WO2019148739A1

WO2019148739A1 - Comprehensive processing method and system for blurred image

Info

Publication number: WO2019148739A1
Application number: PCT/CN2018/091164
Authority: WO
Inventors: 白海玲
Original assignee: 上海康斐信息技术有限公司
Priority date: 2018-01-31
Filing date: 2018-06-13
Publication date: 2019-08-08
Also published as: CN108305230A

Abstract

Disclosed in the present invention are a comprehensive processing method and system for a blurred image. The method comprises: S10, identifying the blur type of a blurred image according to the elongation degree of a blurred image spectrogram, the blur type comprising defocus blur and motion blur; S20, if the blur type of the blurred image is defocus blur, obtaining a blur kernel by using a knife-edge method, and if the blur type of the blurred image is motion blur, obtaining the blur kernel by an adaptive preset algorithm; and S30, obtaining a clear image by a hyper-Laplace prior deconvolution algorithm on the basis of the blur kernel. According to the present invention, the blur types of blurred images are distinguished, and corresponding blur kernel estimation algorithms are set according to the difference between degradation mechanisms of different blur types, so that estimated blur kernels are more in line with actual situations; and especially for images of defocus blur, accurate blur kernel estimation based on a Gaussian model is performed. The present invention is suitable for restoration of more real blurred images, with the restoration method being simple and more targeted and the restoration being accurate, quick, and high-quality.

Description

Fuzzy image comprehensive processing method and system

The present application claims the priority of the Chinese Patent Application No. 201810094023.1, the entire disclosure of which is incorporated herein by reference.

Technical field

The present invention relates to the field of image processing technologies, and in particular, to a fuzzy image integration processing method and system.

Background technique

Image acquisition may result in inaccurate focus, or relative motion of the camera and the camera, camera distortion, air diffraction, etc., resulting in unsatisfactory images, losing useful information in the image, that is, image degradation. . It hinders the subsequent processing analysis, especially in the face of some non-replicable scenes. Therefore, for such cases, restoring the original appearance of the scene is an urgent problem to be solved. Although the actual design can add optical devices, improve transmission equipment and other hardware means such as electronic image shift compensation, optical image shift compensation, mechanical image shift compensation to improve the quality of the captured image, but because the device process is too complicated, The impact of many factors such as high cost and long adjustment time is not universal. Therefore, using image restoration technology to improve image quality from image compensation is a correct, low-cost, low-cost option. .

Image restoration needs to understand the type, mechanism and process of image degradation. Only by knowing these a priori knowledge, it is possible to determine the exact point spread function (PSF), also called fuzzy kernel, to establish a degradation model. The degradation model can be used to reverse the degradation process by algorithm to achieve the restoration of the blurred image.

When the image is restored, if there is no reasonable way to understand the cause of the image degradation and the type of the blur, the most suitable restoration algorithm can not be selected to help the image restoration. The point spread function estimation is not accurate, and the selected restoration algorithm is not appropriate, which will result in low quality. Restoration, not only can not get better fuzzy image restoration effect, but also make the blurred image more fuzzy, so in addition to the accurate estimation of the point spread function is important, it also needs to obtain accurate prior knowledge, that is, the image Causes of degradation and types of blurring, etc.

Image restoration algorithms are mathematically defined as a class of ill-posed inverse problems. Knowing the cause, the result is a positive problem, and the known result is the inverse problem; the image restoration algorithm is an algorithm that knows the reason for the reverse of the result. The ill-posed inverse problem means that the reverse push process is very unstable, that is, it is affected by slight noise, which will cause very large interference to the final speculation, resulting in incorrect results. Therefore, the image is restored to minimize the impact of interference on the reverse process.

The image restoration algorithms commonly used in the prior art include RL filtering, constrained least squares filtering, Wiener filtering, and regular filtering. Common fuzzy image types include motion blur and defocus blur. When recovering clear images, by distinguishing image blur types It is a better method to select a targeted image restoration method for recovery. Whether the criteria for distinguishing image blur types are accurate, and whether the image restoration method is appropriate for the actual image restoration is repeated consideration and design during the restoration process.

