.9760 0.9499 0.9761 0.9705 0.9847
    CC 0.9891 0.9749 0.9903 0.9889 0.9916
    Bird 1.7× MSE 71.808 156.785 61.920 74.543 63.665
    PSNR 29.569 26.178 30.212 29.407 30.092
    SSIM 0.914 0.847 0.920 0.907 0.916
    FSIM 0.9802 0.9511 0.9814 0.9753 0.9868
    CC 0.9921 0.9824 0.9932 0.9919 0.9929
    Clock 1.7× MSE 44.927 104.802 42.651 45.213 36.236
    PSNR 31.606 27.927 31.831 31.578 32.539
    SSIM 0.961 0.932 0.964 0.959 0.962
    FSIM 0.9875 0.9693 0.9862 0.9841 0.9921
    CC 0.9950 0.9880 0.9952 0.9950 0.9959
    Butterfly 1 1.7× MSE 97.058 249.318 87.326 95.717 74.768
    PSNR 28.260 24.163 28.719 28.321 29.394
    SSIM 0.939 0.875 0.946 0.936 0.947
    FSIM 0.9849 0.9600 0.9836 0.9806 0.9904
    CC 0.9929 0.9813 0.9936 0.9932 0.9945
    Bee 1.7× MSE 37.375 106.105 29.757 36.878 26.581
    PSNR 32.405 27.873 33.395 32.463 33.885
    SSIM 0.964 0.916 0.970 0.964 0.970
    FSIM 0.9916 0.9696 0.9907 0.9895 0.9952
    CC 0.9961 0.9886 0.9968 0.9962 0.9972
    Carrousel 1.7× MSE 77.964 183.253 72.883 78.882 62.649
    PSNR 29.212 25.500 29.505 29.161 30.162
    SSIM 0.952 0.913 0.955 0.949 0.953
    FSIM 0.9823 0.9592 0.9809 0.9786 0.9882
    CC 0.9948 0.9877 0.9951 0.9948 0.9958
    Sunflower 1.7× MSE 27.130 75.759 23.167 26.798 18.957
    PSNR 33.796 29.336 34.482 33.850 35.353
    SSIM 0.963 0.932 0.967 0.960 0.964
    FSIM 0.9907 0.9763 0.9906 0.9883 0.9945
    CC 0.9980 0.9945 0.9983 0.9981 0.9986
    Puppet 1.7× MSE 13.923 34.978 13.314 13.920 10.774
    PSNR 36.694 32.693 36.888 36.694 37.807
    SSIM 0.967 0.938 0.970 0.966 0.970
    FSIM 0.9932 0.9813 0.9913 0.9915 0.9958
    CC 0.9975 0.9937 0.9976 0.9975 0.9981
    Eagle 1.7× MSE 38.275 83.231 39.204 40.307 33.764
    PSNR 32.302 28.928 32.197 32.077 32.846
    SSIM 0.945 0.905 0.946 0.942 0.945
    FSIM 0.9907 0.9746 0.9880 0.9885 0.9932
    CC 0.9958 0.9908 0.9957 0.9956 0.9963
    Sheep 1.7× MSE 34.092 73.470 31.319 35.503 32.320
    PSNR 32.804 29.470 33.173 32.628 33.036
    SSIM 0.939 0.885 0.947 0.934 0.941
    FSIM 0.9909 0.9721 0.9903 0.9885 0.9941
    CC 0.9963 0.9919 0.9966 0.9961 0.9965
    Giraffe 1.7× MSE 20.370 62.295 18.833 18.987 14.790
    PSNR 35.041 30.186 35.382 35.346 36.431
    SSIM 0.969 0.929 0.973 0.969 0.972
    FSIM 0.9927 0.9719 0.9905 0.9908 0.9957
    CC 0.9977 0.9930 0.9979 0.9979 0.9983
    Tiger 1.7× MSE 102.932 201.494 101.077 107.526 102.251
    PSNR 28.005 25.088 28.084 27.816 28.034
    SSIM 0.878 0.806 0.883 0.867 0.874
    FSIM 0.9852 0.9578 0.9816 0.9820 0.9892
    CC 0.9851 0.9702 0.9853 0.9846 0.9851
    Cat 1.7× MSE 33.662 62.693 31.593 35.278 34.880
    PSNR 32.859 30.159 33.135 32.656 32.705
    SSIM 0.917 0.867 0.923 0.910 0.912
    FSIM 0.9886 0.9687 0.9883 0.9861 0.9914
    CC 0.9851 0.9702 0.9853 0.9846 0.9851
    Guitar 1.7× MSE 84.518 158.135 68.741 88.767 66.554
    PSNR 28.861 26.141 29.759 28.648 29.899
    SSIM 0.932 0.896 0.941 0.925 0.938
    FSIM 0.9833 0.9659 0.9854 0.9784 0.9914
    CC 0.9927 0.9863 0.9931 0.9924 0.9924
    Dragonfly 1.7× MSE 22.511 52.164 20.837 22.784 18.960
    PSNR 34.607 30.957 34.942 34.554 35.353
    SSIM 0.968 0.942 0.971 0.967 0.969
    FSIM 0.9931 0.9804 0.9917 0.9913 0.9957
    CC 0.9924 0.9877 0.9947 0.9932 0.9949
    Church 1.7× MSE 62.413 139.430 57.258 64.413 55.518
    PSNR 30.178 26.687 30.552 30.041 30.687
    SSIM 0.935 0.879 0.940 0.931 0.938
    FSIM 0.9784 0.9458 0.9759 0.9729 0.9848
    CC 0.9970 0.9930 0.9972 0.9970 0.9975
    Tower 1.7× MSE 37.601 79.741 35.826 39.162 35.180
    PSNR 32.379 29.114 32.589 32.202 32.668
    SSIM 0.943 0.906 0.946 0.938 0.940
    FSIM 0.9793 0.9497 0.9765 0.9750 0.9844
    CC 0.9927 0.9834 0.9933 0.9925 0.9935
    Butterfly 2 1.7× MSE 66.696 178.208 66.707 66.773 48.798
    PSNR 29.890 25.622 29.889 29.885 31.247
    SSIM 0.961 0.919 0.965 0.959 0.966
    FSIM 0.9855 0.9561 0.9802 0.9818 0.9910
    CC 0.9953 0.9900 0.9955 0.9951 0.9956
    House 1.7× MSE 74.397 146.167 68.578 78.266 74.777
    PSNR 29.415 26.482 29.769 29.195 29.393
    SSIM 0.913 0.866 0.919 0.905 0.904
    FSIM 0.9757 0.9447 0.9746 0.9695 0.9811
    CC 0.9908 0.9745 0.9906 0.9910 0.9932
    Lion 1.7× MSE 61.003 116.098 58.801 64.213 61.114
    PSNR 30.277 27.483 30.437 30.055 30.269
    SSIM 0.884 0.812 0.889 0.872 0.880
    FSIM 0.9849 0.9613 0.9835 0.9809 0.9894
    CC 0.9892 0.9785 0.9900 0.9888 0.9891
    Stained Glass 1.7× MSE 245.082 492.326 221.710 254.356 226.619
    PSNR 24.238 21.208 24.673 24.076 24.578
    SSIM 0.871 0.797 0.879 0.858 0.867
    FSIM 0.9570 0.9198 0.9591 0.9467 0.9726
    CC 0.9907 0.9821 0.9910 0.9903 0.9907
    Colorful 1.7× MSE 66.784 150.067 63.122 67.522 56.865
    PSNR 29.884 26.368 30.129 29.836 30.582
    SSIM 0.923 0.873 0.927 0.919 0.922
    FSIM 0.9880 0.9702 0.9869 0.9853 0.9917
    CC 0.9809 0.9602 0.9826 0.9805 0.9821
    Newspaper 1.7× MSE 89.945 181.939 82.357 94.110 79.896
    PSNR 28.591 25.532 28.974 28.394 29.106
    SSIM 0.917 0.868 0.924 0.911 0.920
    FSIM 0.9827 0.9579 0.9804 0.9782 0.9885
    CC 0.9881 0.9754 0.9891 0.9877 0.9894
    Wheel 1.7× MSE 176.331 335.015 180.833 187.125 191.301
    PSNR 25.668 22.880 25.558 25.409 25.314
    SSIM 0.827 0.738 0.826 0.807 0.802
    FSIM 0.9701 0.9253 0.9611 0.9648 0.9743
    CC 0.9908 0.9745 0.9906 0.9910 0.9932

