Quantitative evaluation method of pushbroom hyperspectral imaging stripe noise under controlled light environment

Abstract

A kind of push-broom hyperspectral imaging stripe noise quantitative evaluation method under controlled light environment, steps include:1) stripe noise generation mechanism research;2) smooth object hyperspectral image stripe noise analysis;2) complex object hyperspectral image stripe noise analysis;4) stripe noise evaluation method verification.The present application can satisfy the stripe noise evaluation of original hyperspectral image and hyperspectral image after being processed by different denoising methods, and provide technical support for index space visualization based on chemometrics.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for quantitatively evaluating the noise of push-broom hyperspectral images, which is a method for evaluating the stripe noise of push-broom hyperspectral imaging systems.

[0002] The method can evaluate the stripe noise of the system, can be used to improve the performance of the fast comparison despeckling algorithm, and provides a strong foundation for improving the accuracy based on the visualization of the chemometric index space, and belongs to the field of hyperspectral image processing. BACKGROUND

[0003] With the continuous development of detection technology, traditional detection technology gradually cannot meet the needs of industrial detection due to its defects such as time-consuming and labor-intensive. It is urgent to find new efficient non-destructive detection methods.

[0004] Hyperspectral imaging technology is a new optical non-destructive detection technology that combines image and spectral information. As a non-contact non-destructive detection technology, it has the ability to detect external defects and internal components, and has been widely used in non-destructive detection fields such as food and medicinal materials.

[0005] At present, the acquisition methods based on hyperspectral imaging equipment are mainly divided into point scanning, push scanning and area array, among which push scanning is most suitable for rapid non-destructive detection. However, push scanning hyperspectral imaging has regular stripe noise in the scanning process, which seriously affects the image quality. Therefore, a large number of methods for removing and evaluating stripe noise have been proposed, but there is no method for quantitatively evaluating the stripe noise of push scanning hyperspectral imaging under controlled light environment.

[0006] The existence of stripe noise seriously affects the quality of index space visualization, so there is an urgent need for a stripe noise evaluation method specifically used in the visualization field.

[0007] In the prior art, the Chinese patent with the publication number CN101599170B "Image noise evaluation method and image noise evaluation device" firstly preliminarily estimates the noise level of the image, then adaptively extracts a uniform flat area according to the preliminary estimation result, and then adaptively iteratively optimizes the uniform flat area, and finally optimizes and combines a plurality of noise evaluation indexes to obtain a comprehensive and robust noise evaluation method. The method has higher comprehensive accuracy and wider adaptability, and especially in complex motion scenes, it can use several frames or even one frame of image to quickly and accurately give the noise level of the current image. However, this method is not suitable for evaluating the regular stripe noise of push scanning hyperspectral imaging in the scanning process. SUMMARY

[0008] In order to solve the problem that the stripe noise in the index space visualization is difficult to evaluate, the present application provides a push-broom hyperspectral imaging stripe noise quantitative evaluation method under a controlled light environment, and the stripe noise is obtained and evaluated by using a difference method and a numerical statistical method.The specific steps of the present application are as follows:

[0009] A push-broom hyperspectral imaging stripe noise quantitative evaluation method under a controlled light environment, and the steps include:

[0010] 1) obtaining the stripe noise of the eigenimage of the push-broom hyperspectral imaging system, and amplifying the stripe noise; the eigenstripe noise is helpful to detect the position of the stripe noise in the original hyperspectral image, and facilitates the acquisition of the stripe noise;

[0011] 2) obtaining the stripe noise of the measured image of the push-broom hyperspectral imaging system;

[0012] 3) analyzing and judging whether the pixel is disturbed by the stripe noise, and quantitatively evaluating the stripe noise;

[0013] In step 1), a set of standard reflectance plates are used to capture the stripe noise of the push-broom hyperspectral imaging system; then the noise is obtained by amplifying the noise, which is used to evaluate the eigenstripe noise of the push-broom hyperspectral imaging system:

[0014] In step 2), the steps of obtaining the stripe noise of the measured image include:

[0015] 2.1) obtaining the stripe characteristics of the hyperspectral image of the complex object by difference calculation, and using the difference image to describe the stripe noise in the hyperspectral image; capturing the pixel points disturbed by the stripe noise, and encoding the pixel points;

[0016] 2.2) obtaining the proportion of the obvious stripe noise in the complex object by using a numerical statistical method and encoding;

[0017] In step 3), the pixel points disturbed by the stripe noise in the image are counted, and the stripe noise in the image is quantitatively described; the quantitative value of the stripe noise is used to evaluate the stripe noise of the push-broom hyperspectral imaging.

