interpolation methods based on the discrete ideas can better preserve image structure of edge. However, it is less than
satisfactory in maintaining image details. The discrete method can’t amplify any multiple of the image, and which has a
high cost of time. The interpolation methods based on the continuous idea can obtain richer image detail information. But,
it can’t maintain image edge structure well. In order to inherit advantages of the discrete method and the continuous
method, a new method of rational function image interpolation based on gradient optimization is proposed. Method
Firstly, a novel type of bivariate rational interpolation function is constructed. With shape parameters varying, the
function has different forms of expression. It is an organic unity of polynomial model and rational model. The constructed
continuous rational function interpolation model inherits the advantages of the continuous method, and reduces the
appearance of the jagged edge due to smoothness to some extent. Secondly, according to the regional characteristics, the
image is divided into texture region and smooth region automatically by using isoline method. If the interpolation unit has
at least an isoline, the unit belongs to the texture region. If the interpolation unit doesn’t have isolines, the unit belongs to
the smooth region. The smooth region is interpolated by polynomial model, and the texture region is interpolated by
rational model. Finally, according to the
operator, the image gradient of the interpolation unit is
calculated, and the direction of texture region is determined. According to the image gradient and texture direction, the
weight influence factor of every interpolation unit is obtained. And then the center of the image patch with different
directions is optimized by convoluting with weight matrix. Result A rational function image interpolation algorithm
based on gradient optimization is proposed. The proposed algorithm is tested form three aspects: objective data, visual
effect and time complexity. Compared with the state-of-the-art interpolation algorithm, the average PSNR of the proposed
method is 1.5dB, 0.36dB, 0.14dB, 0.28dB, 1.11dB and 0.95dB higher than that of Bicubic, RSAI, DFDF, NARM, NEDI
and Lee’s, respectively. The average SSIM of the proposed method is 0.0968, 0.0072, 0.0076, 0.0052, 0.0141 and 0.0237
higher than that of Bicubic, RSAI, DFDF, NARM, NEDI and Lee’s, respectively. The image reconstructed by proposed
method has richer texture detail and sharper edge structure. The average run time of the proposed method is 7 seconds. It
is 3.28 times, 5.26 times, 53.28 times, 43.53 times and 418.54 times faster than that of DFDF, NEDI, RSAI, Lee’s and
NARM. For texture images such as Baboon, Barbara and Metal, the proposed method is highly competitive not only in
objective data but also in visual effect. Conclusion We construct a type of bivariate rational interpolation function in this
paper. On the basis of this model, an image interpolation algorithm based on gradient optimization is presented, which not
only can scale the reconstructed image indefinitely, but also have lower time complexity. Experimental results show that
proposed algorithm preserves image details and structure of edge effectively, and has a high quality of interpolation
image.
Key words: image interpolation; rational function; gradient optimization; regional division; isoline method