     | Show Table
    DownLoad: CSV
    Table 2.  Comparison of 2.2× magnification interpolation results with four non-integer methods.
    Image Scale Criterion AREA Nearest Bicubic Bilinear Proposed
    Zebra 2.2× MSE 88.816 206.365 50.628 76.558 46.788
    PSNR 28.646 24.984 31.087 29.291 31.429
    SSIM 0.903 0.848 0.922 0.902 0.925
    FSIM 0.8965 0.8323 0.8975 0.8956 0.8975
    CC 0.94262 0.97264 0.97502 0.97506 0.97513
    Bench 2.2× MSE 109.459 208.750 74.247 106.305 64.341
    PSNR 27.738 24.935 29.424 27.865 30.046
    SSIM 0.906 0.858 0.923 0.901 0.928
    FSIM 0.8993 0.8463 0.9026 0.8963 0.9030
    CC 0.93808 0.96811 0.97060 0.97168 0.97178
    Bird 2.2× MSE 75.389 142.926 53.552 74.440 49.275
    PSNR 29.358 26.580 30.843 29.413 31.205
    SSIM 0.898 0.839 0.921 0.893 0.926
    FSIM 0.8986 0.8398 0.9022 0.8959 0.9022
    CC 0.95257 0.97605 0.97802 0.97807 0.97807
    Clock 2.2× MSE 47.199 95.372 31.825 44.876 28.044
    PSNR 31.391 28.337 33.103 31.611 33.652
    SSIM 0.956 0.931 0.965 0.955 0.966
    FSIM 0.9432 0.9102 0.9449 0.9420 0.9452
    CC 0.98293 0.98430 0.98424 0.98428 0.94669
    Butterfly 1 2.2× MSE 103.625 228.543 61.539 94.279 56.611
    PSNR 27.976 24.541 30.239 28.387 30.602
    SSIM 0.924 0.866 0.949 0.924 0.951
    FSIM 0.9154 0.8609 0.9174 0.9142 0.9174
    CC 0.97389 0.97616 0.97630 0.97641 0.96715
    Bee 2.2× MSE 40.711 93.276 22.502 36.456 20.329
    PSNR 32.034 28.433 34.609 32.513 35.050
    SSIM 0.954 0.912 0.972 0.958 0.973
    FSIM 0.9389 0.8971 0.9398 0.9385 0.9397
    CC 0.98425 0.98562 0.98565 0.98569 0.98569
    Carrousel 2.2× MSE 81.653 166.234 55.864 77.604 49.037
    PSNR 29.011 25.924 30.660 29.232 31.226
    SSIM 0.943 0.909 0.954 0.942 0.956
    FSIM 0.9189 0.8715 0.9202 0.9176 0.9200
    CC 0.96606 0.98261 0.98406 0.98396 0.98402
    Sunflower 2.2× MSE 29.199 66.942 16.333 26.295 14.375
    PSNR 33.477 29.874 36.000 33.932 36.555
    SSIM 0.958 0.932 0.967 0.957 0.968
    FSIM 0.9525 0.9226 0.9530 0.9519 0.9530
    CC 0.98441 0.99236 0.99301 0.99304 0.99305
    Puppet 2.2× MSE 14.994 31.026 9.121 13.794 8.177
    PSNR 36.372 33.214 38.530 36.734 39.005
    SSIM 0.962 0.936 0.972 0.962 0.973
    FSIM 0.9582 0.9342 0.9593 0.9573 0.9591
    CC 0.98381 0.99165 0.99240 0.99247 0.99247
    Eagle 2.2× MSE 38.812 73.190 30.466 39.360 26.828
    PSNR 32.241 29.486 33.293 32.180 33.845
    SSIM 0.939 0.905 0.947 0.936 0.949
    FSIM 0.9490 0.9227 0.9496 0.9487 0.9494
    CC 0.97598 0.98676 0.98786 0.98733 0.98732
    Sheep 2.2× MSE 35.217 65.627 26.054 35.116 24.500
    PSNR 32.663 29.960 33.972 32.676 34.239
    SSIM 0.929 0.883 0.948 0.926 0.950
    FSIM 0.9389 0.9011 0.9415 0.9372 0.9410
    CC 0.97840 0.98865 0.98950 0.98944 0.98945
    Giraffe 2.2× MSE 22.205 53.935 12.282 18.577 11.382
    PSNR 34.666 30.812 37.238 35.441 37.569
    SSIM 0.959 0.923 0.972 0.963 0.973
    FSIM 0.9408 0.8983 0.9416 0.9405 0.9415
    CC 0.97902 0.98989 0.99078 0.99074 0.99079
    Tiger 2.2× MSE 104.056 182.616 89.693 105.770 82.569
    PSNR 27.958 25.515 28.603 27.887 28.963
    SSIM 0.858 0.793 0.874 0.846 0.880
    FSIM 0.9147 0.8711 0.9171 0.9134 0.9167
    CC 0.92847 0.95987 0.96311 0.96260 0.96287
    Cat 2.2× MSE 34.597 57.043 30.381 35.208 28.759
    PSNR 32.740 30.569 33.305 32.664 33.543
    SSIM 0.904 0.861 0.916 0.898 0.918
    FSIM 0.9368 0.9007 0.9391 0.9355 0.9385
    CC 0.96745 0.98197 0.98320 0.98325 0.98326
    Guitar 2.2× MSE 87.945 147.489 62.059 88.695 52.371
    PSNR 28.689 26.443 30.203 28.652 30.940
    SSIM 0.923 0.891 0.939 0.917 0.944
    FSIM 0.9292 0.8952 0.9369 0.9246 0.9376
    CC 0.97516 0.98627 0.98703 0.98839 0.98857
    Dragonfly 2.2× MSE 23.496 46.874 15.687 22.488 14.230
    PSNR 34.421 31.421 36.175 34.611 36.599
    SSIM 0.963 0.940 0.971 0.963 0.972
    FSIM 0.9635 0.9400 0.9648 0.9627 0.9646
    CC 0.98166 0.99039 0.99118 0.99121 0.99124
    Church 2.2× MSE 64.442 126.812 46.051 63.768 42.204
    PSNR 30.039 27.099 31.498 30.085 31.877
    SSIM 0.925 0.878 0.942 0.922 0.945
    FSIM 0.8850 0.8210 0.8870 0.8844 0.8864
    CC 0.95441 0.97650 0.97845 0.97832 0.97838
    Tower 2.2× MSE 38.860 71.499 29.564 38.900 26.737
    PSNR 32.236 29.588 33.423 32.231 33.860
    SSIM 0.938 0.907 0.946 0.932 0.948
    FSIM 0.8976 0.8486 0.8989 0.8968 0.8985
    CC 0.97411 0.98649 0.98745 0.98758 0.98760
    Butterfly 2 2.2× MSE 70.169 157.232 41.024 65.577 36.018
    PSNR 29.669 26.165 32.000 29.963 32.566
    SSIM 0.951 0.914 0.967 0.952 0.969
    FSIM 0.9107 0.8601 0.9119 0.9095 0.9119
    CC 0.92690 0.96274 0.96583 0.96551 0.96565
    House 2.2× MSE 101.537 130.791 61.677 77.698 57.183
    PSNR 28.065 26.965 30.230 29.227 30.558
    SSIM 0.880 0.861 0.912 0.893 0.913
    FSIM 0.8872 0.8298 0.8904 0.8857 0.8897
    CC 0.94697 0.97134 0.97348 0.97367 0.97358
    Lion 2.2× MSE 62.116 104.635 51.312 63.442 47.896
    PSNR 30.199 27.934 31.029 30.107 31.328
    SSIM 0.866 0.803 0.887 0.854 0.891
    FSIM 0.9146 0.8718 0.9184 0.9121 0.9176
    CC 0.95686 0.97589 0.97759 0.97748 0.97759
    Stained Glass 2.2× MSE 253.151 445.729 198.079 252.215 181.234
    PSNR 24.097 21.640 25.162 24.113 25.548
    SSIM 0.851 0.789 0.871 0.837 0.875
    FSIM 0.8532 0.7783 0.8610 0.8473 0.8601
    CC 0.90247 0.94826 0.95220 0.95272 0.95295
    Colorful 2.2× MSE 70.714 137.821 50.291 67.007 46.159
    PSNR 29.636 26.738 31.116 29.870 31.488
    SSIM 0.906 0.861 0.919 0.904 0.922
    FSIM 0.9347 0.8934 0.9360 0.9339 0.9356
    CC 0.96315 0.98112 0.98265 0.98278 0.98282
    Newspaper 2.2× MSE 92.886 166.308 71.003 93.358 63.312
    PSNR 28.451 25.922 29.618 28.429 30.116
    SSIM 0.904 0.859 0.919 0.899 0.924
    FSIM 0.9132 0.8684 0.9161 0.9105 0.9165
    CC 0.93967 0.96737 0.96981 0.97002 0.97009
    Wheel 2.2× MSE 180.859 297.585 158.770 188.328 148.868
    PSNR 25.557 23.395 26.123 25.382 26.403
    SSIM 0.799 0.725 0.815 0.776 0.817
    FSIM 0.8580 0.8014 0.8615 0.8559 0.8607
    CC 0.81563 0.89332 0.90193 0.90154 0.90170

