| As an important field of image processing,traditional image interpolation algorithm has a very wide range of applications at present.Compared with the image interpolation algorithm based on learning,the traditional image interpolation algorithm has the advantages of low algorithm complexity and fast processing speed.Much commercial software,such as Microsoft office and Adobe Photoshop,integrate traditional image interpolation algorithms such as nearest neighbor interpolation,bilinear interpolation and bicubic interpolation for image scaling.In addition,many printer drivers are also using this technology.The research goal of this paper is to improve the interpolation accuracy and image quality under the premise of maintaining the processing speed advantage of traditional image interpolation algorithm.This paper mainly studies the bicubic interpolation algorithm which is widely used in the traditional interpolation algorithm.The bicubic interpolation algorithm has two implementation methods in the application,which are 16 point-ordinary bicubic interpolation algorithm and 16 point-convolution bicubic interpolation algorithm.The difference between them is that the solution of interpolation kernel is different.The16 point-ordinary bicubic interpolation algorithm solves the interpolation kernel according to the bicubic term formula of the interpolation kernel,uses the pixel value of 16 points in the image and the derivative relationship to construct the interpolation parameter equations about bicubic term,so as to obtain the value of 16 interpolation parameters.The 16 point-convolution bicubic interpolation algorithm simplifies the calculation of two-dimensional plane interpolation kernel to two dimensions in thedirection and thedirection respectively.The coefficients of the four interpolation points in the corresponding direction are calculated from two one-dimensional Spaces respectively,and the values of the 16 coefficients in the interpolation kernel are obtained by multiplying the coefficients in the two directions.In this paper,the principles of the above two bicubic interpolation algorithms are studied and programmed.Experiments show that the 16 point-convolution bicubic interpolation algorithm is better than the 16 point-ordinary bicubic interpolation algorithm in interpolation performance,and the image reconstruction quality is higher.In real life,16 point-convolution bicubic interpolation algorithm is more widely used than 16 point-ordinary bicubic interpolation algorithm.In this paper,the 16 point-convolution bicubic interpolation algorithm is studied and further improved.The main contributions of this paper are as follows.1.The size of the interpolation kernel of the 16 point-ordinary bicubic interpolation algorithm is 4×4.In general,the larger the support range is,the more fuzzy the interpolation effect is.In this paper,a bicubic interpolation algorithm with smaller kernel support is proposed.The correlation equation of 16 coefficients of bicubic term formula of interpolation kernel is obtained by pixel value and derivative relationship of every 9 points in low-resolution image,which effectively reduces the range of kernel support.In this paper,the 9 point-bicubic interpolation algorithm is tested.The experimental results show that the reduced kernel support does not get higher peak signal to noise ratio,but the structure similarity is improved.This paper analyzes the results.(This content corresponds to Chapter 3 of this article)2.In the traditional 16 point-convolution bicubic interpolation algorithm,the interpolation kernel parameter(6 is usually set to-0.5 to approximate the Taylor second-order expansion of the original function to obtain a relatively smooth interpolation kernel function.In this paper,the experimental observation shows that for most natural images,when the value of(6 is-0.75 for bicubic interpolation,the reconstruction effect will be better than(6=-0.5(the peak signal-to-noise ratio score and the structure similarity score are higher),which shows that(6=-0.5 is not the optimal interpolation parameter value,and the value of(6 can be further optimized.In order to improve the performance of image interpolation,this paper proposes a new method to improve the peak signal-to-noise ratio of interpolation image.In this method,we directly improve the mean square of the image,and construct the higher-order optimization function of the univariate cubic equation about the interpolation kernel parameter(6,so as to effectively improve the value of the peak signal-to-noise ratio of the interpolation image.The value of the optimized parameter(6 can be obtained by solving the one variable cubic equation.Through the experiment,it is found that the optimized parameter(6 is used to reconstruct the image,and the obtained image has better performance in peak signal-to-noise ratio and structure similarity.At the same time,the advantages of this method can be proved in the visual effect,and the detail processing is more clear and smooth.(This content corresponds to Chapter 4 of this paper)3.In order to further improve the mathematical expression ability of the image mean square error optimization function,this paper introduces a new variable(7 to replace the item(6~2 in the objective function on the basis of the fourth chapter,so that the high-order optimization problem can be transformed into the least square problem,which effectively improves the expression ability of the mathematical model and the interpolation accuracy of the image.In this paper,the partial derivatives of(7 and(6 in the improved optimization function are solved,and then the values of(7 and(6 in the interpolation kernel function are obtained.After the experimental test,the new parameter(7 is introduced,the reconstructed image obtained by this method has higher peak signal-to-noise ratio and structure similarity than the method in the previous chapter,and the interpolation performance of the algorithm is further improved.(This content corresponds to Chapter 5 of this article)... |
PDF Full Text Request