A non-uniformity correction method for asphalt infrared images based on dictionary learning
Through dictionary learning-based methods, a complete dictionary was constructed and sparse representation was performed, which solved the problems of loss of detail and poor noise correction effects in asphalt infrared image non-uniformity correction, and achieved efficient non-uniformity correction effects.
Patent Information
- Application Number
- CN202210505562.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-10
- Publication Date
- 2025-04-29
- Estimated Expiration
- 2042-05-10
AI Technical Summary
The prior art has problems such as loss of detail, poor noise correction effect, and dependence on large amounts of data in the non-uniformity correction of asphalt infrared images.
Using a dictionary-based learning method, a complete dictionary was constructed and sparse representation was performed to achieve inhomogeneity correction through dictionary initialization, OPM algorithm update and Ljung-box noise inspection.
Effectively removes non-uniform noise in asphalt infrared images, protects the details of the image, reduces the demand for large amounts of data, and improves correction effect and accuracy.
Smart Images
Figure CN115147292B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of infrared thermal imaging detection, and in particular to an asphalt infrared image non-uniformity correction method based on dictionary learning. Background Art
[0002] Asphalt binder in asphalt mixture is a thermoplastic material with high temperature sensitivity. Its morphology varies greatly under different temperature conditions. It has significant viscous flow deformation characteristics at high temperatures, and has solid properties at low temperatures, showing strong deformation resistance. Therefore, the construction quality of asphalt pavement depends largely on the control of pavement temperature during construction. Uneven asphalt pavement temperature will directly affect the compaction, strength, density and service life of the asphalt surface layer. Infrared thermal imaging technology has gradually been used in temperature separation detection during asphalt pavement construction due to its advantages such as non-contact temperature measurement, small instrument size, weak environmental impact and fast response speed. However, due to the manufacturing process, the photoelectric response rate of each detection pixel on the focal plane is different, which will cause non-uniform noise in the infrared image.
[0003] There are three main methods for infrared image non-uniformity correction. The first is reference-based non-uniformity correction. This method uses the uniform radiation pattern of a black body as a reference to calculate the non-uniformity correction coefficient. Although the calibration technology can quickly calculate the correction coefficient for the non-uniformity of the current noise pattern of the infrared focal plane, the non-uniformity of the infrared focal plane has the nature of time-drifting. When the operating conditions of the infrared detector change, such as changes in the focal plane temperature, changes in the reading circuit voltage, etc., the response coefficient of each detection pixel will change, resulting in the non-uniformity correction coefficient originally calculated by the calibration technology becoming no longer applicable. The second is statistical model-based non-uniformity correction. This method was first based on time domain high-pass filtering to achieve non-uniformity correction. It is believed that scene information changes belong to high-frequency information, while fixed noise patterns are basically unchanged, and time domain low-frequency information, and then a high-pass filter is designed to achieve non-uniformity correction. However, this method is very dependent on scene motion and requires the accumulation of a large number of image frames at each pixel position. The grayscale value, so the correction speed is very slow. The third is the non-uniformity correction based on neural network. This method uses deep learning methods for non-uniformity correction and has the advantages of high precision and strong robustness. However, due to the lack of training data sets with different noise types, the generalization of this method is poor.
[0004] In summary, the existing dictionary learning methods for infrared image non-uniformity correction have the following main problems: 1) In the process of asphalt infrared image acquisition, there is a lack of clear images for supervision. 2) The details and textures in asphalt images are complex, and the existing methods have the problem of detail loss. 3) The fixed tolerance in the sparse representation process will lead to a decrease in the noise reduction effect. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides an asphalt infrared image non-uniformity correction method based on dictionary learning, the purpose of which is to effectively remove the non-uniform noise existing in the asphalt pavement infrared image while protecting the image texture details.
[0006] The technical solution adopted by the present invention is as follows:
[0007] 1. A method for correcting non-uniformity of asphalt infrared images based on dictionary learning, characterized by comprising:
[0008] S1, collecting infrared images of asphalt pavement during paving;
[0009] S2, dictionary initialization:
[0010] S21, graying the collected infrared image of the asphalt pavement;
[0011] S22, dividing the grayscale image into a number of image blocks, performing a two-dimensional DCT transformation and an inverse transformation on each image block, and then merging the image blocks to generate a DCT dictionary;
[0012] S23, performing singular value decomposition on the grayscale image, and using the decomposed singular value matrix as an SVD dictionary, wherein each column vector of the decomposed singular value matrix is used as a dictionary atom of the SVD dictionary;
[0013] S24, perform noise check on the SVD dictionary through Ljung-box to filter out dictionary atoms without noise;
[0014] S25, splicing the DCT dictionary with the selected dictionary atoms without noise to form an initial dictionary for dictionary learning;
[0015] S3, updating the initial dictionary through the OPM algorithm to obtain an over-complete dictionary of the asphalt pavement infrared image;
[0016] S4. Perform sparse representation on the asphalt pavement infrared image according to the overcomplete dictionary to achieve non-uniformity correction.
[0017] Further technical solutions are:
[0018] In step S24, the noise check is performed on the SVD dictionary by Ljung-box to screen out dictionary atoms without noise, including:
[0019] Two hypotheses are proposed:
[0020] Assume H 0 : Each dictionary atom is distributed independently;
[0021] Assume H 1 : The dictionary atoms are not independently distributed;
[0022] Calculate the hypothesis H 0 The test statistic Q of is as shown below, and the statistic follows the chi-square distribution:
[0023]
[0024] Where, d k represents the kth dictionary atom in the SVD dictionary, L is the length of each dictionary atom, m is the maximum number of lag steps involved in calculating the statistic, is the autocorrelation coefficient of the atom at the lag j step;
[0025] Calculate the average chi-square distribution probability value p of the dictionary atoms from 0 to [L / 2] steps of lag k :
[0026]
[0027] In the formula, χ 2 is a chi-square distribution;
[0028] When the corresponding chi-square distribution probability p of the test statistic Q k When it is greater than the set value, it is considered that H 0 The assumption is true, that is, the atom is composed of random noise, and it is an invalid atom; otherwise, the assumption H 1 The atom passes the white noise test and is a valid atom that can be used to sparsely represent the infrared image.
[0029] In step S4, the sparse representation of the asphalt pavement infrared image is performed according to the overcomplete dictionary to achieve non-uniformity correction, including:
[0030] S41. Set the objective function of sparse representation:
[0031]
[0032] Where f is the target infrared image to be corrected for non-uniformity, D 0 is an overcomplete dictionary, A is the coefficient matrix of the sparse representation of the overcomplete dictionary, μ and σ are the coefficients and sparsity of the sparse items respectively;
[0033] S42, use the orthogonal matching pursuit algorithm to track and update the coefficient matrix A: the coefficient matrix A is compared with the overcomplete dictionary D 0 Multiply to obtain the corrected image generated in the next iteration step And calculate the corrected image Non-uniformity index:
[0034]
[0035] In the formula, is the corrected image is the non-uniformity index of , β is the constant used for adjustment, is the corrected image The output value after convolution with the Laplace operator is: Is to generate the corrected image Output value through Gaussian-Laplace transform;
[0036] S43, calculating the difference of the non-uniformity index in adjacent iterations, if the difference does not meet the iteration termination condition, updating the sparsity in the objective function, repeating S42 to S43 until the iteration termination condition is met, and outputting the asphalt infrared image after non-uniformity correction.
[0037] In step S43, the sparsity in the objective function is updated using the following formula:
[0038]
[0039] In the formula, σ k is the sparsity at the kth step, σ min is the initial sparsity, NEI is the non-uniformity index value of the current iteration, NEI max is the initial non-uniformity index value, and ε is the minimum value.
[0040] In step S1, an infrared thermal imager is used to collect an infrared image of the asphalt pavement while keeping the collection height and the collection angle perpendicular to the ground unchanged, and the image collection is completed before rolling during the paving process.
[0041] The beneficial effects of the present invention are as follows:
[0042] The present invention makes full use of the fact that dictionary learning does not require a large amount of training data, and improves the effect of the non-uniformity correction method based on dictionary learning through Ljung-box noise test and adaptive tolerance, while enhancing the ability to retain image details. It can meet the needs of subsequent detection tasks based on asphalt infrared images, improve the actual engineering application value of infrared images in road surface detection, and has the following significant improvements:
[0043] (1) By using dictionary learning for non-uniformity correction of asphalt infrared images, the demand for large amounts of data is reduced and the implementation difficulty of the method is greatly reduced.
[0044] (2) The initial dictionary is screened through the Ljung-box noise test to improve the effect of subsequent non-uniformity correction.
[0045] (3) The adaptive tolerance based on the image non-uniformity index further improves the accuracy of image non-uniformity correction and the preservation of image details.
[0046] (4) It can improve the accuracy of subsequent tasks based on asphalt infrared images.
[0047] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 This is a flowchart of a method for correcting asphalt infrared image non-uniformity according to an embodiment of the present invention.
[0049] Figure 2 It is the variation curve of the non-uniformity correction index NEI according to the present invention with the non-uniformity intensity.
[0050] Figure 3 It is a schematic diagram of the result of asphalt infrared image after non-uniformity correction in an embodiment of the present invention.
[0051] Figure 4 It shows the change trend of tolerance and non-uniformity correction index NEI during the iterative process of non-uniformity correction in the embodiment of the present invention.
[0052] Figure 5 This is a schematic diagram comparing the non-uniformity correction results of the method of the present invention and other methods. DETAILED DESCRIPTION
[0053] The specific implementation of the present invention is described below with reference to the accompanying drawings.
[0054] This embodiment is a method for correcting asphalt infrared image non-uniformity based on dictionary learning, see Figure 1 , including the following steps:
[0055] S1, collecting infrared images of asphalt pavement during paving;
[0056] S2, dictionary initialization:
[0057] S21, graying the collected infrared image of the asphalt pavement;
[0058] S22, dividing the grayscale image into a number of image blocks, performing a two-dimensional DCT transform (discrete cosine transform) and an inverse transform on each image block, and then merging the image blocks to generate a DCT dictionary;
[0059] S23, performing singular value decomposition on the grayscale image, and using the decomposed singular value matrix as an SVD dictionary, wherein each column vector of the decomposed singular value matrix is used as a dictionary atom of the SVD dictionary;
[0060] S24, perform noise check on the SVD dictionary through Ljung-box to filter out dictionary atoms without noise;
[0061] S25, splicing the DCT dictionary with the selected dictionary atoms without noise to form an initial dictionary for dictionary learning;
[0062] S3, updating the initial dictionary by using the orthogonal matching pursuit (OPM) algorithm to obtain an over-complete dictionary of the asphalt pavement infrared image;
[0063] S4. Perform sparse representation on the asphalt pavement infrared image according to the overcomplete dictionary to achieve non-uniformity correction.
[0064] This embodiment makes full use of the fact that dictionary learning does not require a large amount of training data, and improves the effect of the non-uniformity correction method based on dictionary learning through Ljung-box noise test and adaptive tolerance, while enhancing the ability to retain image details. It can meet the needs of subsequent detection tasks based on asphalt infrared images and improve the actual engineering application value of infrared images in road surface detection.
[0065] Preferably, step S1 specifically includes:
[0066] An infrared thermal imager (Fluke Tix660) was used to collect infrared images of the asphalt pavement while maintaining a collection height of 40 cm from the ground and a collection angle perpendicular to the ground. The image collection was completed before rolling during the paving process.