     | Show Table
    DownLoad: CSV

    In 1.7× interpolation experiment of 25 images, our method obtains best MES index in 17 of 25 images, best PSNR index among 18 of 25 interpolated HR images, best SSIM index in 4 of 25 images, best FSIM index in 19 of 25 images and best CC index in 17 of 25 images. In 2.2× interpolation experiment, proposed method achieved best performance over MSE, PSNR and SSIM index in all of 25 images, best FSIM index in 9 of 25 images and best CC index in 15 of 25 images.

    In comparison with other edge-directed methods NEDI [21], improved NEDI (i-NEDI) [34] and ICBI [29] are invited to evaluate the result of the interpolation algorithm. All of 25 images with 512 × 512 pixels are the references for computing the MSE, PSNR and SSIM, the 25 images with 256 × 256 pixels are enlarged by propose method and other interpolation methods. In order to prevent the pixel-shift in ICBI and our proposed method, the left-top 511 × 511 part of original images (and the results of other methods) is used to compute the MSE, PSNR and SSIM. The experimental results are tabulated in Table 3.

    Table 3.  Comparison of 2× magnification interpolation results with four excellent methods.
    Image Scale Criterion NEDI Bicubic i-NEDI ICBI SGAR Proposed
    Zebra MSE 151.503 148.421 152.44 144.954 146.616 145.153
    PSNR 26.327 26.416 26.3 26.518 26.469 26.513
    SSIM 0.863 0.869 0.866 0.872 0.871 0.872
    FSIM 0.9547 0.9510 0.9512 0.9523 0.9522 0.9519
    CC 0.9876 0.9880 0.9876 0.9874 0.9881 0.9878
    Bench MSE 161.811 157.186 147.207 145.943 147.807 146.195
    PSNR 26.041 26.167 26.452 26.489 26.434 26.481
    SSIM 0.869 0.876 0.883 0.883 0.882 0.883
    FSIM 0.9533 0.9510 0.9526 0.9531 0.9532 0.9525
    CC 0.9880 0.9843 0.9827 0.9842 0.9847 0.9832
    Bird MSE 117.786 114.073 116.195 112.192 112.748 110.911
    PSNR 27.42 27.559 27.479 27.631 27.610 27.681
    SSIM 0.857 0.867 0.867 0.873 0.873 0.873
    FSIM 0.9521 0.9533 0.9529 0.9532 0.9534 0.9530
    CC 0.9843 0.9874 0.9868 0.9869 0.9874 0.9872
    Clock MSE 80.781 77.446 75.958 74.529 75.464 74.407
    PSNR 29.058 29.241 29.325 29.408 29.353 29.415
    SSIM 0.936 0.938 0.942 0.941 0.940 0.940
    FSIM 0.9721 0.9703 0.9719 0.9713 0.9712 0.9709
    CC 0.9875 0.9915 0.9908 0.9914 0.9919 0.9912
    Butterfly 1 MSE 168.61 162.884 172.659 159.534 161.906 157.403
    PSNR 25.862 26.012 25.759 26.102 26.038 26.161
    SSIM 0.884 0.891 0.886 0.893 0.891 0.895
    FSIM 0.9622 0.9604 0.9611 0.9615 0.9608 0.9609
    CC 0.9915 0.9880 0.9873 0.9870 0.9881 0.9877
    Bee MSE 61.689 60.26 63.025 57.927 58.620 56.663
    PSNR 30.229 30.331 30.136 30.502 30.450 30.598
    SSIM 0.907 0.925 0.908 0.913 0.911 0.929
    FSIM 0.9733 0.9719 0.9711 0.9723 0.9719 0.9719
    CC 0.9881 0.9937 0.9932 0.9932 0.9932 0.9938
    Carrousel MSE 130.863 126.369 127.243 123.344 124.221 120.354
    PSNR 26.963 27.114 27.084 27.22 27.189 27.326
    SSIM 0.913 0.920 0.920 0.920 0.919 0.924
    FSIM 0.9618 0.9628 0.9635 0.9628 0.9634 0.9634
    CC 0.9911 0.9915 0.9914 0.9916 0.9916 0.9918
    Sunflower MSE 50.277 42.696 44.174 41.959 42.132 38.675
    PSNR 31.117 31.827 31.679 31.903 31.885 32.256
    SSIM 0.916 0.934 0.918 0.919 0.918 0.940
    FSIM 0.9788 0.9775 0.9782 0.9785 0.9782 0.9781
    CC 0.9963 0.9968 0.9968 0.9969 0.9968 0.9971
    Puppet MSE 23.822 21.949 21.833 20.907 21.189 21.05
    PSNR 34.361 34.717 34.74 34.928 34.870 34.898
    SSIM 0.944 0.949 0.948 0.950 0.949 0.951
    FSIM 0.9813 0.9806 0.9805 0.9808 0.9810 0.9805
    CC 0.9958 0.9961 0.9961 0.9963 0.9963 0.9963
    Eagle MSE 67.783 65.684 64.984 64.385 65.075 64.216
    PSNR 29.82 29.956 30.003 30.043 29.997 30.054
    SSIM 0.905 0.909 0.910 0.911 0.910 0.911
    FSIM 0.9758 0.9756 0.9753 0.9755 0.9757 0.9753
    CC 0.9926 0.9928 0.9929 0.9929 0.9929 0.9929