The patent document disclosed in the publication No. CN104331871A discloses "an image deblurring method and apparatus", comprising: performing blur region detection on a to-be-processed image, determining a blur region image, and determining a blur type of the blur region, if the blur region If the blur type of the image is defocus blur, the defocus blur parameter estimation algorithm based on differential image autocorrelation is used to determine the defocus radius. If the blur type of the blur region image is motion blur, the motion blur based on cepstrum analysis The parameter estimation algorithm determines the blur direction and the blur scale, and substitutes the estimated parameters into the classical image restoration algorithm to obtain the restored image.

The patent uses different methods to estimate the parameters necessary for restoration according to different types of blur. However, when the defocus radius is estimated based on the disc model, the disc model has a large limitation. For the defocused image of the real shot, It is difficult to recover a clear image; in real life, many complex motions cannot determine the direction and scale, that is, they are not linear motion, so the motion blur parameter estimation algorithm based on cepstrum analysis is used to determine the blur direction and blur. The scale does not provide perfect and realistic prior knowledge for image restoration; the classic image restoration algorithm includes Wiener filtering and LR filtering. Wiener filtering is not ideal when the peak signal-to-noise ratio of the blurred image is small, and the LR filtering is not ideal. It is sensitive to noise, and the restored image has a significant ringing effect, that is, the adaptation range is not wide enough.

Summary of the invention

The technical problem to be solved by the present invention is to provide a fuzzy image comprehensive processing method suitable for more real fuzzy image restoration, simple restoration method, strong pertinence, accurate restoration, high speed and high quality, in view of the above-mentioned deficiencies of the prior art. And system.

In order to achieve the above object, the technical solution adopted by the present invention is:

A method for comprehensively processing blurred images, comprising the following steps:

S10: Identify a blur type of the blurred image according to the elongation of the blurred image spectrogram, where the blur type includes defocus blur and motion blur;

S20: if the blurred image is out-of-focus blur, the edge kernel method is used to obtain the blur kernel; if the blurred image is motion blur, the adaptive preset algorithm obtains the blur kernel, and the preset algorithm includes at least one algorithm;

S30: Obtain a clear image by using a super Laplacian prior deconvolution algorithm based on the fuzzy kernel.

Further, the method further includes the following steps:

S40: processing the clear image by using a super-resolution reconstruction technique to obtain a high-resolution image.

Further, if the blurred image is defocused in the step S20, acquiring the fuzzy kernel by using the edge method includes the following steps:

S201: extract an optimal edge edge image based on a gradient criterion;

S202: Acquire a point spread equation according to the optimal edge edge image, where the point spread equation is the blur kernel.

Further, the step S201 includes the following steps:

Performing an edge detection image by performing Canny edge detection on the blurred image;

Performing a Hough transform on the edge detection image to obtain a step edge image;

Cutting a blade edge image of each edge centering on a center point of each edge in the step edge image, the size of the edge edge image selecting a preset size according to the size and blur degree of the defocused image form;

The gradient values of the edge image are calculated one by one, and the optimal edge image is extracted based on the gradient values.

Further, the calculating the gradient value of the edge image and extracting the optimal edge image according to the gradient value comprises the following steps:

Using a least squares method to linearly fit the edges in the edge image to obtain a edge line;

Extracting an intersection formed by an edge of the edge image and a straight line of the edge as a new edge edge point;

The absolute value of the difference between the average pixel values of the two sides of the new edge edge point is counted as the gradient value of the edge edge image, and the edge edge image having the largest gradient value is extracted as the optimal edge edge image.

Further, the step S202 includes the following steps:

Using an least square method to linearly fit the edges in the optimal edge image to obtain an optimal edge line;

Obtaining a vertical distance of each pixel point in the optimal edge edge image to the optimal edge line as an abscissa, and a gray value of each pixel point is an ordinate to form a scatter plot;

Using the Fermi function to linearly fit the scatter plot to obtain an edge spread function;

The point spread equation is calculated using the edge spread function.

Further, in the step S20:

If the blurred image is motion blur, the adaptive regularization method based on sparse prior is used to estimate the fuzzy kernel.

A fuzzy image integrated processing system, comprising:

a fuzzy type identification module, configured to identify a blur type of the blurred image according to an elongation of the blurred image spectrogram, the blur type including defocus blur and motion blur;

An estimation module, configured to acquire a fuzzy kernel by using a blade edge method if the blurred image is out-of-focus blur; and obtain a fuzzy kernel by using an adaptive preset algorithm if the blurred image is motion blur, the preset algorithm includes at least one algorithm;

An image restoration module is configured to obtain a clear image by using a super Laplacian prior deconvolution algorithm based on the fuzzy kernel.

