[0041] Example 1

[0042] like Figure 1-2 As shown, in embodiment 1, a kind of image restoration method of blurred image specifically comprises the following steps:

[0043] S11: performing wavelet decomposition on the image to be restored to obtain a multi-band decomposition image;

[0044] S12: Calculate the autocorrelation function of the decomposed image in each frequency band, and then obtain the initial point spread function of the decomposed image;

[0045] S13: Perform R-L iterative processing on the decomposed image according to the initial point spread function;

[0046] S14: Perform wavelet inverse transform on the decomposed image after R-L iterative processing to realize image restoration processing.

[0047] The present invention uses wavelet decomposition to separate the components of each frequency band in the image to be restored (fuzzy image), and performs R-L iterative processing on the decomposed image based on the point spread function of each decomposed image, which can effectively reduce the phenomenon of image noise amplification and at the same time, better Save image detail information, improve the accuracy of deconvolution algorithm, and then get a clear restored image.

[0048] Further, performing wavelet decomposition on the image to be restored is specifically performing dual-tree complex wavelet decomposition on the image to be restored twice. Specifically, the highest decomposition level of dual-tree complex wavelet decomposition for the image to be restored is M, M≤10, and M=2 in this embodiment. After two decompositions of the image to be restored, multiple high-frequency decomposed images and multiple A low frequency band decomposes the image.

[0049]More specifically, in this embodiment, the blurred image is recorded as image I, and the size of image I is m*n. After image I undergoes the first dual-tree complex wavelet transform (decomposition), the image I is filtered And after downsampling, the size obtained is The real decomposition image I xx , imaginary decomposition image I' xx , based on the first dual-tree complex wavelet decomposition of the image I, the second dual-tree complex wavelet decomposition is performed, that is, after filtering and down-sampling processing, the size is The real decomposition image I xxx , imaginary decomposition image I' xxx ,like Figure 3-4 as shown, image 3 middle h 0 、h 1 is the conjugate orthogonal filter pair of tree A, g 0 , g 1 is the conjugate orthogonal filter pair of tree B, and ↓ represents downsampling. filter pair h 0 、h 1 The corresponding real scaling function φ h (t) and wavelet function for:

[0050]

[0051]

[0052] filter pair g 0 , g 1 The corresponding real scaling function φ g (t) and wavelet function for:

[0053]

[0054]

[0055] This embodiment Figure 4 In, I 01 , I 10 , I 11 The high-frequency detail images in the directions of ±15°, ±45°, and ±75° of the tree A part after the original image I completes a dual-tree complex wavelet transform, I' 01 , I' 10 , I' 11 Respectively, the original image I completes a dual-tree complex wavelet transform and the high-frequency detail images in the directions of ±15°, ±45°, and ±75° of the tree B part, I 01 , I 10 , I 11 and I' 01 , I' 10 , I' 11 The image dimensions of the decomposed image are I 001 , I 010 , I 011 After the original image I completes the first dual-tree complex wavelet transform, the low-frequency image I 00 The high-frequency detail images in the directions of ±15°, ±45°, and ±75° obtained after double-tree complex wavelet decomposition again, I' 001 , I' 010 , I' 011 After the original image I completes the first dual-tree complex wavelet transform, the low-frequency image I' 00 The high-frequency detailed images in the directions of ±15°, ±45°, and ±75° obtained after double-tree complex wavelet decomposition again, I 001 , I 010 , I 011 and I' 001 , I' 010 , I' 011 The image dimensions of the decomposed image are while I' 000 , I 000 is the low-frequency information obtained after two times of dual-tree complex wavelet decomposition, and the image size is

[0056] The invention performs double-tree complex wavelet decomposition on the image to be restored twice, and the decomposed image has time-frequency local analysis characteristics, and has approximate translation invariance and multi-directional selectivity, can better retain image details, and avoid Gibuss phenomenon.

[0057] Further, the calculation formula of the autocorrelation function R of each frequency band decomposition image in step S12 is:

[0058] R=g(x,y)*g(x,y)

[0059] Among them, g(x, y) is the decomposed image of each frequency band, * represents the complex conjugate, and x, y represent the pixel coordinates of the image respectively.

[0060] Further, decompose the initial point spread function (PSF) h of the image 0 The calculation formula is:

[0061] h 0 =R-min(R)+ε[max(R)-min(R)]

[0062] Among them, ε represents the dynamic change percentage of the autocorrelation function R, which is generally a small non-negative integer, and the value here is 0.01.

[0063] Further, performing R-L iterative processing on the decomposed image according to the initial point spread function also includes:

[0064] The corresponding number of R-L iterations is determined according to the size and frequency band of each decomposed image. The present invention determines the corresponding number of R-L iterations according to the size and frequency band of the decomposed image, and can maximize the removal of noise in each decomposed image according to the characteristics of each decomposed image.

[0065] Further, the number of R-L iterations of each decomposed image follows:

[0066] The number of R-L iterations of the large-scale image < the number of R-L iterations of the high-frequency decomposition image < the number of R-L iterations of the low-frequency decomposition image, that is, the number of R-L algorithm iterations is used for the high-frequency detail part, and the low-frequency image part uses the R-L algorithm for multiple iterations. Larger images require fewer iterations of the R-L algorithm.

[0067] In this example, I 000 size is The low-frequency image of I 001 , I 010 , I 011 and I' 001 , I' 010 , I' 011 is the size of The high-frequency detail image of I 01 , I 10 , I 11 and I' 01 , I' 01 , I' 11 is the size of High-frequency detailed images, such as image 3 , Figure 5 As shown, will decompose the image I 01 , I 10 , I 11 , I' 01 , I' 01 , I' 11 The number of iterations is set to N, then the decomposition I 001 , I 010 , I 011 and I' 001 , I' 010 , I' 011 The number of iterations is 2N, decomposing the image I 000 The number of iterations is 3N. Furthermore, the number of R-L iterations of the large-scale image is greater than or equal to 10 (the number of iterations of the large-scale high-frequency decomposition image (I 01 ) 001 ) 000 )), then in this embodiment, the decomposition image I 01 , I 10 , I 11 , I' 01 , I' 01 , I' 11 The number of iterations is set to 10, then the decomposition I 001 , I 010 , I 011 and I' 001 , I' 010 , I' 011 The number of iterations is 20, decomposing the image I 000 The number of iterations is 30.

[0068] Further, the calculation formula for R-L iterative processing of the decomposed image according to the initial point spread function is:

[0069]

[0070]

[0071] Among them, n is the number of iterations, h(x,y) is the point spread function, h 0 is the initial point spread function, and f(x,y) is the restored image.

[0072] Finally, the dual-tree complex wavelet inverse transform is performed on the decomposed image f(x, y) that has completed R-L iterative processing, namely image 3 The inverse operation of , a clear restored image M can be obtained.