    Sheep MSE 77.659 53.97 57.411 52.907 53.019 52.839
    PSNR 29.229 30.809 30.541 30.896 30.886 30.901
    SSIM 0.887 0.898 0.892 0.900 0.900 0.902
    FSIM 0.9711 0.9742 0.9725 0.9745 0.9746 0.9743
    CC 0.9914 0.9940 0.9936 0.9941 0.9939 0.9941
    Giraffe MSE 39.514 37.885 41.076 37.192 37.490 37.067
    PSNR 32.163 32.346 31.995 32.426 32.392 32.441
    SSIM 0.938 0.943 0.938 0.942 0.942 0.944
    FSIM 0.9749 0.9752 0.9740 0.9754 0.9754 0.9752
    CC 0.9955 0.9957 0.9954 0.9958 0.9958 0.9958
    Tiger MSE 183.369 171.484 173.456 166.573 167.652 167.954
    PSNR 25.498 25.789 25.739 25.915 25.887 25.879
    SSIM 0.782 0.801 0.799 0.808 0.806 0.808
    FSIM 0.9605 0.9616 0.9596 0.9616 0.9617 0.9614
    CC 0.9732 0.9749 0.9746 0.9756 0.9760 0.9754
    Cat MSE 59.565 56.445 57.088 55.344 55.440 55.431
    PSNR 30.381 30.615 30.565 30.7 30.693 30.693
    SSIM 0.852 0.864 0.861 0.866 0.866 0.868
    FSIM 0.9704 0.9717 0.9702 0.9717 0.9720 0.9715
    CC 0.9870 0.9877 0.9875 0.9879 0.9883 0.9879
    Guitar MSE 123.978 108.122 107.388 100.344 101.682 98.526
    PSNR 27.197 27.792 27.821 28.116 28.058 28.195
    SSIM 0.883 0.897 0.895 0.898 0.897 0.906
    FSIM 0.9627 0.9662 0.9672 0.9691 0.9686 0.9686
    CC 0.9904 0.9916 0.9917 0.9922 0.9925 0.9924
    Dragonfly MSE 33.055 31.859 32.668 30.377 30.714 30.02
    PSNR 32.938 33.098 32.99 33.305 33.257 33.357
    SSIM 0.950 0.953 0.951 0.953 0.952 0.955
    FSIM 0.9828 0.9827 0.9825 0.9832 0.9833 0.9828
    CC 0.9955 0.9957 0.9956 0.9959 0.9954 0.9959
    Church MSE 107.069 103.032 102.779 99.755 100.272 100.064
    PSNR 27.834 28.001 28.012 28.141 28.119 28.128
    SSIM 0.886 0.890 0.895 0.897 0.896 0.895
    FSIM 0.9471 0.9446 0.9445 0.9449 0.9446 0.9446
    CC 0.9873 0.9878 0.9878 0.9882 0.9881 0.9881
    Tower MSE 65.253 62.152 60.207 59.533 60.026 59.289
    PSNR 29.985 30.196 30.334 30.383 30.347 30.401
    SSIM 0.899 0.903 0.907 0.907 0.906 0.908
    FSIM 0.9490 0.9487 0.9481 0.9479 0.9479 0.9475
    CC 0.9918 0.9922 0.9924 0.9925 0.9928 0.9925
    Butterfly 2 MSE 149.758 110.03 112.736 106.043 108.439 105.334
    PSNR 26.377 27.716 27.61 27.876 27.779 27.905
    SSIM 0.924 0.930 0.930 0.934 0.933 0.934
    FSIM 0.9565 0.9586 0.9581 0.9592 0.9588 0.9588
    CC 0.9783 0.9841 0.9837 0.9846 0.9841 0.9848
    House MSE 119.413 113.792 118.699 112.433 113.106 110.616
    PSNR 27.36 27.57 27.386 27.622 27.596 27.693
    SSIM 0.852 0.861 0.855 0.863 0.862 0.865
    FSIM 0.9459 0.9462 0.9445 0.9461 0.9462 0.9457
    CC 0.9824 0.9833 0.9825 0.9834 0.9832 0.9837
    Lion MSE 106.518 99.932 105.505 99.009 99.477 98.498
    PSNR 27.857 28.134 27.898 28.174 28.154 28.197
    SSIM 0.790 0.809 0.799 0.812 0.811 0.814
    FSIM 0.9601 0.9625 0.9606 0.9627 0.9629 0.9623
    CC 0.9836 0.9846 0.9838 0.9848 0.9855 0.9848
    Stained Glass MSE 400.909 387.65 387.415 379.12 378.228 377.372
    PSNR 22.1 22.246 22.249 22.343 22.353 22.363
    SSIM 0.787 0.798 0.802 0.806 0.805 0.804
    FSIM 0.9245 0.9228 0.9304 0.9278 0.9270 0.9272
    CC 0.9683 0.9692 0.9691 0.9698 0.9706 0.9700
    Colorful MSE 114.975 112.91 113.786 110.039 110.317 108.504
    PSNR 27.525 27.603 27.57 27.715 27.704 27.776
    SSIM 0.851 0.857 0.853 0.857 0.857 0.861
    FSIM 0.9719 0.9713 0.9710 0.9718 0.9719 0.9715
    CC 0.9893 0.9895 0.9894 0.9898 0.9903 0.9899
    Newspaper MSE 154.087 143.778 144.271 138.333 139.746 138.528
    PSNR 26.253 26.554 26.539 26.722 26.677 26.715
    SSIM 0.860 0.871 0.872 0.876 0.875 0.876
    FSIM 0.9575 0.9587 0.9589 0.9596 0.9595 0.9592
    CC 0.9794 0.9815 0.9806 0.9815 0.9821 0.9814
    Wheel MSE 286.172 271.404 277.259 266.933 267.872 265.138
    PSNR 23.565 23.795 23.702 23.867 23.852 23.896
    SSIM 0.711 0.724 0.726 0.737 0.736 0.734
    FSIM 0.9357 0.9334 0.9295 0.9316 0.9314 0.9314
    CC 0.9332 0.9371 0.9353 0.9377 0.9374 0.9382