Further, the system further includes:

a high resolution processing module that processes the sharp image using a super-resolution reconstruction technique to obtain a high resolution image;

The estimation module includes:

a defocusing unit for extracting an edge-edge image based on a gradient criterion and acquiring a point-diffusion equation according to the optimal edge-edge image if the blurred image is defocused, and the point-diffusion equation is The blur kernel;

A motion blur unit is used to acquire a fuzzy kernel based on a sparse priori regularization method if the blurred image is motion blur.

Further, the defocus blur unit includes:

An edge detection subunit, configured to perform Canny edge detection on the blurred image to obtain an edge detection image;

a line detection subunit, configured to perform a Hough transform on the edge detection image to obtain a step edge image;

An image intercepting subunit for intercepting a blade edge image of each edge centering on a center point of each edge of the step edge image, the size of the edge edge image being according to the defocused image size and The degree of blur selects a predetermined size to be formed;

An image extraction subunit, configured to calculate a gradient value of the edge edge image one by one, and extract an optimal edge edge image according to the gradient value;

a linear fitting sub-unit for linearly fitting edges in the optimal edge-edge image by using a least squares method to obtain an optimal edge line;

a scatter plot sub-unit, configured to obtain a vertical distance of each pixel point in the optimal edge-edge image to the optimal edge line as an abscissa, and a gray value of each pixel point is an ordinate, forming a dispersion Dot map

An edge diffusion function sub-unit for linearly fitting the scatter plot with a Fermi function to obtain an edge spread function;

A calculation subunit is configured to calculate a point spread equation using the edge spread function.

After adopting the above technical solution, the beneficial effects of the present invention are:

Determining the type of fuzzy image according to the elongation is not only simple, but also suitable for the clear definition of the fuzzy image type in the actual real scene, which can ensure the accuracy of the fuzzy type identification and help to accurately estimate the fuzzy kernel with different degradation mechanisms. .

The edge edge method is used to adaptively extract the optimal edge image of the defocused image, and the optimal edge image is obtained based on the gradient criterion. On the one hand, it is beneficial to improve the speed of defocus image restoration, on the other hand, it is beneficial to Avoid blindly extracting the edge image, reduce human interference, and improve the accuracy of image restoration.

The fuzzy kernel estimated from the optimal edge-edge image is also closer to the true degradation model, and the recovery speed is faster and more accurate.

The adaptive sparse priori-based regularization method estimates the fuzzy kernel of the motion blurred image, and iterates the target multiple times until convergence, which results in more stable and ideal results, thus solving the ill-posed problem in the image restoration problem.

The super Laplacian prior deconvolution algorithm is used to obtain clear images. The super Laplace a priori is used as a regular term. The accuracy of the fuzzy kernel is relatively low, and the heavy tail of the natural image gradient can be well satisfied. Distribution, reducing the ringing effect and quickly recovering high quality images.

The use of super-resolution reconstruction technology to process the clear image can restore more image details of real objects, help to improve image quality, and help image recognition and image data acquisition and analysis after a series of image restoration operations.

DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the prior art, the drawings are as follows:

1 is a flowchart of a method for comprehensively processing a blurred image according to Embodiment 1 of the present invention;

2 is a flowchart of a method for comprehensively processing a blurred image according to Embodiment 2 of the present invention;

3 is a block diagram of a fuzzy image integrated processing system according to Embodiment 3 of the present invention;

4 is an overall block diagram of a defocus blur unit according to Embodiment 3 of the present invention;

5 is a front and rear image display of a defocused image using the present invention, wherein (a) is a blurred image, and (b) is a restored image;

Figure 6 is a front and rear image display of a motion blurred image using the present invention, wherein (a) is a blurred image and (b) is a restored image.

Detailed ways

The technical solutions of the present invention are further described below with reference to the accompanying drawings, but the present invention is not limited to the embodiments.

Image restoration is the process of restoring a clear, high-quality original image from a blurred, noisy, low-quality, poorly-resolution, degraded image. Image restoration firstly analyzes the cause of image distortion according to the image distortion phenomenon. For example, the fuzzy type of fuzzy image is generally divided into two categories: defocus blur and motion blur. Among them, defocus blur is caused by the fact that the image plane is not on the corresponding focal plane; the motion blur is caused by the relative motion between the imaging system and the target. Understand the causes of image distortion and then create different distortion models for different distortion causes, also known as degenerate models, and finally invert them to restore the original clear image.

The invention distinguishes the fuzzy type of the blurred image according to the elongation of the blurred image spectrogram, and based on different blur type images, different distortion causes, and performs different algorithms for defocus blur and motion blur to establish respective degradation models, which are more targeted. Accurately estimate the fuzzy kernel, and use the super Laplacian prior deconvolution algorithm with less ringing effect and fast calculation speed to recover the image.

Example 1

As shown in FIG. 1 , this embodiment provides a method for comprehensive processing of a blurred image, which includes the following steps:

Although the causes of defocus blur and motion blur distortion are different, it is difficult for the light to accurately distinguish the two types of blur from the naked eye. Therefore, it is necessary to find other breakthrough points for the separation of focal blur and motion blur. After research, it is found that defocus blur and There is a big difference in the spectrogram of the blurred image of the motion blur. The relevant geometric features are extracted from the spectrogram, and the threshold is set to distinguish the fuzzy type. The method is simple and easy to implement. It should be noted that the distinction is made here. The fuzzy type is not simply to distinguish the type of fuzzy degradation, but it is hoped to differentiate according to the type, to establish a degradation model that is more consistent with its degradation process, and to estimate a more realistic fuzzy kernel.

In the actual scene, the blurred image does not necessarily have defocus blur or motion blur in the complete sense, or it may be a mixture of the two, accompanied by some other types of noise, so it is more convenient to extract from the spectrogram. Related geometric features generally perform preprocessing such as smoothing, image enhancement, binarization, etc., and these preprocessing are only used to eliminate noise interference and do not affect the differentiation of true fuzzy image degradation types. In this embodiment, the fuzzy type of the blurred image is identified according to the elongation of the blurred image spectrogram, and the elongation is specifically defined as:

T=W*L/A

Where A is the area of the strip or circle in the spectrogram, and W and L are the width and length of the smallest rectangle surrounding the target respectively. According to the above formula, it can be seen that the elongation can easily distinguish the circular and strip targets. That is, the closer to the circle, the smaller the value of the elongation. The experimental research shows that the spectrogram of motion blur is strip-shaped, and the speckle spectrum is circular. According to the formula of elongation, the elongation of defocus blur is smaller than the elongation of motion blur, and a threshold is set. When the elongation is greater than the threshold, the blur type of the image is motion blur, and vice versa.

The method of distinguishing fuzzy types by elongation is simple, the calculation is convenient and fast, and the accuracy of identification is high, which can lay a good foundation for the later use of fuzzy kernel estimation methods for fuzzy kernel estimation.

The estimation of the fuzzy kernel plays a vital role in the restoration process of the blurred image. If the accuracy of the acquired fuzzy kernel is high, the subsequent restoration can adopt a simpler processing method.

The edge is a kind of image feature, which is the most uncertain place in the image, and the most concentrated image information. The edge is also an important basis for image segmentation, and also an important reference for texture analysis and image recognition. The blur of the defocused image mainly refers to the blur of the edge. The edge edge method is used to obtain the step edge first, and the step edge is derived to calculate the point spread function, so that the fuzzy kernel closer to the real data can be obtained.

In addition, in this embodiment, if the blurred image is motion blur, adaptively selecting a preset algorithm to estimate the fuzzy kernel, the preset algorithm may be a fuzzy kernel based on two-dimensional discrete wavelet transform and cepstrum analysis. Estimation algorithm, maximum likelihood method, Bayesian estimation algorithm, regularization algorithm, maximum entropy method, partial differential equation based algorithm and many other fuzzy kernel estimation algorithms.

Regardless of whether it is a defocused blur image or a motion blur blur image, after calculating the blur kernel, the super Laplacian prior deconvolution algorithm is used to obtain a clear image, that is, using super Laplac as the image first. Knowledge, modeling, and rapid restoration of images for high-quality, clear images.

Example 2

As shown in FIG. 2, the difference from Embodiment 1 is that the present embodiment provides a method for comprehensively processing a blurred image, and the method further includes the following steps:

Specifically, if the blurred image is defocused in the step S20, acquiring the fuzzy kernel by using the edge method includes the following steps:

S201: extract an optimal edge edge image based on a gradient criterion;

S202: Acquire a point spread equation according to the optimal edge edge image, where the point spread equation is the blur kernel, and the point spread equation is a point spread function, referred to as PSF, which is a spatial function representation of the fuzzy kernel.

The step S201 includes the following steps:

The Canny edge detection is performed on the blurred image to obtain the edge detection image. The reason why the Canny edge detection operator is selected is that the algorithm has a good suppression effect on noise, single line response, high positioning accuracy, and some parameters involved in the algorithm have Versatility, which can be used for post-image restoration, Canny edge detection operator can detect the general edges in the image. The quasi-determined position of the edge plays a role in the effective estimation of the subsequent transfer function (MTF) and point spread function (PSF).

The edge detection image is subjected to Hough transform to acquire a step edge image.

Perform a Hough transform on the edge detection image after edge detection. At this time, some obvious Hough transform peak points appear. According to these peak points, you can find and link to the line corresponding to the peak point, and Mark the specific spatial positions of these lines in the image. These lines are also step edges. There are multiple step edges in the regular step edge image. In order to avoid artificially blindly selecting a step edge as the fuzzy kernel estimation. According to the embodiment, the optimal edge edge image extraction based on the gradient criterion is provided.

A blade edge image of each edge is taken centering on a center point of each edge in the step edge image, and the size of the edge edge image is preferably formed by a predetermined size according to the size of the defocused image and the degree of blur.

Specifically, determining the coordinates of the center point of each edge in the step edge image, and taking the center point as a center point, intercepting a certain edge image with a partial edge, that is, how many independent step edges are in the step edge image Actually, there are many edge-edge images. It should be noted that the edge-edge image size should be appropriate, and should include all the important information of the blurred image point spread function; therefore, it is necessary to consider the defocused image when determining the edge-edge image size. The size and blur degree, the edge image block is too small may not contain sufficient information, too large may cause the selected edge to be too curved, the calculation deviation is large, and the calculation process is time consuming, therefore, in actual application, Different images should be properly sized to the edge of the image.

The gradient value r of the edge image is calculated one by one, and the optimal edge image is extracted based on the gradient value r.

Specifically, the gradient value r of the edge image is calculated one by one, and an optimal edge image is extracted in a plurality of the edge images according to the magnitude of the gradient value r. It should be noted that, when calculating the gradient value r of the edge edge image, in order to reduce the gradient criterion calculation error and improve the accuracy of the optimal edge edge image extraction, it is necessary to ensure the number of columns of the edge edge of the edge edge image. The range of the line is larger than the range of the number of lines in which it is located. If the range of the number of columns in the original edge image is smaller than the range of the number of lines in which it is located, the angle is rotated so that the range of the number of columns in which the edge point is located is larger than the range of the number of columns. After the number of rows, as a new edge image, the gradient value r calculation is performed.

Preferably, said calculating the gradient value r of said edge image and extracting the optimal edge image based on the gradient value r comprises the steps of:

First, the edge in the edge image is linearly fitted by the least squares method to obtain the edge line.

In general, the edge in the selected edge image is likely not a straight line in the strict sense, or the edge edge point distribution model caused by the edge detection error is not a straight line, so the edge is assumed in this embodiment. The point obeys the straight line model, and the edge point of the edge in the edge image is straight-line fitted by the least square method. The expression of the fitted line is:

y=ax+b

Where n is the number of edge points of the edge, x _k is the number of edge points, and y _k represents the relative position of the edge line. First, the linear coefficients a and b are obtained to determine the fitted edge line.

Next, an intersection formed by overlapping the edge in the edge image with the edge of the edge is extracted as a new edge edge point.

Finally, the absolute value of the difference between the average pixel values of the two sides of the new edge edge point is counted as the gradient value r of the edge edge image, and the edge edge image with the largest gradient value r is extracted as the optimal edge edge image. .

Specifically, the gradient value r is calculated as follows:

r=|gc_1-gc_2|

Where m and n are the length and width of the edge image, respectively, r _j (j=1, 2, . . . , n) represents the number of lines in the edge of each edge of the edge, and e _i,j represents the position ( Gray value at i, j).

According to the above formula, the gradient value r calculated by using the edge edge information clearly reflects the distribution of the gray value of the bright and dark areas on both sides of the edge. The larger the r value, the gray value on both sides of the edge. The greater the contrast and the higher the similarity of the gray values on the same side, the more favorable the estimation of the fuzzy kernel and the subsequent image restoration. Therefore, the edge edge image having the largest gradient value r is extracted as an optimum edge edge image.

The step S202 includes the following steps:

Using the least square method to linearly fit the edge in the optimal edge image to obtain the optimal edge line; the optimal edge line here is consistent with the calculation formula of the edge line in step S201. If there is a calculation before the edge line is saved, it can be read directly for calculation.

After obtaining the optimal edge line, the vertical distance d of each pixel in the optimal edge image to the optimal edge line is obtained as the abscissa, and the gray value of each pixel is the ordinate. Scatter plot.

In general, the accurate acquisition of the edge spread function (ESF) scatter plot is a necessary process for fitting the ESF curve to each pixel (i, j) in the optimal edge-edge image block (i = 1, 2, ..., m; j=1,2,...,n) to the fitted edge edge straight line ax+by+c=0 vertical distance d as the abscissa (in pixels), normalized grayscale of each pixel The value is used as the ordinate, so that the ESF scatter plot can be formed smoothly. Among them, the distance from the point to the line is calculated as:

The scatter plot is linearly fitted using the Fermi function to obtain an edge spread function.

After forming the distance-gray ESF scatter plot, the edge spread function ESF can be fitted. In view of the limitation of the edge information of the original blurred image reflected by the ESF scatter plot, the present embodiment selects to be efficient and robust to random noise. The improved Fermi function to fit the ESF, the expression is as follows:

A point spread equation (PSF) is calculated using the edge spread function, which is also a point spread function.

Deriving the edge spread function to obtain the line spread function LSF in the x direction,

Further, the line spread function LSF in the y direction can also be obtained in the above manner.

After obtaining the LSF in the x and y directions, the MTF can be obtained by Fourier transform, and then the point spread equation is obtained after convolving the MTF.

In general, the PSF model is considered to be isotropic, so the two-dimensional PSF is separable, so the PSF can also be quickly calculated by:

PSF(x,y)=LSF(x)×LSF(y)

In summary, the adaptive edge-edge method based on Gaussian model is used to estimate the fuzzy kernel, and the research shows that the degradation type of defocus blur is also coincident with the Gaussian model. Therefore, the method is active compared with the traditional edge-edge method. The optimal edge-edge image is selected as the basic parameter of fuzzy kernel estimation, and the estimated fuzzy kernel energy is closer to the real fuzzy kernel.

The reason for the motion blur is mainly caused by the relative displacement change of the pixel, and in actual use, when the motion blur is generated, the relative motion between the target scene and the imaging device is not a uniform motion, so preferably, the In step S20:

If the blurred image is motion blur, the adaptive sparse prior method is used to estimate the fuzzy kernel, and the sparsity of the image gradient domain is used as the regular constraint. Firstly, the image pyramid of the blurred image is established, and then the fuzzy kernel and the clear image optimal value of each layer of image are calculated layer by layer by the method of alternating iteration until the last layer calculates the best fuzzy kernel as the parameter used for image restoration.

Specific steps are as follows:

(1) input a blurred image g, a fuzzy kernel size m, wherein the fuzzy kernel size is a given value;

(2) determining the number of decomposition layers according to the size of the fuzzy kernel, and performing bilateral filtering on each layer of the image;

(3) The filtered image is reprocessed by the impulse filter, which can suppress the noise well and strengthen the edge information of the image;

(4) Gradient image y obtained by first order derivation;

(5) Using the Iteration Shrinkage Thresholding Algorithm (ISTA) algorithm to update the solution to the f-sub-problem, where f is the clear image optimal value of each layer of image;

(6) The unconstrained iterative Re-weighted Least Squares (IRLS) algorithm is used to solve the h sub-problem, where h is the fuzzy kernel of each layer of image;

(7) Add the fuzzy kernel h obtained by this layer to the iterative process of the next layer until the last layer obtains a relatively stable fuzzy kernel as the best fuzzy kernel.

Example 3

As shown in FIG. 3, the embodiment provides a fuzzy image synthesis processing system, which is used to provide a physical implementation basis of the method in Embodiment 2, including:

The fuzzy type identification module 100 is configured to identify a blur type of the blurred image according to the elongation of the blurred image spectrogram, where the blur type includes defocus blur and motion blur;

The estimation module 200 is configured to obtain a fuzzy kernel by using a blade edge method if the blurred image is out-of-focus blur, and obtain a blur kernel by using an adaptive preset algorithm if the blurred image is motion blur, and the preset algorithm includes at least one Algorithm

The image restoration module 300 is configured to obtain a clear image by using a super Laplacian prior deconvolution algorithm based on the fuzzy kernel.

Optionally, the system further includes:

The high-resolution processing module 400 processes the clear image by using a super-resolution reconstruction technique to obtain a high-resolution image;

The estimating module 200 includes:

The defocusing unit 210 is configured to extract an edge edge image based on a gradient criterion and obtain a point spread equation according to the optimal edge edge image if the blurred image is out of focus, and the point diffusion equation That is, the fuzzy kernel;

The motion blur unit 220 is configured to obtain a fuzzy kernel based on a sparse priori regularization method if the blurred image is motion blur.

Optionally, as shown in FIG. 4, the defocus blur unit 210 includes:

The edge detection sub-unit 211 is configured to perform Canny edge detection on the blurred image to obtain an edge detection image.

a line detection sub-unit 212, configured to perform a Hough transform on the edge detection image to acquire a step edge image;

An image intercepting sub-unit 213, configured to intercept a blade edge image of each edge centering on a center point of each edge of the step edge image, the size of the edge edge image being according to the size of the defocused image And the degree of blurring is preferably formed by a predetermined size;

An image extraction sub-unit 214, configured to calculate a gradient value r of the edge-edge image one by one, and extract an optimal edge-edge image according to the gradient value r;

a linear fitting sub-unit 215, configured to perform linear fitting on the edge in the optimal edge-edge image by using a least square method to obtain an optimal edge line;

a scatter plot sub-unit 216, configured to obtain a vertical distance d of each pixel point in the optimal edge-edge image to the optimal edge line as an abscissa, and a gray value of each pixel point is an ordinate, Forming a scatter plot;

An edge diffusion function sub-unit 217, configured to perform a linear fit on the scatter plot by using a Fermi function to obtain an edge spread function;

A calculation subunit 218 is configured to calculate a point spread equation using the edge spread function.

In the fuzzy image integration processing system provided by the embodiment, the fuzzy type recognition module 100 firstly identifies the blur type of the blurred image according to the elongation of the blurred image spectrogram; and then estimates the blur type to defocus blur by the estimation module 200 using different estimation methods. Or a blurred kernel of the blurred image of the motion blur; and then the image restoration module 300 obtains a clear image according to the restoration of the fuzzy kernel, and finally the clear image is processed by the high-resolution processing module 400 to obtain richer details and higher resolution, which is more advantageous. Identify and analyze images.

The estimation module 200 provided by the embodiment has a more accurate and fast fuzzy kernel calculation process for the blur type image with defocus blur, and can reduce human interference, avoid blindly extracting relevant parameters, and the possibility of reducing the reliability of the fuzzy kernel estimation. At the same time, based on the calculation and extraction of the derivation calculation process parameters, the evaluation accuracy and speed of the image evaluation factor modulation transfer function (MTF) can be improved, which is more suitable for the restoration of real blurred images than the prior art. Specifically, FIG. 5 and FIG. 6 are front and rear comparison images of the blurred image after the above processing, and FIG. 5 shows the integrated processing of the defocused image, wherein (a) is a blurred image, and (b) is a restored image; 6 shows the image processing of motion blur, in which (a) is a blurred image and (b) is a restored image.

The specific embodiments described herein are merely illustrative of the spirit of the invention. A person skilled in the art can make various modifications or additions to the specific embodiments described or in a similar manner, without departing from the spirit of the invention or as defined by the appended claims. The scope.

Claims

A method for comprehensively processing a blurred image, comprising the steps of:

S10: Identify a blur type of the blurred image according to the elongation of the blurred image spectrogram, where the blur type includes defocus blur and motion blur;

S20: if the blurred image is out-of-focus blur, the edge kernel method is used to obtain the blur kernel; if the blurred image is motion blur, the adaptive preset algorithm obtains the blur kernel, and the preset algorithm includes at least one algorithm;

S30: Obtain a clear image by using a super Laplacian prior deconvolution algorithm based on the fuzzy kernel.
A method for comprehensively processing blurred images according to claim 1, further comprising the steps of:

S40: processing the clear image by using a super-resolution reconstruction technique to obtain a high-resolution image.
The method of processing a blurred image according to claim 1, wherein if the blurred image is defocused in the step S20, acquiring the fuzzy kernel by using the edge method comprises the following steps:

S201: extract an optimal edge edge image based on a gradient criterion;

S202: Acquire a point spread equation according to the optimal edge edge image, where the point spread equation is the blur kernel.
The method of processing a blurred image according to claim 3, wherein the step S201 comprises the following steps:

Performing an edge detection image by performing Canny edge detection on the blurred image;

Performing a Hough transform on the edge detection image to obtain a step edge image;

Cutting a blade edge image of each edge centering on a center point of each edge in the step edge image, the size of the edge edge image selecting a preset size according to the size and blur degree of the defocused image form;

The gradient values of the edge image are calculated one by one, and the optimal edge image is extracted based on the gradient values.
A fuzzy image synthesis processing method according to claim 4, wherein said calculating a gradient value of said edge image and extracting an optimal edge image based on the gradient value comprises the following steps:

Using a least squares method to linearly fit the edges in the edge image to obtain a edge line;

Extracting an intersection formed by an edge of the edge image and a straight line of the edge as a new edge edge point;

The absolute value of the difference between the average pixel values of the two sides of the new edge edge point is counted as the gradient value of the edge edge image, and the edge edge image having the largest gradient value is extracted as the optimal edge edge image.
The method of processing a blurred image according to claim 3, wherein the step S202 comprises the following steps:

Using an least square method to linearly fit the edges in the optimal edge image to obtain an optimal edge line;

Obtaining a vertical distance of each pixel point in the optimal edge edge image to the optimal edge line as an abscissa, and a gray value of each pixel point is an ordinate to form a scatter plot;

Using the Fermi function to linearly fit the scatter plot to obtain an edge spread function;

The point spread equation is calculated using the edge spread function.
A method for comprehensively processing blurred images according to claim 1, wherein in step S20:

If the blurred image is motion blur, the adaptive sparse priori-based regularization method obtains the fuzzy kernel.
A fuzzy image integrated processing system, comprising:

a fuzzy type identification module, configured to identify a blur type of the blurred image according to an elongation of the blurred image spectrogram, the blur type including defocus blur and motion blur;

An estimation module, configured to acquire a fuzzy kernel by using a blade edge method if the blurred image is out-of-focus blur; and obtain a fuzzy kernel by using an adaptive preset algorithm if the blurred image is motion blur, the preset algorithm includes at least one algorithm;

An image restoration module is configured to obtain a clear image by using a super Laplacian prior deconvolution algorithm based on the fuzzy kernel.
A fuzzy image synthesis processing system according to claim 8, wherein the system further comprises:

a high resolution processing module that processes the sharp image using a super-resolution reconstruction technique to obtain a high resolution image;

The estimation module includes:

a defocusing unit for extracting an edge-edge image based on a gradient criterion and acquiring a point-diffusion equation according to the optimal edge-edge image if the blurred image is defocused, and the point-diffusion equation is The blur kernel;

A motion blur unit is used to acquire a fuzzy kernel based on a sparse priori regularization method if the blurred image is motion blur.
A blurred image synthesis processing system according to claim 8, wherein the defocus blur unit comprises:

An edge detection subunit, configured to perform Canny edge detection on the blurred image to obtain an edge detection image;

a line detection subunit, configured to perform a Hough transform on the edge detection image to obtain a step edge image;

An image intercepting subunit for intercepting a blade edge image of each edge centering on a center point of each edge of the step edge image, the size of the edge edge image being according to the defocused image size and The degree of blur selects a predetermined size to be formed;

An image extraction subunit, configured to calculate a gradient value of the edge edge image one by one, and extract an optimal edge edge image according to the gradient value;

a linear fitting sub-unit for linearly fitting edges in the optimal edge-edge image by using a least squares method to obtain an optimal edge line;

a scatter plot sub-unit, configured to obtain a vertical distance of each pixel point in the optimal edge-edge image to the optimal edge line as an abscissa, and a gray value of each pixel point is an ordinate, forming a dispersion Dot map

An edge diffusion function sub-unit for linearly fitting the scatter plot with a Fermi function to obtain an edge spread function;

A calculation subunit is configured to calculate a point spread equation using the edge spread function.