[0018] The present application studies the generation mechanism of the stripe noise, analyzes the stripe noise of the hyperspectral image of the smooth object, analyzes the stripe noise of the hyperspectral image of the complex object, and verifies the evaluation method.

[0019] The present application can evaluate the stripe noise of the original hyperspectral image and the hyperspectral image processed by different denoising methods, and provide technical support for the index space visualization based on chemometrics. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 Spectral image of 99% standard reflectance plate

[0021] Figure 2(a) is a standard reflectance plate spectral image

[0022] Figure 2(b) is a ginkgo biloba leaf spectral image

[0023] Figure 2(c) is a standard reflectance plate difference map

[0024] Figure 2(d) is a ginkgo biloba leaf difference map

[0025] Figure 2(e) is a standard reflectance plate difference column mean

[0026] Figure 2(f) is a ginkgo biloba leaf difference column mean

[0027] Figure 2(g) is a standard reflectance plate fringe noise encoding map

[0028] Figure 2(h) is a ginkgo biloba leaf fringe noise encoding map

[0029] Figure 3 Comparison of fringe index for original and de-fringed spectral images

[0030] Figures 4(a) and 4(b) are icv and urv values of the visualization results generated by LRTD and pixel-wise fringe removal algorithm, respectively

[0031] Figure 5 Visualization results after low-rank tensor decomposition and pixel-wise fringe removal. DETAILED DESCRIPTION

[0032] The present application is a push-broom hyperspectral imaging fringe noise quantification and evaluation method under controlled light environment, the steps of which include:

[0033] 1) Analyzing the fringe noise of the hyperspectral image of a standard reflectance plate;

[0034] 2) Designing a hyperspectral image fringe noise evaluation method for complex objects;

[0035] 3) Designing a fringe noise quantification and evaluation method;

[0036] 4) Verifying the fringe noise quantification and evaluation method;

[0037] In step 1), to clearly show the fringe noise distribution of the hyperspectral image, a set of 99% standard reflectance plates is used to capture the fringe noise of the push-broom hyperspectral imaging system;

[0038] Because the stripe noise itself belongs to column regular distribution in image distribution, that is, the stripe noise keeps stable interference to the same column pixels in the line scanning process. The interference degree of the stripe noise to the pixel gray value is not easy to obtain in the original image, so it is necessary to obtain the noise by amplifying the noise. The difference between each pixel point and the left and right pixel points is calculated, and the difference image is generated, the stripe noise can be amplified, and the calculation formula is as follows:

[0039]

[0040] Wherein, D ij is the pixel difference value of the i-th row and the j-th column, p ij is the pixel gray value of the i-th row and the j-th column.

[0041] Because the 99% standard reflectivity plate is a uniform sample, therefore, there should be no uneven area in the ideal spectral image, that is, the difference value of each pixel point is close to 0. By this method, the stripe noise can be greatly amplified.

[0042] In step 2), because the actual measured object is quite different from the standard reflectivity plate, the information in the image is more complex, and the stripe noise cannot be obtained by the difference calculation method. The stripe noise of the image is obtained by the following method:

[0043] (1) Calculate the difference image, and the calculation formula is as follows:

[0044]

[0045] Compared with the standard reflectivity plate image, the same column pixels in the actual measured object are interfered by the stripe noise, because the same column pixel gray values are not at the same gray level, so that the difference value will appear the phenomenon of positive and negative alternation, in order to solve the interference of the phenomenon to the acquisition of the stripe noise, the absolute difference value is selected in the difference calculation formula.

[0046] (2) Compare the difference values of adjacent pixels, and code the pixel points:

[0047]

[0048] Wherein, E ij is the stripe coding value of the i-th row and the j-th column pixel point.

[0049] (3) Because of the complexity of the image, the coding result of (2) cannot obtain the same result in the same column, therefore, the coding result is analyzed by numerical analysis method, and the result is optimized. The result shows that when the difference value of 80% of the pixel points in the same column is greater than that of the left and right sides, it can be determined that the pixel point is interfered by the stripe noise.

[0050]

[0051] wherein E i is the i-th row stripe encoding value, n is the j-th column pixel number.

[0052] In step 3), the degree of interference of the hyperspectral image by the stripe noise is quantitatively represented according to the encoding result.

[0053]

[0054] wherein S r represents the degree of interference of the image by the stripe noise, N s is the number of pixel points interfered by the stripe noise, and N is all pixel points in the image.

[0055] In step 4), since the spatial index visualization method has the effect of amplifying the stripe noise, the hyperspectral images after different denoising treatments are used for visualization, and the reliability of the stripe noise evaluation method is verified in reverse according to the result. The hyperspectral image stripe noise evaluation method is verified in reverse by taking the spatial index visualization based on chemometrics as the verification means.

[0056] Step 1) describes the hyperspectral imaging system intrinsic stripe noise evaluation method.

[0057] In formula 2, the stripe characteristics of the hyperspectral image of the complex object obtained by the difference calculation are described by the difference image as the stripe noise in the hyperspectral image.

[0058] In formula 3, according to the stripe noise generation mechanism, the pixel points interfered by the stripe noise are captured.

[0059] In formula 4, the numerical statistics method is used to analyze the proportion of the obvious stripe noise in the complex object, and the image is encoded.

[0060] In formula 5, the pixel points interfered by the stripe noise in the image are counted, and the stripe noise in the image is quantitatively described.

[0061] The present method will be described below in combination with the drawings and specific embodiments.

[0062] The distribution state of the stripe noise is relatively simple, that is, there is a large difference between the pixel and the left and right adjacent pixels. The pixel whether is affected by the stripe noise can be confirmed by judging the change difference between the pixel and the surrounding two pixels. However, this method is only suitable for the images of the surface uniform sample, and the standard reflectivity plate used for the image reflectivity correction of the present application is a relatively suitable sample. By setting the gray scale range to 49000 to 53000, the pixel points affected by the stripe noise can be found from the image, such as Figure 1The gray values of the two sides are lower due to the non-uniform illumination of the light source on both sides. The stripes in the image can be more clearly distinguished, and the appearance of the stripe noise is also always present in a column.

[0063] However, the samples in actual image acquisition do not have a uniform surface, and the stripe noise cannot be directly analyzed. The present application designs the following method to evaluate the stripe noise of the image:

[0064] (I) Calculate the difference between each pixel point and the left and right pixel points, and generate a difference image;

[0065] (II) Judge the pixel points on the difference image. If the left and right pixel points of the pixel point are greater than or less than the difference value of the pixel point, then the pixel point is coded as 1; if the above conditions are not met, then the pixel point is coded as 0; generate an image code.

[0066] (III) Calculate the average value of the coding value of each column and divide it by the total number of pixel points. If it is greater than 0.8, it is determined that the column pixel is disturbed by stripe noise; if it is less than 0.8, it is determined that the column pixel is not disturbed by stripe noise.

[0067] In this example, the method is used to analyze the standard reflectance plate and the leaf image of ginkgo biloba. The results are shown in FIG. 2(a) and FIG. 2(b). FIG. 2(a) and FIG. 2(b) are the original spectral images of the standard reflectance plate and the leaf, respectively. The image of the standard reflectance plate is scaled, while the leaf does not scale due to the large difference in pixels within the image. Therefore, the spectral image of the leaf cannot directly visually see the stripe noise. After difference calculation, the stripe noise of the spectral image is significantly amplified. From FIG. 2(c), it can be found that due to the existence of stripe noise, the difference image of the standard reflectance plate presents a large data fluctuation. From FIG. 2(d), it can be found that the fluctuation in the difference image of the ginkgo biloba leaf is smaller, and the distribution of the stripes can be seen but not enough to judge the stripes. The difference column average of FIG. 2(c) and FIG. 2(d) can obtain the results of FIG. 2(e) and FIG. 2(f), and the difference column average more clearly shows the stripe noise on the hyperspectral image. The above results are obtained according to the results of the first step of the evaluation method. This step amplifies the stripe noise in the image, which is more convenient for observation.

[0068] Since the first step only plays a huge amplification effect on the fringe noise in the standard reflectance plate, it does not show good results for the ginkgo leaf spectral image. In order to have a good effect on the fringe noise in the complex image, the second step and the third step of the method are used to calculate and count the difference image. By comparing the size of the pixel points in the difference image with the left and right pixel points, it can be judged whether the pixel points in the original spectral image are different from the left and right pixel points, that is, the suspected fringe pixel points. In the third step, the number of suspected pixel points in each column is counted. Since the image information in the ginkgo leaf is relatively complex, the pixel points in the column direction are uniformly analyzed during analysis. The fringe noise always interferes with the entire column of pixels when interfering with the image, so whether a pixel point is interfered by the fringe noise can be analyzed from the entire column of pixels. However, the ginkgo leaf image contains texture information that the standard reflectance does not have, so there are pixel points in the same column that do not meet the conditions of the second step. Through analysis, it is found that the number of such pixel points in the fringe column and not meeting the group conditions is less than 20%, so the threshold is set to 80% in the third step. The results obtained by this method are basically consistent with the column of the fringe noise. Figures 2(g) and 2(h) are the fringe encoding diagrams of the standard reflectance plate and the ginkgo leaf spectral image, respectively, in which black represents that the pixel point is not interfered by the fringe noise, and white represents that the pixel point is interfered by the fringe noise. This method can quantify the interference of the fringe noise on the spectral image, and the interference degree of the standard reflectance plate spectral image is 45.8%, and the interference degree of the ginkgo leaf spectral image is 3.5%. The calculation formula of the numerical value is as follows:

[0069]

[0070] Wherein, S r represents the interference degree of the image by the fringe noise, N s is the number of pixel points interfered by the fringe noise, and N is all the pixel points in the image. And the generated encoding image is consistent with the detected fringe in the original spectral image and the difference image. It is proved that the method is a good fringe noise detection method.

[0071] In the present application, two methods of LRTD and pixel-by-pixel fringe removal are used to correct the fringe noise of the hyperspectral image. In order to detect the effect of the two methods on the fringe noise removal, the above-mentioned fringe noise detection method is applied to the generated hyperspectral image.

[0072] The fringe index of the two fringe removal algorithms is obtained by evaluating the fringe noise of all samples at different wavelengths, as shown in the following table: Figure 3The strip index of LRTD is lower than that of the original spectral image, especially in the first half of the spectral image. The strip index of the pixel-wise strip removal method is greatly improved, and the strip index of the spectral image of most channels is below 1%. To verify the effectiveness of the method, the hyperspectral images generated by LRTD and the pixel-wise strip removal method are applied to the visualization of the total flavonoids content of Ginkgo biloba leaves, and a comparison is made.

[0073] The icv and urv values of the visualization results generated by LRTD and the pixel-wise strip removal algorithm are shown in FIG. 4(a) and FIG. 4(b). It can be found that when the principal component number increases from 8 to 9, the unreasonable value of the pixel-wise strip removal algorithm increases instantaneously. The difference between the two strip removal algorithms is large before and after the principal component number 8. When the principal component number is 1-8, the unreasonable prediction value of the pixel-wise strip removal algorithm is obviously lower than that of LRTD, and when the principal component number is higher than 8, LRTD has better performance. There is no big difference in image quality between the two strip removal algorithms.

[0074] Figure 5 For the principal component number 8, the visualization prediction images and their evaluation indexes after applying the two strip removal algorithms are shown. It is found that after strip noise removal, compared with the original visualization prediction result, the prediction accuracy has a slight decrease, but the visualization quality has a large increase; compared with the visualization prediction result after optimization but without strip noise removal, the visualization quality is obviously improved due to the weakening of the strip noise, but the prediction accuracy after strip noise removal decreases.

[0075] After strip noise removal, the range of unreasonable values is smaller, so the pseudo-color range of Figure 5 is adjusted to 0-9 mg / g. It can be found that the visualization predictions of the images generated by the two algorithms are basically consistent. Comparing the visualization prediction results of the two methods, it can be found that there is still some strip noise in the image of LRTD, while the strip noise is greatly eliminated after the pixel-wise strip removal processing. This conclusion is consistent with the result of the strip noise evaluation method proposed in the present application, which proves the effectiveness of the strip noise evaluation method proposed in the present application. In terms of visualization evaluation indexes, the pixel-wise strip removal method is better than LRTD, and the RMSE roi of the pixel-wise strip removal method is 0.6174 mg / g, the icv is 2.6052, and the urv is 0.65%.

Claims

1. A method for quantitatively evaluating the striping noise of pushbroom hyperspectral imaging under controlled light environment, comprising the following steps: 1) obtaining the striping noise of the intrinsic image of the pushbroom hyperspectral imaging system and amplifying the striping noise; 2) obtaining the striping noise of the measured image of the pushbroom hyperspectral imaging system; 3) analyzing and judging whether the pixels are disturbed by the striping noise and quantitatively evaluating the striping noise; In step 1), a set of standard reflectance plates are used to capture the striping noise of the pushbroom hyperspectral imaging system; then the noise is obtained by amplifying the noise, which is used to evaluate the intrinsic striping noise of the pushbroom hyperspectral imaging system: In step 2), the steps for obtaining the striping noise of the measured image include: 2.1) obtaining the striping characteristics of the hyperspectral image of the complex object by difference calculation, and using the difference image to describe the striping noise in the hyperspectral image; capturing the pixel points disturbed by the striping noise and encoding the pixel points; 2.2) obtaining the proportion of obvious striping noise in the complex object by numerical statistics and encoding; In step 3), the pixel points disturbed by the striping noise in the image are counted, and the striping noise in the image is quantitatively described; the striping noise of the pushbroom hyperspectral imaging is evaluated by the quantitative value of the striping noise; In step 2.1), for the measured image, the difference image is calculated, and the calculation formula is as follows: wherein, is the first row and the first column pixel difference value, is the first row and the first column pixel gray value; The difference values of adjacent pixels are compared, and the pixel points are encoded: wherein is the is the is the 2.2) when 80% of the pixel points in the same column have difference values greater than those on the left and right sides, it is determined that the pixel points are disturbed by the striping noise: wherein is the first row stripe encoding value, is the first column pixel point number; In step 3), according to the encoded results, the degree of disturbance of the hyperspectral image by the striping noise is quantitatively represented: wherein, represents the degree of interference of the image by the stripe noise, is the number of pixel points interfered by the stripe noise, is all pixel points in the image; Then by Numerical evaluation of pushbroom hyperspectral imaging striping noise.

2. The method of claim 1, wherein the method is performed in a controlled light environment. In step 1), the difference between each pixel point and the left and right pixel points is calculated, and the difference image is generated, then the striping noise is amplified, and the calculation formula is as follows: in, For the first Line number The pixel difference value of the column, For the first Line number The pixel grayscale value of the column.

3. The method according to claim 1 or 2, characterized in that It also includes step 4) verification of the striping noise quantitative evaluation method: using spatial index visualization based on chemometrics as a verification means to verify the hyperspectral image striping noise evaluation method in reverse.

4. The method of claim 1, wherein the method is performed in a controlled light environment. In step 1), a set of 99% standard reflectance plates are used to capture the striping noise of the pushbroom hyperspectral imaging system.

Images