     | Show Table
    DownLoad: CSV

    In 2× interpolation experiment, proposed method together with our previous SGAR method achieved best MSE in 18 of 25 images, best PSNR in 18 of 25 images, best SSIM index in 21 of 25 images, best FSIM index in 8 of 25 images and best CC index in 21 of 25 images. According to the PNSR index, 7 of the 25 interpolated HR images obtained the best effect with the ICBI method, and the other 18 interpolated HR images obtained the best effect with the method proposed in this paper. The ratio was as high as 72%. For these 18 images, the PSNRs of the proposed method were improved to different degrees compared to those of the other four methods. The maximum, minimum and average increments were 1.672, 0.005 and 0.267 dB, respectively, and the maximum, minimum and average relative increments were 5.79, 0.02 and 0.94%, respectively. Compared with NEDI, the maximum, minimum and average increases in PSNR were 1.672, 0.234 and 0.547 dB, respectively, and the relative maximum, minimum and average increases were 5.79, 0.78 and 1.94%, respectively. Compared with bicubic, the maximum, minimum and average increases in PSNR were 0.429, 0.063 and 0.182 dB, respectively, and the relative maximum, minimum and average increases were 1.45, 0.22 and 0.63%, respectively. Compared with i-NEDI, the maximum, minimum and average increases in PSNR were 0.577, 0.051 and 0.281 dB, respectively, and the relative maximum, minimum and average increases were 1.82, 0.17 and 0.98%, respectively. Compared with ICBI, the maximum, minimum and average increases in PSNR were 0.353, 0.005 and 0.06 dB, respectively, and the relative maximum, minimum and average increases were 1.11, 0.02 and 0.21%, respectively. According to the SSIM index, 17 of 25 images obtain the best effect with the ICBI method, compare with the second best ICBI method, the maximum, minimum and average increments were 0.02099, 0.00019 and 0.00251, respectively, and the maximum, minimum and average relative increments were 2.284, 0.021 and 0.283%, respectively. Compared with bicubic, which accommodate non-integer scaling factors, the maximum, minimum and average increases in SSIM were 0.00976, 0.00124 and 0.00461, respectively, and the relative maximum, minimum and average increases were 1.35, 0.13 and 0.52%, respectively. Two points should be specified: First, both bicubic and the proposed method can achieve non-integer image magnification. Compared with the bicubic method, the proposed method improved the interpolation effect to a certain extent in all 25 experimental images. Second, both ICBI and the proposed method use the iterative curvature method. Thus, the two methods achieved the best interpolation effect among these edge-directed methods. The difference is that the method presented in this paper uses the SGAR model to predict the unknown pixels in the first step the increasing trend is obvious.

    Compared with the most popular non-integer interpolate methods and other conventional edge-directed methods, the HR images obtained by the proposed method have significantly fewer blurring effects. As is shown in Figures 5 and 6, these defects are especially obvious among non-integer methods. In the visual comparison of 2× magnification, these defects also exist in Figures 7(a) and 8(a). It can be seen that the interpolation method based on the SGAR model has a better image description ability than the interpolation method based on the B-spline theory. The isotropic low-pass filter can enhance the smoothing effect, while the SGAR model is more suitable for revealing the distribution pattern of the image. Furthermore, the interpolation method based on the SGAR model has better color fidelity than NEDI, as shown in Figure 9. As shown in Figures 7 (d) and 8 (d), the HR image interpolated by ICBI has more visual defects. To analyze the reason, the proposed method used a curvature iterative method based on the discrete features of the image to describe the curvature change of anchor pixels in more directions, which helps eliminate interpolation defects such as artifacts in edge areas and texture-rich areas.

    Figure 4.  Images used in the experiments.
    Figure 5.  Visual comparisons of 2.2× magnification on the bench image.
    Figure 6.  Visual comparisons of 2.2× magnification on the newspaper image.
    Figure 7.  Visual comparisons of 2× magnification on the sunflower image.
    Figure 8.  Visual comparisons of 2× magnification on the Carrousel image.
    Figure 9.  Color infidelity in comparing methods.

    In machine vision system, especially in online visual inspection system. Due to the vibration and overheating problems, image noise will inevitably in the poor working conditions. In order to better combine with the potential application scenarios, gaussian noise with mean value of 0 and variance of 0.005 is added to gray image. The images used in this part of the experiment are shown in Figure 10.

    Figure 10.  Noise images.

    We add the latest learning-based Meta-SR [44] to the comparison. In comparing with the traditional Bicubic interpolation method and Meta-SR [44], our previous SGAR method and the proposed method have made better progress in objective indicators on all of 5 images. Although, on some objective indicators, the edge-based method like NEDI and i-NEDI are better results than ours, these methods also magnify the impact of noise on visual quality. These artifacts are especially obvious in Figure 11.

    Figure 11.  Visual comparisons of 2× magnification on the noise image.

    Frankly, the naive for-loop and iterative process make the processing time relatively long, the proposed method is much slower than ICBI, let along other popular non-integer interpolation methods well-established in OpenCV. However, the processing speed can be improved with parallel threads simultaneously process on NVIDIA CUDA devices and the research will continue in future works.

    Table 4.  Comparison of 2× magnification interpolation results on noise images.
    Image Scale Criterion NEDI Bicubic i-NEDI ICBI Meta-SR SGAR Proposed
    Image 1 MSE 444.632 461.051 407.923 416.077 1240.412 361.525 402.718
    PSNR 21.651 21.493 22.025 21.939 17.195 22.549 22.081
    SSIM 0.421 0.417 0.428 0.401 0.225 0.422 0.418
    FSIM 0.763 0.758 0.757 0.752 0.752 0.768 0.762
    CC 0.923 0.921 0.932 0.930 0.820 0.939 0.932
    Image 2 MSE 521.417 548.597 491.627 528.783 1295.417 467.874 531.563
    PSNR 20.959 20.738 21.214 20.898 17.007 21.430 20.875
    SSIM 0.433 0.420 0.430 0.399 0.228 0.419 0.415
    FSIM 0.770 0.761 0.758 0.751 0.690 0.766 0.762
    CC 0.923 0.920 0.929 0.924 0.836 0.932 0.923
    Image 3 MSE 439.341 489.277 441.469 466.365 1158.269 327.368 440.931
    PSNR 21.703 21.235 21.682 21.444 17.493 22.980 21.687
    SSIM 0.464 0.461 0.475 0.436 0.238 0.507 0.454
    FSIM 0.807 0.799 0.796 0.791 0.734 0.839 0.802
    CC 0.784 0.771 0.803 0.789 0.605 0.939 0.797
    Image 4 MSE 391.395 414.673 383.998 402.288 1054.442 411.358 391.838
    PSNR 22.205 21.954 22.288 22.085 17.901 21.989 22.200
    SSIM 0.437 0.440 0.450 0.417 0.241 0.457 0.434
    FSIM 0.784 0.786 0.782 0.776 0.713 0.805 0.786
    CC 0.898 0.894 0.905 0.900 0.780 0.810 0.902
    Image 5 MSE 358.182 368.718 354.381 376.873 900.455 349.492 347.985
    PSNR 22.590 22.464 22.636 22.369 18.586 22.696 22.715
    SSIM 0.505 0.527 0.539 0.504 0.304 0.437 0.521
    FSIM 0.836 0.842 0.838 0.833 0.787 0.789 0.843
    CC 0.931 0.930 0.935 0.930 0.858 0.912 0.935

