Brain tumor tissue hyperspectral data non-uniform noise calibration method

Abstract

The invention relates to a noise calibration method, in particular to a brain tumor tissue hyperspectral data non-uniform noise calibration method, and solves the problem that the elimination effect is not ideal when brain tumor tissue non-uniform stripe noise elimination is carried out through light source-free uniform output response and a calibration method based on mathematical statistical analysis. According to the method, a sample image and a corresponding dark background image are collected in the same environment, then weighted least square filtering processing is carried out on the dark background image, the filtered dark background image can keep edge information, then the original dark background image and the filtered image are subtracted to obtain the edge information, and the edge information is obtained. Averaging the full wave band of the edge information to obtain a weighting coefficient of a corresponding pixel position, and then separating out the fixed noise content of the instrument according to the weighting coefficient in a wave band-by-wave manner; weighted least square filtering processing of embedded bilateral filtering is carried out on a sample image, and then the sample image after filtering processing is subtracted from the fixed noise content to obtain a final clear image.

Description

Technical Field

[0001] The invention relates to a noise calibration method, and in particular to a method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue. Background Art

[0002] Hyperspectral images are often used in the imaging display of cancerous tissues because they can display rich spatial and spectral information. In daily use, the hyperspectral imaging system sensor will show non-uniform response output under the influence of multiple factors such as ambient temperature. The existence of non-uniformity reduces the imaging quality and spectral accuracy of hyperspectral images, and seriously reduces the information extraction and target recognition of cancerous sample images, especially the information extraction and target recognition of brain tumor tissue sample images. Therefore, effective non-uniformity correction must be implemented during imaging in the hyperspectral imaging system to improve the accuracy of image quantitative analysis. In the currently commonly used spatial push-scan hyperspectral cancerous tissue imaging system, the image non-uniformity is manifested in the form of streak noise throughout the sampled image. Therefore, removing the streak noise can largely achieve the non-uniformity correction of hyperspectral images.

[0003] There are two main methods for removing non-uniform stripe noise in brain tumor tissue sample images. One is to use the uniform output response without light source as a reference image to eliminate non-uniform stripe noise. The problem is that the stripe noise in the dark background is not estimated based on the position of the dark pixel, so the effect of eliminating non-uniform stripe noise is not ideal. The other is to process the sample image based on mathematical statistical analysis, mainly through filtering, transformation and other methods. The problem is that it is not based on the characteristics of the instrument itself, so the effect of eliminating non-uniform noise is very limited.

[0004] Therefore, in order to better achieve the removal of non-uniform noise, the non-uniform noise removal method of the spatial push-broom hyperspectral imaging system must combine the above two methods at the same time. On the one hand, it focuses on the imaging mechanism of the instrument and analyzes the causes of the instrument non-uniform noise. On the other hand, it analyzes the data rules from the perspective of mathematical statistics, so that the front-end imaging and back-end data processing are organically combined, thereby improving the effect of non-uniform noise elimination. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that both the uniform output response without a light source and the calibration method based on mathematical statistical analysis have unsatisfactory elimination effects when performing non-uniform stripe noise elimination of brain tumor tissue, and to provide a non-uniform noise calibration method for hyperspectral data of brain tumor tissue.

[0006] In order to achieve the above-mentioned invention object, the present invention provides the following technical solutions:

[0007] A method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue is special in that it includes the following steps:

[0008] S1: Collect sample images and corresponding dark background images in the same environment;

[0009] S2: Perform weighted least squares filtering on the dark background image;

[0010] S3: Subtract the original dark background image from the dark background image processed by weighted least squares filtering to obtain edge information of the dark background image, and take the average of the edge information in the entire band to obtain the weighting coefficient of the corresponding pixel position;

[0011] S4: Separate the fixed noise content of the instrument band by band according to the weighting coefficient of the corresponding pixel position;

[0012] S5: Perform weighted least squares filtering with embedded bilateral filtering on the sample image;

[0013] S6: Subtract the filtered sample image from the separated fixed noise content of the instrument to obtain a final clear image, thereby completing the non-uniform noise calibration of the brain tumor tissue hyperspectral data.

[0014] Furthermore, step S1 is specifically as follows: collecting sample images x in the same environment (i,j,B) And the corresponding dark background image x (i,j,B) , where (i,j) is the spatial dimension position and B is the spectral dimension position.

[0015] Furthermore, step S2 is specifically: performing weighted least squares filtering on the dark background image to obtain a filtered dark background image y (i,j,B) ;

[0016]

[0017]

[0018] in: is the gradient of the dark background image along the X direction after filtering, is the gradient of the dark background image along the Y direction after filtering, * is the pixel point, ω i is the weight of the dark background image in the X direction, ω j is the weight of the dark background image in the Y direction, and is the universal weight, ε is the small coefficient, λ is the regularization term coefficient, and y is the dark background image after filtering.

[0019] Furthermore, step S3 is specifically as follows: the original dark background image x (i,j,B)and the dark background image y after weighted least squares filtering (i,j,B) Subtract to get the edge information e′ of the dark background image (i,j,B) , and for the edge information e′ (i,j,B) Take the average of the entire band to obtain the weighting coefficient K of the corresponding pixel position (i,j,B) ;

[0020] e′ (i,j,B) =x (i,j,B) -y (i,j,B)

[0021]

[0022] Where: M×N is the size of the dark background image.

[0023] Further, step S4 is specifically: according to the obtained weighting coefficient K of the corresponding pixel position (i,j,B) Separate the fixed noise content N of the instrument band by band (i,j,B) ;

[0024] N (i,k,B) =K (i,j,B) × (i,j,B) / (dark_m) B

[0025] Where: (dark_m) B is the mean value of a single band of the dark background image.

[0026] Furthermore, step S5 is specifically as follows: performing weighted least squares filtering with embedded bilateral filtering on the sample image to obtain a sample image γ after filtering. (i,j,B) ;

[0027]

[0028] Where: (i,j) is the sample image, γ (i,j) is the sample image after filtering. is the sample image gradient, is the sample image gradient after filtering, ω x,(i,j) is the weight of the sample image along the X direction after filtering, ω y,(i,j) is the weight of the sample image along the Y direction after filtering, f BLF Represents the bilateral operator weight.

[0029] Furthermore, step S6 is specifically as follows: the sample image γ after filtering is (i,j,B) The fixed noise content N of the isolated instrument (i,j,B) Subtract and get the final clear image X′ (i,j,B) ;

[0030] X′ (i,j,B) =γ (i,j,B) -N (i,j,B)

[0031] Completed the inhomogeneous noise calibration of brain tumor tissue hyperspectral data.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) The present invention provides a method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue. First, a sample image and a dark background image of the brain tumor tissue are collected under the same environment. This ensures that the influencing factors of the two images are basically consistent, which is more accurate in later calculations. After the sample image and the dark background image are collected, the two images are filtered respectively. A weighted least squares filtering method is used for the dark background image. This filtering method has the characteristic of maintaining edge information. The edge information of the original dark background image can be retained to the maximum extent by using this filtering method. Then, the edge information can be obtained by subtracting the filtered dark background image from the original dark background image. At this time, the edge information obtained is averaged across all bands to obtain the weighting coefficient of the corresponding pixel position. Then, the fixed noise content of the instrument is separated band by band based on the obtained weighting coefficient. At this time, the processing and calculation of the dark background image is completed.

[0034] (2) The present invention provides a method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue. For a sample image, a weighted least squares filtering process with embedded bilateral filtering is used. After obtaining the sample image after filtering, the fixed noise content of the instrument separated from the dark background image is subtracted to obtain a clear image.

[0035] (3) The present invention provides a non-uniform noise calibration method for brain tumor tissue hyperspectral data. Based on the imaging characteristics of the spatial push-scan hyperspectral brain tumor tissue imaging system, it can more accurately separate the fixed noise content of the instrument itself from the sample image, thereby obtaining a clean and clear hyperspectral brain tumor image. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a working principle diagram of an embodiment of a method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to the present invention;

[0037] Figure 2 The figure is a comparison diagram of the grayscale distribution of brain tumor tissue before and after denoising using an embodiment of a method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and exemplary embodiments.

[0039] The working principle diagram of the non-uniform noise calibration method of brain tumor tissue hyperspectral data of the present invention is as follows: Figure 1 As shown, firstly, it is necessary to collect the sample image and the dark background image of the brain tumor tissue under the same environment, so as to ensure that the influencing factors are basically consistent. After collecting the sample image and the dark background image, the two images are filtered respectively.

[0040] The dark background image is processed by weighted least squares filtering. According to the characteristics of weighted least squares filtering that maintains edge information, this filtering method can retain the edge information of the original dark background image to the maximum extent. Then, the edge information can be obtained by subtracting the filtered dark background image from the original dark background image. At this time, the edge information is averaged across all bands to obtain the weighting coefficient for the corresponding pixel position. The fixed noise content of the instrument is separated band by band based on the obtained weighting coefficient. At this time, the processing calculation of the dark background image is completed.

[0041] For the sample image, weighted least squares filtering with embedded bilateral filtering is used. After obtaining the filtered sample image, it is subtracted from the fixed noise content of the instrument separated from the dark background image to obtain a clear image.

[0042] The specific steps include:

[0043] S1: Collect sample images x in the same environment (i,j,B) And the corresponding dark background image x (i,j,B) , where (i,j) is the spatial dimension position and B is the spectral dimension position.

[0044] S2: Perform weighted least squares filtering on the dark background image to obtain the filtered dark background image y (i,j,B) ;

[0045]

[0046]

[0047] in: is the gradient of the dark background image along the X direction after filtering, is the gradient of the dark background image along the Y direction after filtering, * is the pixel point, ω i is the weight of the dark background image in the X direction, ω j is the weight of the dark background image in the Y direction, and is the universal weight, ε is the small coefficient, λ is the regularization term coefficient, and y is the dark background image after filtering.

[0048] S3: The original dark background image x (i,j,B)and the dark background image y after weighted least squares filtering (i,j,B) Subtract to get the edge information e′ of the dark background image (i,j,B) , and for the edge information e′ (i,j,B) Take the average of the entire band to obtain the weighting coefficient K of the corresponding pixel position (i,j,B) ;

[0049] e′ (i,j,B) =x (i,j,B) -y (i,j,B)

[0050]

[0051] Where: M×N is the size of the dark background image.

[0052] S4: According to the obtained weighting coefficient K of the corresponding pixel position (i,j,B) Separate the fixed noise content N of the instrument band by band (i,j,B) ;

[0053] N (i,j,B) =K (i,j,B) × (i,j,B) / (dark_m) B

[0054] Where: (dark_m) B is the mean value of a single band of the dark background image.

[0055] S5: Perform weighted least squares filtering with embedded bilateral filtering on the sample image to obtain the sample image γ after filtering. (i,j,B) ;

[0056]

[0057] Where: (i,j) is the sample image, γ (i,j) is the sample image after filtering. is the sample image gradient, is the sample image gradient after filtering, ω x,(i,j) is the weight of the sample image along the X direction after filtering, ω y,(i,j) is the weight of the sample image along the Y direction after filtering, f BLF Represents the bilateral operator weight.

[0058] S6: The sample image after filtering is γ (i,j,B) The fixed noise content N of the isolated instrument (i,j,B) Subtract and get the final clear image X′ (i,j,B) ;

[0059] X′ (i,j,B) =γ(i,j,B) -N (i,j,B)

[0060] Completed the inhomogeneous noise calibration of brain tumor tissue hyperspectral data.

[0061] This gives us a clear image X′ (i,j,B) The grayscale distribution comparison of brain tumor tissue before and after denoising is shown in the figure below: Figure 2 As shown, it can be clearly seen that the image after denoising is clearer.

