An optimized iterative method and system based on blurred image restoration

Abstract

The present invention discloses an image restoration optimization method and system based on a strong light background, which relates to the technical field of image restoration. The main technical points of the present invention mainly include: realizing strong light background suppression through a wavefront coding system with a built-in mask plate, and collecting a coded image after background strong light suppression of an object to be detected; using an improved L-R image restoration algorithm to restore the coded image, and combining wavelet transform for residual iteration; and using a secondary threshold to remove two main types of noise and retain weak target data; on this basis, establishing a reference-free image quality evaluation system; and using a sliding window to calculate its arithmetic mean, and at the same time setting a patience iteration mechanism to control the effective iteration times of the L-R algorithm; finally, calculating the maximum value of the image evaluation parameters in each iteration to determine the optimal iteration times and obtaining a clear target image under a strong light background. The present invention can effectively eliminate the interference of a strong light background on target imaging, shorten the algorithm iteration time, restore the detail information of the target image under a strong light background, significantly improve the quality of the target image under a strong light background, and has excellent image restoration ability.

Description

Technical Field

[0001] The present invention relates to the technical field of blurred image restoration, and particularly relates to an optimized iterative method and system for blurred image restoration. Background Art

[0002] The strong backlight detection technology has important applications in the field of optical imaging, especially playing a key role in scenarios such as high dynamic range imaging, autonomous driving, and remote sensing. When detecting a target under a strong backlight background, it may cause the loss of texture details in the detected target area or severe light saturation phenomenon, affecting the quality of the target image and the accuracy of subsequent target recognition.

[0003] The strong backlight detection technology uses optical instruments for optical imaging. By introducing wavefront coding into a traditional imaging system, the signal-to-noise ratio and image contrast of the image can be effectively improved, thereby efficiently improving the imaging quality of the target under a strong backlight background and effectively retaining the structural detail information. In recent years, due to the continuous development of optical imaging technology, the problem of image restoration in special environments has received great attention from researchers. The optical imaging technology based on wavefront coding has gradually become a research hotspot among many scholars in the field of optical imaging due to its excellent performance under complex lighting conditions.

[0004] For example, the prior art with the document number CN112801917A discloses a device and method for monitoring the rotation speed of a rotating object and restoring a blurred image based on the Laguerre-Gaussian mode. Among them, the rotating object rotation speed monitoring device includes: a rotating object optical imaging module and an object rotation speed measurement module, and the blurred image restoration device includes a rotating object rotation speed monitoring device, a Laguerre-Gaussian spectrum calculation module, a Laguerre-Gaussian spectrum correction module, and a blurred image restoration module. It can realize the rotation speed monitoring of a high-speed rotating object and the restoration processing of a rotationally blurred image, and then realize the clear imaging of a high-speed rotating object. However, it still cannot eliminate the influence of the strong light background on the target imaging system. Summary of the Invention

[0005] In view of the problems in the prior art that the influence of the strong light background on the target imaging system cannot be eliminated and the quality of the target image under the strong light background is poor, the present invention proposes an optimized iterative method and system for blurred image restoration.

[0006] The optimized iterative method for blurred image restoration proposed by the present invention includes:

[0007] Step 1: The detector collects the encoded image after suppressing the background strong light of the object to be detected: When collecting the target image under the strong light background, first use the wavefront coding optical system with an internal mask plate to suppress the strong light background, and then use the detector to collect the encoded image after suppressing the background strong light of the object to be detected;

[0008] Step 2. Restore the encoded image using the improved L - R image restoration algorithm;

[0009] Step 3. Perform residual iteration in combination with wavelet transform;

[0010] Step 4. Remove two main types of noise using a quadratic threshold and retain the dim target data;

[0011] Step 5. Establish a no - reference image quality evaluation system;

[0012] Step 6. Calculate its arithmetic mean using a sliding window, and at the same time set a patience iteration mechanism to control the effective iteration times of the L - R algorithm;

[0013] Step 7. Calculate the maximum value of the image evaluation parameters in each iteration to determine the optimal iteration times;

[0014] Step 8. Perform algorithm iteration restoration to obtain a clear target image under a strong light background.

[0015] The specific implementation process of each step is as follows:

[0016] Step 1. The detector collects the encoded image after suppressing the background strong light containing the object to be detected;

[0017] When collecting the target image under a strong light background, first use a wavefront coding system with a built - in mask plate to achieve background strong light suppression, and then use the detector to collect the encoded image after suppressing the background strong light containing the object to be detected;

[0018] Based on the fact that the target image under a strong light background is vulnerable to strong light interference, which affects the normal imaging of the detector. The optimized iterative method based on fuzzy image restoration first obtains the encoded image through the wavefront coding optical system, which is in a Poisson distribution state; then uses the maximum likelihood criterion as the standard and completes image restoration through an optimized iterative method; finally, a clear target image under a strong light background is obtained.

[0019] Step 2. Restore the encoded image using the improved L - R image restoration algorithm;

[0020] Based on the wavefront coding optical system under a strong backlight background, a linear solution scheme is adopted, and the simple Lucy - Richardson algorithm is used to solve the decoding characteristics of wavefront coding. The traditional L - R algorithm is an iterative algorithm based on Thomas Bayes theory analysis. The image collected under a strong light background passes through the wavefront coding system with a mask plate to obtain the encoded image, which is used as the input image of the image restoration algorithm. When the influence of noise is small or can be ignored, its solution is unique; when collecting images under a strong light background, noise cannot be completely eliminated, and it is necessary to suppress noise to the greatest extent.

[0021] When restoring an image, an optical imaging model indicating the presence of noise is usually adopted where G and R represent the best restored image and the original image obtained by acquisition, ξ represents the optical imaging system noise, and h represents the point spread function (PSF) of the optical system; substituting this model into the iterative calculation formula of the improved image restoration algorithm, the calculation formula can be obtained as follows

[0022]

[0023] where and represent convolution and correlation operations. Based on the improved L - R algorithm, the target image is restored in an iterative form. Therefore, it is necessary to re - define the residual α in each iteration before performing the algorithm iteration (n) , and the calculation formula is

[0024]

[0025] Step 3: Combine wavelet transform for residual iteration

[0026] Use the improved L - R algorithm for image restoration, combine two - dimensional wavelet transform, and perform residual iteration elimination on the image noise signal

[0027] The restoration algorithm contains a noise term in each iterative calculation, which will affect the quality of the subsequent restored image. Through the combination of the L - R algorithm and wavelet transform, residual secondary noise reduction is performed during the image restoration iteration process. It is known that in the wavelet domain, the amplitude distribution states of the noise term and the wavelet coefficients of the signal are different, and they are divided into wavelet coefficients converted from noise and signal. First, by setting a hard threshold (set the threshold according to the standard deviation of the noise), the noise term is reduced, and a large amount of noise existing in the image is effectively filtered to obtain the residual τ after the first noise removal (n) .