     | Show Table
    DownLoad: CSV

    Based on our and others’ previous work, this paper introduced a new method for image interpolation via integration. The new method is based on the SGAR model and can accommodate arbitrary scaling factors. First, the paper discussed how to use the SGAR model to describe the image window, including the establishment of a linear autoregressive model, the SGAR model and the relationship between anchor pixels and their neighboring pixels. By grouping the gradient directions, the adaptive extension direction of the image window was determined, and an image window adaptive extension method based on the gradient angles for the SGAR model was formed. Because the product terms were introduced into the SGAR model, the degree of freedom of the model was increased, and the ability to describe the model was enhanced, but this may cause overfitting problems. Therefore, an elastic network was introduced into the solution of the objective function to address the overfitting problem. Finally, the curvatures were calculated in eight directions, and the interpolation results were updated accordingly to improve the interpolation accuracy. Experiments on 25 images show that the objective measures of the proposed method were improved to a certain extent. Subjective visual effect evaluations were carried out, and much better results were achieved. Therefore, the method presented in this paper improved the objective index of image interpolation and enhanced the subjective visual effect.

    Mr. Hu reports grants from National Natural Science Foundation of China, grants from Postgraduate Research & Practice Innovation Program of Jiangsu Province, during the conduct of the study.

    This work was supported in part by the National Natural Science Foundation of China under Grant No. 51705238 and the Postgraduate Research & Practice Innovation Program of Jiangsu Province Grant No. SJCX19_0491.

    We declare that we do not have any commercial or associative interest that represents a conflict of interest in connection with the work submitted.