[0062] In step S5, weighted least squares filtering with embedded bilateral filtering is performed on the sample image, and the theoretical calculation method of the calculation formula is as follows:

[0063] The matrix representation of the sample image after filtering is as follows:

[0064] (χ-γ) T (χ-γ)+λ(χ T D x T A x D x χ+f BLF (γ T D y T A y D y γ))

[0065] Among them: A x A y is a k×k diagonal matrix, D x D y is the matrix representation of the discrete difference operator

[0066] The size of the dark background image is M×N. We introduce a variable p=M×N, then p is a one-dimensional vector. The gradient of the sample image after filtering is set to 0, and the equation is:

[0067] (I+λF BLF (L γ ))χ=γ

[0068] Where L γ =D T A x D x +D T A y D y , χ is the sample image, γ is the sample image after filtering, and the weight gradient design of the present invention does not use the logarithmic brightness of the WLS algorithm, but uses the gradient of the original image. It has been proven that this can achieve a smoothing effect more quickly. The weight design is as follows:

[0069]

[0070] Where: α x,(i,j) To correct the lateral gradient, α y,(i,j) To correct the longitudinal gradient, is the original lateral gradient, is the original longitudinal gradient, and ε is the small coefficient.

[0071] The embodiments described above are merely descriptions of specific implementation methods of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue, characterized in that: The following steps are involved: S1: Collect sample images and corresponding dark background images in the same environment; S2: Perform weighted least squares filtering on the dark background image; S3: Subtract the original dark background image from the dark background image processed by weighted least squares filtering to obtain edge information of the dark background image, and take the average of the edge information in the entire band to obtain the weighting coefficient of the corresponding pixel position; S4: Separate the fixed noise content of the instrument band by band according to the weighting coefficient of the corresponding pixel position; S5: Perform weighted least squares filtering with embedded bilateral filtering on the sample image; S6: Subtract the filtered sample image from the separated fixed noise content of the instrument to obtain a final clear image, thereby completing the non-uniform noise calibration of the brain tumor tissue hyperspectral data.

2. The method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to claim 1, characterized in that: Step S1 is specifically as follows: Collect sample images in the same environment (i,j,B) And the corresponding dark background image x (i,j,B) , where (i,j) is the spatial dimension position and B is the spectral dimension position.

3. The method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to claim 2, characterized in that: Step S2 is specifically as follows: Perform weighted least squares filtering on the dark background image to obtain the filtered dark background image y (i,j,B) ; in: is the gradient of the dark background image along the X direction after filtering, is the gradient of the dark background image along the Y direction after filtering, * is the pixel point, ω i is the weight of the dark background image in the X direction, ω j is the weight of the dark background image in the Y direction, and is the universal weight, ε is the small coefficient, λ is the regularization term coefficient, and y is the dark background image after filtering.

4. The method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to claim 3, characterized in that: Step S3 is specifically as follows: The original dark background image x (i,j,B) and the dark background image y after weighted least squares filtering (i,j,B) Subtract to get the edge information e′ of the dark background image (i,j,B) , and for the edge information e′ (i,j,B) Take the average of the entire band to obtain the weighting coefficient K of the corresponding pixel position (i,j,B) ; and' (i,j,B) =x (i,j,B) -and (i,j,B) Where: M×N is the size of the dark background image.

5. The method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to claim 4, characterized in that: Step S4 is specifically as follows: According to the weight coefficient K of the corresponding pixel position (i,j,B) Separate the fixed noise content N of the instrument band by band (i,j,B) ; N (i,k,B) =K (i,j,B) ×x (i,j,B) / (dark_m) B Where: (dark_m) B is the mean of a single band of the dark background image.

6. The method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to claim 5, characterized in that: Step S5 is specifically as follows: The sample image is processed by weighted least squares filtering with embedded bilateral filtering to obtain the filtered sample image γ (i,j,B) ; Where: (i,j) is the sample image, γ (i,j) is the sample image after filtering, is the sample image gradient, is the sample image gradient after filtering, ω x,(i,j) is the weight of the sample image along the X direction after filtering, ω y,(i,j) is the weight of the sample image along the Y direction after filtering, f BLF Represents the bilateral operator weight.

7. The method for calibrating non-uniform noise of hyperspectral data of brain tumor tissue according to claim 6, characterized in that: Step S6 is specifically as follows: The sample image after filtering is γ (i,j,B) The fixed noise content N of the isolated instrument (i,j,B) Subtract to get the final clear image X′ (i,j,B) ; X′ (i,j,B) =c (i,j,B) -N (i,j,B) Completed the inhomogeneous noise calibration of brain tumor tissue hyperspectral data.

Images