[0028] Step 4: Use a secondary threshold to remove two main types of noise and retain faint target data

[0029] In the wavelet domain, the amplitude distributions of the wavelet coefficients of noise and signal are different, mainly divided into wavelet coefficients transformed from the first - type noise and wavelet coefficients transformed from the second - type signal. The first - type noise is removed by the hard - threshold method. On this basis, the second - type noise is removed by the intermediate threshold, and threshold smoothing is performed to retain faint target data

[0030] The residual τ after the first noise removal is obtained through the above Step 3 (n) , let The iterative calculation of the algorithm after the first noise reduction can be obtained, and the calculation formula is

[0031]

[0032] The above noise removal scheme can remove a large amount of noise existing in the image and accelerate the iterative convergence of the algorithm, but it cannot further remove some of the noise existing in the wavelet coefficients converted from the signal. Therefore, residual second-stage noise reduction is introduced based on the above noise reduction algorithm, mainly aiming at the second type of partial noise doped in the signal. The relevant parameters of wavelet transform are reset, the "db4" wavelet signal is used, and its decomposition level is set to 5. At the same time, the intermediate threshold (obtained by the adaptive method) is used to perform residual second-stage noise reduction on the image data after the first-stage noise reduction. This scheme can more accurately remove a small amount of noise in the signal, and the residual ε after the second-stage noise reduction (n) The calculation formula is

[0033]

[0034] where ζ (n-1) represents the partial noise doped in the second type of signal. Substituting ζ (n-1) into the iterative equation, the calculation formula is

[0035]

[0036] Based on the improved L-R algorithm for residual second-stage noise reduction, it can efficiently remove a large amount of noise existing in the image, thereby further improving the quality of the restored image. At the same time, considering that using the quadratic threshold method to remove noise interference will affect the continuity of the image. The wavelet coefficient C Smooth is selected for threshold smoothing processing. The calculation formula is

[0037]

[0038] In the formula, C(i,j) represents the value of the wavelet coefficient at the position (i,j) in the matrix C, threshold represents the threshold set in the text; sign(C(i,j)) represents the sign function of the coefficient C(i,j).

[0039] Adjustment factors α and γ are added at the threshold connection to control the threshold segmentation degree and interval respectively. In the text, choosing α = 0.5 and γ = 1.5 can make the threshold processing softer, avoid excessive noise reduction, thereby retaining more image details and structural information, and achieving the best smoothing and noise reduction effect.

[0040] Step Five: Establish a no-reference image quality evaluation system;

[0041] Establish a no-reference image quality evaluation system. The specific process is as follows: It is known that each frame of the acquired image is composed of multiple pixels, and each pixel is correlated with each other. The image evaluation parameter uses the correlation degree between each pixel to further judge the quality and clarity of the restored image. A no-reference image quality evaluation system is selected. The three evaluation parameters can accurately evaluate the quality of the restored image from different angles, and have been widely used and verified in academic research and practical applications, with high computational efficiency and accuracy. The evaluation system indicators are as follows: The Naturalness Image Quality Evaluator (NIQE) evaluates the quality of the restored image through the Natural Scene Statistics (NSS) model, and can reliably evaluate the image quality in various scenarios. The Blind Image Integrity Notator Using DCT Statistics (BIQI) is a no-reference image quality evaluation method based on Discrete Cosine Transform (DCT) statistics, which can efficiently evaluate the quality of various image distortion types. The Perception based Image Quality Evaluator (PIQE) is a no-reference image quality evaluation method that combines spatial domain and frequency domain features and is perception-driven.

[0042] It is known that each frame of the acquired image is composed of multiple pixels, and each pixel (i, j) is correlated with each other. The image evaluation parameter uses the correlation degree between each pixel to further judge the quality and clarity of the restored image. A no-reference image quality evaluation system is selected. The three evaluation parameters can accurately evaluate the quality of the restored image from different angles, and have been widely used and verified in academic research and practical applications, with high computational efficiency and accuracy. The evaluation system indicators are as follows.

[0043] NIQE (Naturalness Image Quality Evaluator)

[0044] The Naturalness Image Quality Evaluator evaluates the quality of the restored image through the Natural Scene Statistics (NSS) model, and can reliably evaluate the image quality in various scenarios. Its calculation formula is as follows.

[0045]

[0046] In the formula, μ q and ∑ q represent the mean and covariance matrix of the image to be evaluated; μ p and ∑ p represent the mean and covariance matrix obtained by training with the natural scene image set.

[0047] BLIINDS-II (Blind Image Integrity Notator Using DCT Statistics)

[0048] Blind Image Integrity Notator Using DCT Statistics II is a reference-free image quality assessment method based on discrete cosine transform (DCT) statistics, which can efficiently evaluate the quality of various image distortion types. In this algorithm, the image block size is 8x8, and the formula for calculating the image quality score is

[0049]

[0050] where N represents the number of image blocks; w i represents the weight of the i-th image block; q i represents the quality score of the i-th image block, which can be obtained through DCT feature extraction mapping.

[0051] PIQE (Perception based Image Quality Evaluator)

[0052] Perception based Image Quality Evaluator is a perception-driven reference-free image quality assessment method that combines spatial domain and frequency domain features. It selects 16x16 image blocks, which can provide better computational efficiency and accurately capture local features of the image in image analysis. Its formula for the degree of distortion and weighted average based on image blocks is

[0053]

[0054] where N represents the number of image blocks; Q i represents the quality score of the i-th image block.

[0055] Step 6: Calculate its arithmetic mean using a sliding window, and at the same time set a patience iteration mechanism to control the effective number of iterations of the L-R algorithm;

[0056] When the optimal number of iterations is obtained within the set iteration threshold in the optimized iterative method based on fuzzy image restoration, the best restored image can be obtained. In each iteration of the algorithm, the parameter values of the statistical image quality evaluation system are calculated, and the maximum and minimum values S of the three parameters in several iterations are obtained i (i is the number of iterations), and the calculation formula is

[0057] S i =NIQE i +BLIINDS-II i +PIQE i (i = 1, 2,... N)

[0058] To avoid problems such as redundant algorithm iterations caused by blindly designed iteration times and unsatisfactory restored image quality, a feedback loop is designed during the iterative calculation of the algorithm. By designing a sliding window, setting a reasonable step size, and calculating S after each iteration i , it is necessary to further determine the dominant intervals DI (Dominant intervals) of the iterative algorithm. On this basis, a patience mechanics PM (Patience mechanics) is supplemented to limit algorithm iteration redundancy. According to multiple experiments and the distribution of parameters of the image evaluation system after different iteration times, the value of PM is generally set to N / n (when it is not an integer, round up or down). The limit parameter n can be reasonably set according to factors such as the noise level, image blurring degree, and computer resources in the actual scenario, which can effectively avoid the local optimum of the patience iteration mechanism. Then, a sliding window (size = 4) and a step size (step = 2) are used to accelerate the running rate of the feedback loop and improve the timeliness of the algorithm. The parameter calculation formula of the feedback loop is

[0059]

[0060] Step 7: Calculate the maximum value of the image evaluation parameters in each iteration and determine the optimal iteration times;

[0061] When the algorithm iteration feedback loop meets the conditions, the algorithm will continue iterative calculation. When it does not meet the conditions, it enters the patience iteration mechanism PM. When it exceeds PM = N / n, the algorithm iteration process ends, and the maximum value S of the parameters of the image quality evaluation system is searched for during several iterations MAX , if the feedback loop does not meet the conditions and does not exceed the patience iteration mechanism PM, the algorithm iterates until the originally set iteration times stop, and S is searched for within the originally set iteration times MAX , and the calculation formula is

[0062] S MAX =(NIQE i +BLlINDS-II i +PIQE i ) MAX (i = 1,2,...N)

[0063] The above algorithm iteration design completes adaptive optimal iteration and feedback. According to the optimal iteration times of the algorithm, the best restored image is output, which can effectively reduce the algorithm iteration time, improve the quality of the restored image, avoid problems such as noise amplification caused by multiple iterations, and provide a good foundation for subsequent target recognition and positioning

[0064] Step 8: Perform algorithm iteration restoration to obtain a clear target image under strong light background

[0065] According to another aspect of the present invention, an optimized iterative system based on blurred image restoration is proposed. The method includes:

[0066] Image acquisition module: It is configured to achieve strong light background suppression through a wavefront coding system with a built-in mask plate, and acquire a coded image after strong light background suppression of the object to be detected.

[0067] Residual noise reduction module: The traditional L - R algorithm is an iterative algorithm based on Thomas Bayes theory analysis. It is necessary to avoid problems such as noise amplification caused by multiple iterations of the Lucy - Richardson algorithm and unsatisfactory quality of the restored image. Due to the inability to completely eliminate the influence of external noise and other interference factors, the process of image restoration is a many - to - many ill - posed problem, which is difficult to solve for the original image. The wavefront coding optical imaging system under strong backlight uses a linear solution method and uses the L - R algorithm to solve the decoding characteristics of the wavefront coding system. First, the coded image is obtained through the point spread function in the wavefront coding system, which is in a Poisson distribution state. Then, the maximum likelihood criterion is used as the standard, and wavelet transform is combined to remove two types of noise in the residual, and finally, the image restoration is completed by an iterative method.

[0068] The target image acquired under strong backlight is combined with the point spread function model to obtain a coded image, which is used as the input end of the restoration algorithm iteration. When the noise interference is small or negligible, the solution is unique; in actual scenarios, especially under strong backlight, the noise cannot be ignored and noise removal is required. The image restoration under strong backlight combines the L - R algorithm and wavelet transform to perform secondary noise reduction on the residual.

[0069] During image restoration, the optical imaging model representing the existence of noise is usually adopted, where G and R represent the best restored image and the original image acquired, ξ represents the optical imaging system noise, and h represents the point spread function (PSF) of the optical system; substituting this model into the iterative calculation formula of the improved image restoration algorithm, the calculation formula is

[0070]

[0071] where and represent convolution and correlation operations. Based on the improved L - R algorithm, the target image is restored in an iterative form, so it is necessary to re - define the residual α in each iteration before the algorithm iteration (n) , and the calculation formula is

[0072]

[0073] The restoration algorithm contains noise terms in each iterative calculation, which will affect the quality of the subsequent restored image. The residual quadratic noise reduction is carried out in the image restoration iteration process by combining the L-R algorithm with wavelet transform. It is known that in the wavelet domain, the amplitude distribution states of the noise terms and the wavelet coefficients of the signal are different, and they are divided into wavelet coefficients converted from noise and signal. First, by setting a hard threshold (set according to the standard deviation of the noise), the noise terms are reduced, and a large amount of noise existing in the image is effectively filtered out to obtain the residual τ after the first noise removal. (n) , let The iterative calculation of the algorithm after the first noise reduction can be obtained, and the calculation formula is

[0074]

[0075] The above noise removal scheme can remove a large amount of noise existing in the image and accelerate the iterative convergence of the algorithm, but it cannot further remove some of the noise existing in the wavelet coefficients converted from the signal. Therefore, residual second noise reduction is introduced on the basis of the above noise reduction algorithm, mainly aiming at the part of the noise doped in the second type of signal. The relevant parameters of the wavelet transform are reset, and the "db4" wavelet signal is used, and its decomposition level is set to 5; at the same time, the scheme of using the intermediate threshold (obtained by the adaptive method) is used to perform residual second noise reduction on the image data after the first noise reduction. This scheme can more accurately remove a small amount of noise in the signal, and the residual ε after the second noise reduction (n) The calculation formula is

[0076]

[0077] Among them, ζ (n-1) represents the part of the noise doped in the second type of signal. Substituting ζ (n-1) into the iterative equation, the calculation formula can be obtained as

[0078]

[0079] Based on the residual second noise reduction of the improved L-R algorithm, a large amount of noise existing in the image can be efficiently removed, thereby further improving the quality of the restored image. At the same time, considering that using the quadratic threshold method to remove noise interference will affect the continuity of the image. The wavelet coefficient C Smooth is selected for threshold smoothing processing, and the calculation formula is

[0080]

[0081] In the formula, C(i,j) represents the value of the wavelet coefficient at the position (i,j) in the matrix C, threshold represents the threshold set in the text; sign(C(i,j)) represents the sign function of the coefficient C(i,j).

[0082] Adjustment factors α and γ are added at the junction of the set thresholds to control the threshold segmentation degree and interval respectively. In this paper, choosing α = 0.5 and γ = 1.5 can make the threshold processing softer, avoid excessive noise reduction, retain more image details and structural information, and achieve the best smoothing and noise reduction effect.

[0083] (3) Adaptive optimization iteration module

[0084] It is configured to obtain the optimal number of iterations within the set iteration threshold and obtain the best restored image. In each iteration process of the algorithm, the parameter values of the statistical image quality evaluation system are calculated, and the maximum values S of the three parameters in several iterations are obtained i (i is the number of iterations), and the calculation formula is

[0085] S i =NIQE i +BLIINDS-II i +PIQE i (i = 1, 2,...N)

[0086] To avoid problems such as redundant algorithm iteration and unsatisfactory restored image quality caused by blindly designing the number of iterations, a feedback link is designed in the iterative calculation of the algorithm. By designing a sliding window, setting a reasonable step size, and calculating S after each iteration i , it is necessary to further determine the dominant interval DI (Dominant intervals) of the iterative algorithm. On this basis, a patience iteration mechanism PM (Patience mechanics) is supplemented to limit the redundant algorithm iteration. According to multiple experiments and the distribution state of the parameters of the image evaluation system after different numbers of iterations, the value of PM is generally set to N / n (rounding up or down if it is not an integer). The limit parameter n can be reasonably set according to factors such as the noise level, image blur degree, and computer resources in the actual scenario, which can effectively avoid the local optimum of the patience iteration mechanism. Then, a sliding window (size = 4) and a step size (step = 2) are used to accelerate the running speed of the feedback link and improve the timeliness of the algorithm. The parameter calculation formula of the feedback link is

[0087]

[0088] When the algorithm iteration feedback link meets the conditions, the algorithm will continue to perform iterative calculations. When it does not meet the conditions, it enters the patience iteration mechanism PM. When it exceeds PM = N / n, the algorithm iteration process ends, and the maximum value S of the parameters of the image quality evaluation system in several iterations is searched for MAX , if the feedback link does not meet the conditions and does not exceed the patience iteration mechanism PM, the algorithm iterates until the originally set number of iterations stops, and S in the originally set number of iterations is searched for MAX , and the calculation formula is

[0089] S MAX =(NIQE i +BLlINDS-II i +PIQE i ) MAX (i = 1, 2,... N)

[0090] The adaptive optimal iteration and feedback are completed by the above algorithm iteration design. According to the optimal iteration times of the algorithm, the best restored image is output, which can effectively reduce the algorithm iteration time, improve the quality of the restored image, avoid problems such as noise amplification caused by multiple iterations, and provide a good foundation for subsequent target recognition and positioning.

[0091] The beneficial effects of the present invention are as follows:

[0092] The image restoration optimization method and system for strong light background proposed by the present invention first realizes strong light background suppression through a wavefront coding system with a built-in mask plate, and acquires the encoded image after strong light background suppression of the background containing the object to be detected; uses the improved L-R image restoration algorithm to restore the encoded image, combines wavelet transform for residual iteration; and uses a quadratic threshold to remove two types of main noises and retain the dim target data; on this basis, a no-reference image quality evaluation system is established; and the arithmetic mean value is calculated by using a sliding window, and at the same time, a patience iteration mechanism is set to control the effective iteration times of the L-R algorithm; finally, the maximum value of the image evaluation parameters in each iteration is calculated to determine the optimal iteration times, and a clear target image under strong light background is obtained. The present invention can effectively eliminate the interference of strong light background on target imaging, shorten the algorithm iteration time, restore the detail information of the target image under strong light background, significantly improve the quality of the target image under strong light background, and has excellent image restoration ability.

[0093] The advantages of the present invention are as follows:

[0094] (1) The wavefront coding system based on the built-in mask plate has an inhibitory effect on strong light, and can effectively eliminate the influence of strong light background on the target imaging system.

[0095] (2) The image restoration algorithm based on strong light background combines wavelet transform for residual iteration and uses a quadratic threshold denoising method, which can effectively remove two types of main noises in the image, effectively retain the dim target data information, restore the detail information of the target image under strong light background, and significantly improve the quality of the target image under strong light background.

[0096] (3) The optimized iterative algorithm for image restoration based on strong light background establishes a no-reference image quality evaluation system, calculates the arithmetic mean using a sliding window, sets a patience iterative mechanism to control the effective number of iterations of the L-R algorithm, calculates the optimal number of iterations, shortens the algorithm iteration time, has excellent image restoration ability, and provides a good foundation for the subsequent theoretical research on target recognition and positioning. Description of the Drawings

[0097] The present invention can be better understood by referring to the description given in conjunction with the accompanying drawings below. The accompanying drawings, together with the following detailed description, are included in this specification and form a part of this specification, and are used to further illustrate the preferred embodiments of the present invention and explain the principles and advantages of the present invention.

[0098] Figure 1 It is a flowchart of an optimized iterative method for fuzzy image restoration according to an embodiment of the present invention;

[0099] Figure 2 It is an overall framework diagram of a wavefront coding optical imaging system based on a strong backlight background in an embodiment of the present invention;

[0100] Figure 3 It is a comparative analysis diagram of a group of classical image restoration algorithms in an embodiment of the present invention;

[0101] Among them: (a) is the inverse filtering algorithm, (b) is the Wiener filtering algorithm, (c) is the blind deconvolution algorithm, and (d) is the improved L-R restoration algorithm based on fuzzy images;

[0102] Figure 4 It is a parameter analysis diagram of an optimized iterative no-reference evaluation system for image restoration in an embodiment of the present invention;

[0103] Among them: (a) is the NIQE parameter, (b) is the BLIINDS-II parameter, and (c) is the PIQE parameter. Detailed Embodiments

[0104] In order to enable those skilled in the art to better understand the solution of the present invention, the exemplary embodiments or examples of the present invention will be described below in conjunction with the attached Figures 1-4 Obviously, the described embodiments or examples are only a part of the embodiments or examples of the present invention, rather than all of them. All other embodiments or examples obtained by those of ordinary skill in the art based on the embodiments or examples of the present invention without creative work shall fall within the scope of protection of the present invention.

[0105] An embodiment of the present invention provides an optimized iterative method for fuzzy image restoration, as Figure 1As shown, the method includes the following steps:

[0106] Step 1: Implement strong light background suppression through a wavefront coding system with a built-in mask, and collect the encoded image after strong light background suppression of the background containing the object to be detected;

[0107] Step 2: Restore the encoded image using the improved L-R image restoration algorithm;

[0108] Step 3: Perform residual iteration in combination with wavelet transform;

[0109] Step 4: Remove two types of main noises using a quadratic threshold, and retain the dim target data;

[0110] Step 5: Establish a no-reference image quality evaluation system;

[0111] Step 6: Calculate its arithmetic mean using a sliding window, and at the same time set a patience iteration mechanism to control the effective iteration times of the L-R algorithm;

[0112] Step 7: Calculate the maximum value of the image evaluation parameters in each iteration to determine the optimal iteration times;

[0113] Step 8: Perform algorithm iteration restoration to obtain a clear target image under strong light background.

[0114] The method of the present invention starts from Step 1. In Step 1, strong light background suppression is achieved through a wavefront coding system with a built-in mask, and the encoded image after strong light background suppression of the background containing the object to be detected is collected;

[0115] According to an embodiment of the present invention, the image collected under strong light background is encoded after passing through a wavefront coding system with an added mask to obtain an encoded image.

[0116] Then Step 2 is executed. In Step 2, the collected encoded image is restored using the improved L-R image restoration algorithm, as Figure 2 shown;

[0117] According to an embodiment of the present invention, a linear solution scheme is adopted for the wavefront coding optical system under strong backlight background, and the simple Lucy-Richardson algorithm is used to solve the decoding characteristics of wavefront coding. The traditional L-R algorithm is an iterative algorithm based on Thomas Bayes theory analysis. First, the blurred image obtained after passing through the wavefront coding optical system is in a Poisson distribution state, and the maximum likelihood criterion is used as the standard, and then the image restoration is completed by an iterative method.

[0118] When restoring an image, usually An optical imaging model indicating the presence of noise, where G and R represent the best restored image and the original image acquired, ξ represents the optical imaging system noise, and h represents the point spread function (PSF) of the optical system; substituting this model into the iterative calculation formula of the improved image restoration algorithm, the calculation formula is,

[0119]

[0120] where, and represent convolution and correlation operations.

[0121] Then perform step 3. In step 3, use the improved L - R algorithm for image restoration, combined with two - dimensional wavelet transform, to perform residual iterative elimination on the image noise signal.

[0122] According to an embodiment of the present invention,

[0123] Based on the target image acquired under a strong backlight background, combined with the point spread function model, an encoded image is obtained, which is used as the input end of the restoration algorithm iteration. When the noise interference is small or negligible, the solution is unique; in actual scenarios, especially under a strong backlight background, the noise cannot be ignored and noise removal is required. Image restoration under a strong backlight background uses the combination of the L - R algorithm and wavelet transform for residual secondary noise reduction.

[0124] Based on the improved L - R algorithm, the target image is restored in an iterative form, so it is necessary to re - define the residual α in each iteration before performing the algorithm iteration (n) , and the calculation formula is,

[0125]

[0126] Then perform step 4. In step 4, use quadratic thresholding to remove two main types of noise and perform smoothing processing to retain the dim target data;

[0127] According to an embodiment of the present invention, the restoration algorithm contains a noise term in each iterative calculation, which will affect the quality of the subsequent restored image. Residual secondary noise reduction is performed during the image restoration iteration process by combining the L - R algorithm and wavelet transform. It is known that in the wavelet domain, the noise term and the wavelet coefficient amplitude distribution state of the signal are different, and they are divided into wavelet coefficients converted from noise and signal. First, by setting a hard threshold (set the threshold according to the standard deviation of the noise), the noise term is reduced, and a large amount of noise existing in the image is effectively filtered to obtain the residual τ after the first noise removal (n) , let The algorithm iterative calculation after the first noise reduction can be obtained, and the calculation formula is,

[0128]

[0129] The above noise removal scheme can remove a large amount of noise existing in the image and accelerate the iterative convergence of the algorithm, but it cannot further remove some of the noise existing in the wavelet coefficients converted from the signal. Therefore, residual second-stage noise reduction is introduced based on the above noise reduction algorithm, mainly aiming at the second type of partial noise doped in the signal. The relevant parameters of wavelet transform are reset, the "db4" wavelet signal is used, and its decomposition level is set to 5. At the same time, the residual second-stage noise reduction is performed on the image data after the first-stage noise reduction by adopting the scheme of intermediate threshold (obtained by the adaptive method). This scheme can more accurately remove a small amount of noise in the signal, and the residual ε (n) The calculation formula is

[0130]

[0131] where ζ (n-1) represents the partial noise doped in the second type of signal. Substituting ζ (n-1) into the iterative equation, the calculation formula is

[0132]

[0133] Based on the residual second-stage noise reduction of the improved L-R algorithm, a large amount of noise existing in the image can be efficiently removed, thereby further improving the quality of the restored image. At the same time, considering that using the quadratic threshold method to remove noise interference will affect the continuity of the image. The wavelet coefficient C Smooth is selected for threshold smoothing processing, and the calculation formula is

[0134]

[0135] In the formula, C(i, j) represents the value of the wavelet coefficient at the position (i, j) in the matrix C, threshold represents the threshold set in the text; sign(C(i, j)) represents the sign function of the coefficient C(i, j).

[0136] Adjustment factors α and γ are added at the threshold connection to control the threshold segmentation degree and interval respectively. Selecting α = 0.5 and γ = 1.5 in the text can make the threshold processing softer, avoid excessive noise reduction, retain more image details and structural information, achieve the best smoothing and noise reduction effect, perform updated estimation, and reconstruct the denoised data.

[0137] Then step 5 is executed. In step 5, a no-reference image quality evaluation system is established;

[0138] According to the embodiments of the present invention, it is known that each captured image frame is composed of multiple pixel points, and each pixel point (i, j) is correlated. The image evaluation parameter uses the correlation degree between each pixel point to further judge the quality and clarity of the restored image. By selecting a no-reference image quality evaluation system, the three evaluation parameters can accurately evaluate the quality of the restored image from different perspectives, and have been widely used and verified in academic research and practical applications, with high calculation efficiency and accuracy. The evaluation system indicators are as follows.

[0139] NIQE (Naturalness Image Quality Evaluator)

[0140] The Natural Image Quality Evaluator evaluates the quality of the restored image through the Natural Scene Statistics (NSS) model and can reliably evaluate the image quality in various scenarios. Its calculation formula is as follows.

[0141]

[0142] In the formula, μ q and ∑ q represent the mean and covariance matrix of the image to be evaluated; μ p and ∑ p represent the mean and covariance matrix obtained by training with the natural scene image set.

[0143] BLIINDS-II (Blind Image Integrity Notator Using DCT Statistics)

[0144] The Blind Image Integrity Notator Using DCT Statistics II is a no-reference image quality evaluation method based on Discrete Cosine Transform (DCT) statistics and can efficiently evaluate the quality of various image distortion types. In this algorithm, the image block size is 8x8, and its image quality score calculation formula is as follows.

[0145]

[0146] In the formula, N represents the number of image blocks; w i represents the weight of the i-th image block; q i represents the quality score of the i-th image block, which can be obtained through DCT feature extraction mapping.

[0147] PIQE (Perception based Image Quality Evaluator)

[0148] The perceptual image quality evaluator is a perception-driven no-reference image quality evaluation method that combines spatial domain and frequency domain features. It selects 16x16 image blocks, which can provide better computational efficiency in image analysis and can accurately capture the local features of the image. Its distortion degree and weighted average calculation formula based on image blocks are as follows:

[0149]

[0150] In the formula, N represents the number of image blocks; Q i represents the quality score of the i-th image block.

[0151] Then step 6 is executed. In step 6, the arithmetic mean is calculated using a sliding window, and at the same time, a patience iteration mechanism is set to control the effective number of iterations of the L-R algorithm;

[0152] According to the embodiments of the present invention, the optimal number of iterations is obtained within the set iteration threshold to obtain the best restored image. In each iteration process of the algorithm, the parameter values of the statistical image quality evaluation system are calculated, and the maximum and minimum values S i (i is the number of iterations) are obtained. The calculation formula is as follows:

[0153] S i =NIQE i +BLIINDS-II i +PIQE i (i = 1, 2,...N)

[0154] To avoid problems such as redundant algorithm iterations and unsatisfactory restored image quality caused by blindly designing the number of iterations, a feedback link is designed during the iterative calculation of the algorithm. By designing a sliding window and setting a reasonable step size, S i after each iteration is calculated. It is necessary to further determine the dominant intervals DI (Dominant intervals) of the iterative algorithm's advantages. On this basis, a patience iteration mechanism PM (Patience mechanics) is supplemented to limit the algorithm iteration redundancy. According to multiple experiments and the distribution status of the parameters of the image evaluation system after different numbers of iterations, the value of PM is generally set to N / n of the total number of iterations (when it is not an integer, rounding is used). The limit parameter n can be reasonably set according to factors such as the noise level, image blur degree, and computer resources in the actual scenario, which can effectively avoid the local optimum of the patience iteration mechanism. Then a sliding window (size = 4) and a step size (step = 2) are used to accelerate the operation rate of the feedback link and improve the timeliness of the algorithm. The parameter calculation formula of the feedback link is as follows:

[0155]

[0156] Then, step 7 is executed. In step 7, the maximum value of the image evaluation parameter in each iteration is calculated to determine the optimal number of iterations.

[0157] According to an embodiment of the present invention, when the algorithm iteration feedback link meets the conditions, the algorithm will continue to perform iterative calculations. When the conditions are not met, it enters the patience iteration mechanism PM. When it exceeds PM = N / n, the algorithm iteration process ends, and the maximum value S of the parameters of the image quality evaluation system during several iterative processes is found. MAX If the feedback link does not meet the conditions and does not exceed the patience iteration mechanism PM, the algorithm iterates until the originally set number of iterations stops, and S during the originally set number of iterations is found. MAX The calculation formula is S MAX =(NIQE i +BLlINDS-II i +PIQE i )(i = 1, 2,... N) MAX

[0158] Then, step 8 is executed. In step 8, algorithm iteration restoration is performed to obtain a clear target image under a strong light background, as Figure 3 shown.

[0159] According to an embodiment of the present invention, the above algorithm iteration design completes adaptive optimal iteration and feedback. According to the optimal number of algorithm iterations, the improved L-R image restoration optimization iteration method is used to complete image iteration restoration, and the best restored image is output, which can effectively reduce the algorithm iteration time, improve the quality of the restored image, avoid problems such as noise amplification caused by multiple iterations, and provide a good foundation for subsequent target recognition and positioning, as Figure 3 shown.

[0160] In summary, the present invention studies problems such as unclear target imaging, poor quality of restored images, and difficult acquisition of target detail information under a strong light background, solves the problem of strong light interference in target imaging under a strong light background, effectively reduces the algorithm iteration time by using the improved L-R image restoration optimization iteration algorithm, improves the quality of the restored image, and avoids problems such as noise amplification caused by multiple iterations.

[0161] Another embodiment of the present invention proposes an optimized iterative system based on fuzzy image restoration. The system includes:

[0162] (1) Image acquisition module: It is configured to implement strong light background suppression through a wavefront coding system with a built-in mask plate and acquire the encoded image after strong light background suppression of the object to be detected.

[0163] (2) Residual denoising module: It is configured to obtain a coded image through the point spread function in the wavefront coding system, which is in a Poisson distribution state. Then, using the maximum likelihood criterion as the standard, it combines wavelet transform to remove two types of noise from the residual, and finally uses the adaptive optimization iteration method to complete image restoration.

[0164] (3) Adaptive optimization iteration module: It is configured to establish a reference-free image quality evaluation system; and use a sliding window to calculate its arithmetic mean. At the same time, a patience iteration mechanism is set to control the effective iteration times of the L-R algorithm, calculate the optimal iteration times, and complete image restoration through the restoration algorithm.

[0165] The function of the optimized iteration system based on blurred image restoration described in this embodiment can be illustrated by the aforementioned optimized iteration method for blurred image restoration. Therefore, for the parts not detailed in this embodiment, reference can be made to the above method embodiments, which will not be elaborated here.

[0166] Verification of the technical effects of the present invention:

[0167] To test that the wavefront coding optical imaging system with an internal phase plate has a good effect on suppressing strong light and complete the restoration of the target image in complex situations to obtain a restored image with high quality. In this paper, a wavefront coding optical imaging system based on a strong light background is built to suppress strong light interference; and residual secondary noise reduction is performed based on the improved L-R algorithm to obtain a better restoration effect. On this basis, to obtain the best restored image among the effective iteration times, improve the algorithm iteration redundancy, determine the optimal iteration times, and a self-adaptive iterative image restoration algorithm under a strong backlight background is proposed. Through a reference-free image quality evaluation system, image quality evaluation indicators such as NIQE, BLIINDS-II, and PIQE are selected as parameters to limit the algorithm iteration times. It is known that the smaller the above evaluation parameter values, the better the image quality and the richer the detail information; find the optimal iteration times among the set total iteration times, output the optimal restored image, and conduct a physical experiment to verify the effectiveness of the algorithm in this paper. In this simulation experiment, a real backlight environment image is selected as the restoration object, and Gaussian noise with a mean of 0 and a variance of 0.005 is added to simulate the noise existing in the strong backlight environment collected by the optical device. Combining with the improved L-R image restoration algorithm based on wavelet transform and using the "db4" wavelet signal, its decomposition level is set to 5, and the experimental object can be better restored.

[0168] When conducting desktop physical experiments, a homogeneous integrating sphere light source (illuminance value at the light output port is 54500 lx) is used as the solar light source, and the target is placed at the light intake port to simulate the imaging effect of the target under a strong backlight background. The adaptive optimal iterative image restoration algorithm uses cyclic iteration, starting from a low order and iterating to a high order successively. Each time an iteration is performed, the parameter values of the image quality evaluation system are recorded. The total number of iterations set this time is 400 times. Analyze the changes in each parameter as the number of iterations is superimposed, as shown in Figure 4 shown,

[0169] In the reference-free image quality evaluation system, parameters such as the Natural Image Quality Evaluator (NIQE), the Blind Image Integrity Evaluator II (BLIINDS-II) based on DCT statistics, and the Perceptual Image Quality Evaluator (PIQE) will gradually decrease, and the quality of the restored image will gradually become clear; when the iteration reaches a certain number of times (85 times), after the three parameters reach the extreme values, which is the expected optimal number of iterations of the algorithm in this paper, the image quality is the best at this time, and the target clear image with the optimal number of iterations is output.

[0170] Although the present invention has been described based on a limited number of embodiments, those skilled in the art in this technical field will understand that other embodiments can be envisioned within the scope of the present invention thus described. For the scope of the present invention, the disclosure made for the present invention is illustrative rather than restrictive, and the scope of the present invention is defined by the appended claims.

Claims

1. An optimization iterative method based on fuzzy image restoration, characterized in that: The method comprises: Step 1: The detector collects the coded image after the background strong light suppression including the object to be detected: When collecting the target image under the strong light background, firstly, a wavefront coding optical system with a built-in mask plate is used to suppress the strong light background, and then the detector is used to collect the coded image after the background strong light suppression including the object to be detected; Step 2: Use the improved LR image restoration algorithm to restore the encoded image; Step 3: Perform residual iteration in combination with wavelet transform; Step 4: Use secondary threshold to remove two types of main noise and retain the dim target data; Step 5: Establish a no-reference image quality evaluation system; Step 6: Use a sliding window to calculate the arithmetic mean, and set a patient iteration mechanism to control the effective number of iterations of the LR algorithm; Step 7: Calculate the maximum value of the image evaluation parameter in each iteration and determine the optimal number of iterations; Step 8: Perform algorithm iteration and restoration to obtain a clear target image under a strong light background; The specific implementation of step 2 is: Based on the linear solution of the wavefront coding optical system under strong backlight background, the decoding characteristics of the wavefront coding are solved using a simple Lucy-Richardson algorithm. The image collected under strong light background is passed through the wavefront coding system with a mask plate added to obtain a coded image, which is used as the input image of the image restoration algorithm; when the influence of noise can be ignored, its solution is unique; when collecting images under strong light background, the noise cannot be completely eliminated and needs to be suppressed to the greatest extent; When restoring the image, we use represents the optical imaging model with noise, where G and R represent the best restored image and the original image acquired, ξ represents the noise of the optical imaging system, and h represents the point spread function (PSF) of the optical system; substituting this model into the iterative calculation formula of the improved image restoration algorithm, the calculation formula is: in, and Represents convolution and correlation operations. Based on the improved LR algorithm, the target image is restored in an iterative form. Therefore, it is necessary to redefine the residual α in each iteration before iterating the algorithm. (n) , the calculation formula is:

2. The optimization iterative method based on fuzzy image restoration according to claim 1, characterized in that: In step 1, the coded image obtained after being processed by the wavefront coding optical system is in a Poisson distribution state.

3. The optimization iterative method based on fuzzy image restoration according to claim 1, characterized in that: In step three, the improved LR algorithm is used to restore the image, and combined with the two-dimensional wavelet transform, the residual error of the image noise signal is iteratively eliminated; The restoration algorithm contains noise terms in each iterative calculation, which will affect the quality of the image after subsequent restoration. The residual secondary denoising is performed in the image restoration iterative process by combining the LR algorithm with the wavelet transform. The noise term and the wavelet coefficient amplitude distribution of the signal are different in the wavelet domain, and the wavelet coefficients are divided into noise and signal conversion. By setting a hard threshold, that is, setting the threshold according to the standard deviation of the noise, the noise term is reduced, and a large amount of noise in the image is effectively filtered out, and the residual τ after the first noise removal is obtained. (n) .

4. The optimization iterative method based on fuzzy image restoration according to claim 3, characterized in that: The specific process of step 4 is as follows: in the wavelet domain, the wavelet coefficients of noise and signal have different amplitude distributions, which are mainly divided into the wavelet coefficients of the first type of noise transformation and the wavelet coefficients of the second type of signal transformation; the first type of noise is removed by hard thresholding, and on this basis, the second type of noise is removed by using the intermediate threshold, and threshold smoothing is performed to retain the dim target data; Through the above step three, we can get the residual τ after the first noise removal. (n) ,make The algorithm iteration calculation after the first denoising can be obtained, and the calculation formula is: Based on the above denoising algorithm, the residual second denoising is introduced, mainly for the second type of partial noise mixed in the signal, the relevant parameters of the wavelet transform are reset, the "db4" wavelet signal is used, and its decomposition level is set to 5; at the same time, the intermediate threshold (obtained by the adaptive method) is used to perform residual secondary denoising on the image data after the first denoising. This scheme can more accurately remove a small amount of noise in the signal. The residual ε after the secondary denoising is (n) The calculation formula is, Among them, (n-1) represents the part of noise mixed in the second type of signal, and ζ (n-1) Substituting into the iterative equation, the calculation formula is: Based on the improved LR algorithm residual secondary denoising, a large amount of noise in the image can be effectively removed, thereby further improving the quality of the restored image; at the same time, considering that the use of the secondary threshold method to remove noise interference will affect the continuity of the image; the wavelet coefficient C is selected Smooth Perform threshold smoothing, such as the calculation formula is: In the formula, C(i,j) represents the value of the wavelet coefficient at position (i,j) in the matrix C, threshold represents the set threshold; sign(C(i,j)) represents the sign function of the coefficient C(i,j); Adjustment factors α and γ are added at the threshold setting point to control the threshold segmentation degree and interval respectively. Choosing appropriate α and γ can make the threshold processing softer, avoid excessive noise reduction, thereby retaining more image details and structural information, and achieve the best smooth noise reduction effect.

5. The optimization iterative method based on fuzzy image restoration according to claim 4, characterized in that: In step five, a no-reference image quality evaluation system is established, and the specific process is as follows: it is known that each frame of the acquired image is composed of multiple pixels, and each pixel (i, j) is correlated. The image evaluation parameters use the degree of correlation between the pixels to further judge the quality and clarity of the restored image. A no-reference image quality evaluation system is selected, and its evaluation system indicators are: the natural image quality evaluator evaluates the quality of the restored image through the natural scene statistical NSS model, and can reliably evaluate the image quality in a variety of scenes; the blind image integrity evaluator II based on DCT statistics is a no-reference image quality evaluation method based on discrete cosine transform DCT statistics, which can efficiently evaluate the quality of various image distortion types; the perceptual image quality evaluator is a no-reference image quality evaluation method that combines spatial domain and frequency domain features and is perception-driven.

6. The optimization iterative method based on fuzzy image restoration according to claim 5, characterized in that: In step 6, a sliding window is used to calculate the arithmetic mean, and a patient iteration mechanism is set to control the effective number of iterations of the LR algorithm. The specific process is as follows: The optimal iterative method based on fuzzy image restoration can obtain the best restored image when the optimal number of iterations is obtained within the set iteration threshold. In each iteration, the algorithm performs statistical analysis on the parameter values ​​of the image quality evaluation system to obtain the maximum values ​​of the three parameters in several iterations. i , i is the number of iterations, and the calculation formula is, S i =NIQE i +BLINDS-II i +PIQE i i=1,2,...N When performing iterative calculations, a feedback link is designed to calculate S after each iteration by designing a sliding window and setting a reasonable step size. i , it is necessary to further determine the advantage interval DI of the iterative algorithm, and on this basis, supplement it with a patient iteration mechanism PM to limit the algorithm iteration redundancy; according to the distribution state of the image evaluation system parameters after multiple experiments and different numbers of iterations, the value of PM is generally set to the total number of iterations N / n, and rounding is adopted when it is less than an integer. The limit parameter n can be reasonably set according to the noise size, image blur and computer resource factors in the actual scene, which can effectively avoid the local optimum of the patient iteration mechanism; then the sliding window size = 4 and the step size step = 2 are used to accelerate the running rate of the feedback link and improve the timeliness of the algorithm. The parameter calculation formula of the feedback link is, 7. The optimization iterative method based on fuzzy image restoration according to claim 6, characterized in that: In step 7, when the algorithm iteration feedback link meets the conditions, the algorithm will continue to iterate. When the conditions are not met, it will enter the patient iteration mechanism PM. When PM=N / n is exceeded, the algorithm iteration process ends and the maximum value S of the parameter of the image quality evaluation system in the process of several iterations is found. MAX If the feedback link does not meet the conditions and does not exceed the patient iteration mechanism PM, the algorithm will iterate to the original set number of iterations and stop, and find S in the original set number of iterations. MAX , the calculation formula is, S MAX =(NIQE i +BLlINDS-ll i +PIQE i ) MAX i=1,2,...N The above algorithm iterative design completes the adaptive optimal iteration and feedback, and outputs the best restored image according to the optimal number of algorithm iterations.

8. An optimization iterative system based on fuzzy image restoration, characterized in that: The system has a program module corresponding to the steps of any one of claims 1 to 7 above, and executes the steps in the optimization iterative method based on fuzzy image restoration when running; It includes: An image acquisition module corresponding to step 1: configured to realize strong light background suppression through a wavefront coding system with a built-in mask plate, and to acquire a coded image containing the strong light suppressed background of the object to be detected; The residual denoising module corresponding to steps 2 to 4 is configured to adopt a linear solution method based on the wavefront coding optical imaging system under a strong backlight background, and use the LR algorithm to solve the decoding characteristics of the wavefront coding system; first, the coded image is obtained through the point spread function in the wavefront coding system, which is in a Poisson distribution state, and then the maximum likelihood criterion is used as a standard, combined with wavelet transform to perform residual removal of two types of noise, and smoothing processing is performed, and finally the image restoration is completed by an iterative method; The adaptive optimization iteration module corresponding to steps five to eight is configured to calculate the arithmetic mean using a sliding window and set a patient iteration mechanism to control the effective number of iterations of the LR algorithm; finally, the maximum value of the image evaluation parameter in each iteration is calculated to determine the optimal number of iterations to obtain a clear target image under a strong light background.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the optimization iterative method based on blurred image restoration according to any one of claims 1 to 7 when called by a processor.

Images