    [1] Y. Tian, M. M. Kaleemullah, M. A. Rodhaan, B. Song, A. Al-Dhelaan, T. Ma, A privacy preserving location service for cloud-of-things system, J. Parallel Distrib. Comput., 123 (2019), 215-222. doi: 10.1016/j.jpdc.2018.09.005
    [2] W. Wang, W. Zhang, Z. Jin, K. Sun, R. Zou, C. Huang, et al., A Novel Location Privacy Protection Scheme with Generative Adversarial Network, International Conference on Big Data and Security Springer, 2019.
    [3] L. Wang, X. Shu, W. Zhang, Y. Chen, Design and Optimization of Evaluation Metrics in Object Detection and Tracking for Low-Altitude Aerial Video, International Conference on Big Data and Security, 2019.
    [4] Y. Liu, M. Pang, Research on Medical Image Encryption Method Based on Chaotic Scrambling and Compressed Sensing, International Conference on Big Data and Security, 2019.
    [5] B. Song, M. M. Hassan, A. Alamri, A. Alelaiwi, Y. Tian, M. Pathan, et al., A two-stage approach for task and resource management in multimedia cloud environment, Computing, 98 (2016), 119-145.
    [6] Z. Pan, C. N. Yang, V. S. Sheng, N. Xiong, W. Meng, Machine Learning for Wireless Multimedia Data Security, Sec. Commun. Networks, 2019 (2019), 7682306.
    [7] S. Ousguine, F. Essannouni, L. Essannouni, M. Abbad, D. Aboutajdine, A New Image Interpolation Using Laplacian Operator, International Symposium on Ubiquitous Networking, 2016, Singapore.
    [8] S. Pan, L. Han, Y. Tao, Q. Liu, Study on Indicator Recognition Method of Water Meter Based on Convolution Neural Network, International Conference on Big Data and Security, 2019, Springer.
    [9] J. A. Parker, R. V. Kenyon, D. E. Troxel, Comparison of Interpolating Methods for Image Resampling, IEEE Trans. Med. Imaging, 2 (1983), 31-39. doi: 10.1109/TMI.1983.4307610
    [10] R. Hanssen, R. Bamler, Evaluation of interpolation kernels for SAR interferometry, IEEE Trans. Geosci. Remote Sens., 37 (1999), 318-321. doi: 10.1109/36.739168
    [11] T. M. Lehmann, C. Gonner, K. Spitzer, Survey: interpolation methods in medical image processing, IEEE Trans. Med. Imaging, 18 (1999), 1049-1075. doi: 10.1109/42.816070
    [12] D. Y. Han, Comparison of Commonly Used Image Interpolation Methods, Proceedings of the 2nd international conference on computer science and electronics engineering, Atlantis Press, 2013.
    [13] A. Amanatiadis, I. Andreadis, Performance evaluation techniques for image scaling algorithms, IEEE International Workshop on Imaging Systems and Techniques, 2008.
    [14] Y. Li, F. Qi, Y. Wan, Improvements On Bicubic Image Interpolation, 2019 IEEE 4th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC), 2019.
    [15] H. Hsieh, H. Andrews, Cubic splines for image interpolation and digital filtering, IEEE Trans. Acoust. Speech Signal Process., 26 (1978), 508-517. doi: 10.1109/TASSP.1978.1163154
    [16] S. Abbas, M. Irshad, M. Z. Hussain, Adaptive image interpolation technique based on cubic trigonometric B-spline representation, IET Image Process., 12 (2018), 769-777. doi: 10.1049/iet-ipr.2016.0393
    [17] D. D. Muresan, T. W. Parks, Adaptively quadratic (AQua) image interpolation, IEEE Trans. Image Process., 13 (2004), 690-698. doi: 10.1109/TIP.2004.826097
    [18] C. Chen, C. Lai, Iterative Linear Interpolation Based on Fuzzy Gradient Model for Low-Cost VLSI Implementation, IEEE Trans. VLSI Syst., 22 (2014), 1526-1538. doi: 10.1109/TVLSI.2013.2276410
    [19] Q. Wang and R. K. Ward, A New Orientation-Adaptive Interpolation Method, IEEE Trans. Image Process., 16 (2007), 889-900. doi: 10.1109/TIP.2007.891794
    [20] C. M. Zwart, D. H. Frakes, Segment Adaptive Gradient Angle Interpolation, IEEE Trans. Image Process., 22 (2013), 2960-2969. doi: 10.1109/TIP.2012.2228493
    [21] X. Li, M. T. Orchard, New edge-directed interpolation, IEEE Trans. Image Process., 10 (2001), 1521-1527. doi: 10.1109/83.951537
    [22] X. Zhang, X. Wu, Image Interpolation by Adaptive 2-D Autoregressive Modeling and Soft-Decision Estimation, IEEE Trans. Image Process., 17 (2008), 887-896. doi: 10.1109/TIP.2008.924279
    [23] Q. Wang, J. Liu, W. Yang, Z. Guo, Adaptive autoregressive model with window extension via explicit geometry for image interpolation, IEEE International Conference on Image Processing (ICIP), 2015.