[0067] Preferably, step S2 specifically includes:
[0068] S21, grayscale processing is performed on the collected infrared image, and the calculation formula is as follows:
[0069]
[0070] Among them, x and y are the horizontal and vertical coordinates of the pixel, g(x,y) represents the gray value at the position (x,y), and f R (x,y),f G (x,y),f B (x, y) represents the position of red, green, and blue pixels in the image respectively;
[0071] S22, dividing the grayscaled asphalt infrared image into 8*8 image blocks, performing a two-dimensional DCT (discrete cosine transform) transformation and an inverse transformation on each image block, and finally merging the image blocks to generate a DCT dictionary;
[0072] The two-dimensional DCT transform formula is as follows:
[0073]
[0074]
[0075] Where F(u,v) is the coefficient after DCT transformation, c(u) and c(v) are the corresponding compensation coefficients, f(i,j) is the two-dimensional image block, N is the size of the image block, i and j represent the horizontal and vertical coordinates of the image, and u and v represent the coordinates after DCT transformation.
[0076] The inverse transformation is similar to this, and the original image information can be restored through the inverse transformation. The corresponding transformation formula will not be repeated here.
[0077] S23, performing singular value decomposition (SVD) on the grayscale infrared image, the singular value decomposition formula is as follows:
[0078] X=U∑V T
[0079] Where X is the grayscale infrared image, ∑ is a diagonal matrix, U and V are left and right singular value matrices respectively, and each column vector of the U matrix is used as a dictionary atom of the SVD dictionary;
[0080] Step S24, performing a noise check on the SVD dictionary through a Ljung-box to screen out dictionary atoms that do not contain noise, including:
[0081] Two hypotheses are proposed:
[0082] Assume H 0 : Each dictionary atom is distributed independently, that is, the atoms will not affect each other;
[0083] Assume H 1 : The atoms in each dictionary are not distributed independently, that is, there is mutual influence between atoms;
[0084] Calculate the hypothesis H 0 The test statistic Q of is as shown below, and the statistic follows the chi-square distribution:
[0085]
[0086] Where, d k represents the kth dictionary atom in the SVD dictionary, L is the length of each dictionary atom, m is the maximum number of lag steps involved in calculating the statistic, is the autocorrelation coefficient of the atom at the lag j step;
[0087] Calculate the average chi-square distribution probability value p of the atom from 0 to [L / 2] steps of lag k :
[0088]
[0089] In the formula, χ 2 is a chi-square distribution;
[0090] When the corresponding chi-square distribution probability p of the test statistic Q k When it is greater than the set value, it is considered that H 0 The assumption is true, that is, the atom is composed of random noise, and it is an invalid atom; otherwise, the assumption H 1 The atom passes the white noise test and is a valid atom that can be used to sparsely represent the infrared image. The setting value in this embodiment is 0.05.
[0091] S25, concatenating the DCT dictionary with the noise-free dictionary atoms (ie, the effective atoms that can perform sparse representation on the infrared image) to form an initial dictionary for dictionary learning.
[0092] Preferably, step S3 specifically includes:
[0093] Define the optimization goal:
[0094]
[0095] Where Y is the grayscale infrared image, D is the initial dictionary to be learned, X is the coefficient of the initial dictionary, and α is a constant representing the sparsity constraint parameter.
[0096] Decomposing the above formula, we can get:
[0097]
[0098] Where, d j ,d k Represent the jth and kth column dictionary elements respectively, X j ,X k Respectively represent the j-th column and the k-th column of the dictionary coefficient matrix;
[0099] The terms in parentheses are expressed as residual terms E k :
[0100]
[0101] Then the objective function expression of the above optimization objective can be transformed into:
[0102]
[0103] The singular value decomposition method is used to solve the above objective function:
[0104] E k =U′∑′V′ T
[0105] In the above formula, U′, V′ are the left and right singular value matrices after decomposition, and ∑′ is a diagonal matrix.
[0106] Take the first column vector of the matrix U′ as the update atom of the dictionary element, and take the product of the first row vector of the matrix V′ and the first singular value to update the dictionary atom of the corresponding sparse matrix, that is, the initial dictionary X;
[0107] The dictionary elements are continuously updated through the OMP algorithm. When the dictionary satisfies the sparsity constraint, the dictionary update ends and an overcomplete dictionary of the infrared image is output.
[0108] Preferably, step S4 specifically includes:
[0109] S41. Set the objective function of sparse representation:
[0110]
[0111] Where f is the target infrared image to be corrected for non-uniformity, D 0 is an overcomplete dictionary, A is the coefficient matrix of the sparse representation of the overcomplete dictionary, μ and σ are the coefficients and sparsity of the sparse items respectively;
[0112] S42, using the orthogonal matching pursuit algorithm (OMP) to track and update the coefficient matrix A:
[0113] Combine the coefficient matrix A with the overcomplete dictionary D 0 Multiply to obtain the corrected image generated in the next iteration step And calculate the corrected image Non-uniformity index:
[0114]
[0115] In the formula, is the corrected image is the non-uniformity index of , β is the constant used for adjustment, is the corrected image The output value after convolution with the Laplace operator is: Is to generate the corrected image Output value through Gaussian-Laplace transform;
[0116] S43, calculating the difference of the non-uniformity index in adjacent iterations, if the difference does not meet the iteration termination condition, updating the sparsity in the objective function, repeating S42 to S43 until the iteration termination condition is met, and outputting the asphalt infrared image after non-uniformity correction.
[0117] like Figure 2As shown in FIG. 1 , the non-uniformity correction index NEI changes with the non-uniformity intensity. As can be seen from the figure, the non-uniformity correction index NEI increases as the non-uniformity intensity of the image increases.
[0118] Specifically, the difference between the five consecutive iterations is compared to see whether the iteration termination condition is reached. The termination condition is expressed as follows:
[0119]
[0120] In the formula, χ is the iteration termination judgment condition, NEI i is the NEI index value of the i-th iteration, which is set to 0.001 in this example.
[0121] Specifically, the sparsity in the objective function is updated using the following formula:
[0122]
[0123] In the formula, σ k is the sparsity at the kth step, σ min is the initial sparsity, which is set to 90 in this embodiment. NEI is the non-uniformity index value of the current iteration. max is the initial non-uniformity index value, and ε is the minimum value, in order to ensure that the formula is meaningful.
[0124] like Figure 3 The figure is a schematic diagram of the result of non-uniformity correction of asphalt infrared image in this embodiment. Figure 3 (a) and (b) are infrared images of asphalt pavement before and after non-uniform correction, respectively. It can be seen from the figures that the asphalt infrared image non-uniformity correction method proposed in this application can effectively remove the non-uniform noise in the image while retaining the image texture details.
[0125] like Figure 4 As shown in the figure, the non-uniformity index NEI and adaptive tolerance change curves with iteration in the sparse representation iteration process. It can be seen from the figure that as the number of iterations increases, the non-uniform noise in the asphalt infrared image decreases, and the tolerance also decreases adaptively with the reduction of non-uniform noise, thereby improving the accuracy of non-uniformity correction.
[0126] In order to prove the superiority of the method proposed in the present invention, the currently commonly used non-uniformity correction method is compared with the denoising method based on dictionary learning proposed in the present invention. The NEI index after non-uniformity correction is shown in Table 1 below, and the corresponding correction result diagram is shown in Figure 5 As shown in the figure, it can be seen that in the asphalt infrared image non-uniformity correction, the method of the present invention can better remove the non-uniform noise and better preserve the texture details in the image.
[0127] Table 1 NEI index of different non-uniformity correction methods
[0128] method KSV BM3D LLRT LRTA MxDV The present invention <![CDATA[NEI(×10 -2 )]]> 4.511 4.937 5.001 4.113 4.126 <![CDATA[ 3.960 ]]>
[0129] The present invention makes full use of the characteristic that dictionary learning has little reliance on data, improves the dictionary learning denoising method for asphalt infrared images, and realizes efficient and accurate non-uniform correction of asphalt infrared images.
[0130] Those skilled in the art can understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for correcting asphalt infrared image non-uniformity based on dictionary learning, characterized in that: include: S1, collecting infrared images of asphalt pavement during paving; S2, dictionary initialization: S21, graying the collected infrared image of the asphalt pavement; S22, dividing the grayscale image into a number of image blocks, performing a two-dimensional DCT transformation and an inverse transformation on each image block, and then merging the image blocks to generate a DCT dictionary; S23, performing singular value decomposition on the grayscale image, and using the decomposed singular value matrix as an SVD dictionary, wherein each column vector of the decomposed singular value matrix is used as a dictionary atom of the SVD dictionary; S24, perform noise check on the SVD dictionary through Ljung-box to filter out dictionary atoms without noise; S25, splicing the DCT dictionary with the selected dictionary atoms without noise to form an initial dictionary for dictionary learning; S3, updating the initial dictionary through the OPM algorithm to obtain an over-complete dictionary of the asphalt pavement infrared image; S4, performing sparse representation on the asphalt pavement infrared image according to the overcomplete dictionary to achieve non-uniformity correction; In step S24, the noise check is performed on the SVD dictionary by Ljung-box to screen out dictionary atoms without noise, including: Two hypotheses are proposed: Assumption H0: Each dictionary atom is independently distributed; Assumption H1: The dictionary atoms are not independently distributed; Calculate the test statistic Q for hypothesis H0, as shown below. The statistic follows a chi-square distribution: Where, d k represents the kth dictionary atom in the SVD dictionary, L is the length of each dictionary atom, m is the maximum number of lag steps involved in calculating the statistic, is the autocorrelation coefficient of the atom at the lag j step; Calculate the average chi-square distribution probability value p of the dictionary atoms from 0 to [L / 2] steps of lag k : In the formula, χ 2 is a chi-square distribution; When the corresponding chi-square distribution probability p of the test statistic Q k When it is greater than the set value, the H0 hypothesis is considered to be valid, that is, the atom is composed of random noise and is an invalid atom; otherwise, the H1 hypothesis is considered to be valid, the atom passes the white noise test, and is a valid atom that can be used to sparsely represent the infrared image.
2. The asphalt infrared image non-uniformity correction method based on dictionary learning according to claim 1 is characterized in that: In step S4, the sparse representation of the asphalt pavement infrared image is performed according to the overcomplete dictionary to achieve non-uniformity correction, including: S41. Set the objective function of sparse representation: Where f is the target infrared image to be corrected for non-uniformity, D0 is the overcomplete dictionary, A is the coefficient matrix of the sparse representation of the overcomplete dictionary, μ and σ are the coefficient and sparsity of the sparse terms respectively; S42, use the orthogonal matching pursuit algorithm to track and update the coefficient matrix A: multiply the coefficient matrix A by the overcomplete dictionary D0 to obtain the corrected image generated in the next iteration step And calculate the corrected image Non-uniformity index: In the formula, is the corrected image is the non-uniformity index of , β is the constant used for adjustment, is the corrected image The output value after convolution with the Laplace operator is, is the corrected image generated under the current sparsity σ Output value through Gaussian-Laplace transform; S43, calculating the difference of the non-uniformity index in adjacent iterations, if the difference does not meet the iteration termination condition, updating the sparsity in the objective function, repeating S42 to S43 until the iteration termination condition is met, and outputting the asphalt infrared image after non-uniformity correction.
3. The asphalt infrared image non-uniformity correction method based on dictionary learning according to claim 2 is characterized in that: In step S43, the sparsity in the objective function is updated using the following formula: In the formula, σ k is the sparsity at the kth step, σ min is the initial sparsity, NEI is the non-uniformity index value of the current iteration, NEI max is the initial non-uniformity index value, and ε is the minimum value.
4. The asphalt infrared image non-uniformity correction method based on dictionary learning according to claim 1 is characterized in that: In step S1, an infrared thermal imager is used to collect an infrared image of the asphalt pavement while keeping the collection height and the collection angle perpendicular to the ground unchanged, and the image collection is completed before rolling during the paving process.