    [24] W. Yang, J. Liu, S. Yang, Z. Guo, Novel autoregressive model based on adaptive window-extension and patch-geodesic distance for image interpolation, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015.
    [25] W. H. Yang, J. Y. Liu, M. D. Li, Z. Gao, Isophote-Constrained Autoregressive Model With Adaptive Window Extension for Image Interpolation, IEEE Trans. Circuits Syst. Video Technol., 28 (2018), 1071-1086. doi: 10.1109/TCSVT.2016.2638864
    [26] F. Hao, J. Shi, R. Chen, S, Zhu, Z. Zhang, Noise-insensitive and edge-preserving resolution upconversion scheme for digital image based on the spatial general autoregressive model, IET Image Process., 10 (2016), 280-288. doi: 10.1049/iet-ipr.2015.0095
    [27] K. Chang, P. L. K. Ding, B. Li, Single Image Super-resolution Using Collaborative Representation and Non-local Self-Similarity, Signal Process., 149 (2018), 49-61.
    [28] D. Cheng, K. I. Kou, FFT Multichannel Interpolation and Application to Image Super-resolution, Signal Process., 162 (2019), 21-34. doi: 10.1016/j.sigpro.2019.03.025
    [29] A. Giachetti, N. Asuni, Real-Time Artifact-Free Image Upscaling, IEEE Trans. Image Process., 20 (2011), 2760-2768. doi: 10.1109/TIP.2011.2136352
    [30] H. Kim, Y. Cha, S. Kim, Curvature Interpolation Method for Image Zooming, IEEE Trans. Image Process., 20 (2011), 1895-1903. doi: 10.1109/TIP.2011.2107523
    [31] A. Marquina, S. Osher, A New Time Dependent Model Based on Level Set Motion for Nonlinear Deblurring and Noise Removal, International Conference on Scale-Space Theories in Computer Vision, 1999.
    [32] T. F. Chan, J. J. Shen, Image processing and analysis: variational, PDE, wavelet, and stochastic methods, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2005.
    [33] J. Hu, Y. Luo, Single-image superresolution based on local regression and nonlocal self-similarity, J. Electron. Imaging, 23 (2014), 033014. doi: 10.1117/1.JEI.23.3.033014
    [34] N. Asuni, A. Giachetti, Accuracy Improvements and Artifacts Removal in Edge Based Image Interpolation, VISAPP (1), 8 (2008), 58-65.
    [35] W. Dong, L. Zhang, R. Lukac, G. Shi, Sparse Representation Based Image Interpolation With Nonlocal Autoregressive Modeling, IEEE Trans. Image Process., 22 (2013), 1382-1394. doi: 10.1109/TIP.2012.2231086
    [36] F. Hao, J. Shi, Z. Zhang, R, Chen, S. Zhu, Canny edge detection enhancement by general auto-regression model and bi-dimensional maximum conditional entropy, Optik, 125 (2014), 3946-3953. doi: 10.1016/j.ijleo.2014.01.153
    [37] M. Li, J. Liu, J. Ren, Z. Guo, Adaptive General Scale Interpolation Based on Weighted Autoregressive Models, IEEE Trans. Circuits Syst. Video Technol., 25 (2015), 200-211. doi: 10.1109/TCSVT.2014.2347531
    [38] A. Ullah, M. Z. Asghar, A. Habib, S. Aleem, F. M. Kundi, A. M. Khattak, Optimizing the Efficiency of Machine Learning Techniques, International Conference on Big Data and Security, 2019.
    [39] R. M. Rifkin, R. A. Lippert, Notes on regularized least squares, Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory, 2007.
    [40] K. Koh, S. J. Kim, S. Boyd, An interior-point method for large-scale l1-regularized logistic regression, J. Mach. Learn. Res., 8 (2007), 1519-1555.
    [41] J. Friedman, T. Hastie, R. Tibshirani, Regularization paths for generalized linear models via coordinate descent, J. Stat. Software, 33 (2010), 1.
    [42] H. Wang, W. Guan, K. Zhang, Over-Sampling Multi-classification Method Based on Centroid Space, International Conference on Big Data and Security, 2019.
    [43] Y. Gong, I. F. Sbalzarini, Curvature filters efficiently reduce certain variational energies, IEEE Trans. Image Process., 26 (2017), 1786-1798. doi: 10.1109/TIP.2017.2658954
    [44] X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, J. Sun, Meta-SR: A Magnification-Arbitrary Network for Super-Resolution, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.
  • This article has been cited by:

    1. Keruo Jiang, Zhen Huang, Xinyan Zhou, Chudong Tong, Minjie Zhu, Heshan Wang, Deep belief improved bidirectional LSTM for multivariate time series forecasting, 2023, 20, 1551-0018, 16596, 10.3934/mbe.2023739
  • Reader Comments
  • © 2020 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(3859) PDF downloads(86) Cited by(1)

Figures and Tables

Figures(11)  /  Tables(4)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog