Fluorescent microscopic image enhancement method and device based on multi-saliency guide filtering fusion and storage medium

Abstract

The invention relates to a fluorescent microscopic image enhancement method and device based on multi-saliency guide filtering fusion and a storage medium. The method comprises the following steps: inputting a single-layer fuzzy pathological image I; performing correction algorithm on the single-layer fuzzy pathological imaging to obtain a pseudo exposure image sequence; respectively calculating saturation S in three channels for each pixel through a spatial linear saturation adjustment algorithm, carrying out normalization processing on saturation increment to obtain a saturation-adjusted pseudo-exposure image, and carrying out two-scale decomposition through mean filtering to obtain a base layer sequence and a detail layer sequence; the base layer sequence obtains a base layer fusion result through an energy significance guide filtering algorithm, and the detail layer sequence obtains a detail layer fusion result through a Gaussian significance guide filtering algorithm; and performing two-scale reconstruction on the base layer fusion result and the detail layer fusion result to obtain a single-layer deblurring pathological image. According to the method, the color fidelity of the sample can be improved, and the cell texture structure is clearer while noise interference is eliminated.

Description

Technical Field

[0001] The invention relates to an image enhancement method, in particular to a fluorescence microscopic image enhancement method, device and storage medium based on multi-saliency guided filtering fusion. Background Art

[0002] Digital pathology is an emerging pathological diagnosis method in the field of tumor detection that combines whole-slide scanning imaging and pathological image analysis algorithms. It is regarded as the gold standard for pathological diagnosis in the industry. However, a large amount of redundant information in scanning imaging will interfere with the pathological diagnosis process. This requires the use of pathological image analysis algorithms to mine effective deep-level pathological feature data through tasks such as feature detection, segmentation, and extraction. Pathological feature characterization provides a qualitative and quantitative description of the diseased tissue. The accuracy and richness of the features are crucial for the above-mentioned pathological image analysis, which helps to reveal the nature and progression of the disease. However, pathological feature characterization is largely affected by complex environmental interference such as uneven slice thickness, sample jitter, and background noise. Among them, color deviation and clarity reduction caused by uneven slice thickness are the most direct factors affecting the performance of digital pathology imaging systems. When light passes through pathological slices, its transmission and scattering characteristics determine the clarity and contrast of the imaging. When the slice thickness is uneven, the absorption and reflection of light in different areas of the same slice will be different, which will lead to uneven changes in brightness and darkness in the local imaging area, thereby reducing the contrast of digital pathological imaging and increasing the risk of missed detection of diseased cells. Therefore, it is very necessary to study enhancement methods suitable for microscopic imaging characteristics.

[0003] To address the problem of brightness inconsistency during imaging, common image enhancement methods include histogram equalization (HE), Retinex technology, and gamma correction. HE provides a simple and fast method that can expand the dynamic range of the image, thereby improving the overall global contrast. However, this method may lead to loss of image information and distortion of contours. The Retinex model is a mechanism that simulates the perception of color and color constancy by the human retina. Based on this model, researchers have developed a variety of algorithms, including single-scale Retinex, multi-scale Retinex, and multi-scale Retinex with color restoration, which aim to improve the color and contrast performance of images through different scale processing. Although the Retinex algorithm generally improves the quality and contrast of the image, it may unconsciously introduce noise during the brightness enhancement process, which may lead to unclear edges and color deviation. Gamma correction can enhance the contrast of the image, but it may introduce distortion in low-light images. In the field of learning-based enhancement methods, researchers have also conducted a lot of research. Although these methods can restore edge details in low-light images, in complex environments such as digital pathology imaging, they cannot achieve the optimal enhancement effect due to excessive enhancement caused by the influence of brightness inconsistency. Summary of the invention

[0004] In order to solve the problem that the optimal enhancement effect cannot be achieved due to excessive enhancement caused by the influence of brightness inconsistency, the present invention provides a fluorescence microscopic image enhancement method based on multi-saliency guided filtering fusion, which can simply and efficiently achieve pathological image enhancement. The specific technical solution is:

[0005] A fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion comprises the following steps:

[0006] S1, input single-layer fuzzy pathological imaging I;

[0007] S2, the single-layer fuzzy pathological image I is subjected to the Gamma correction algorithm to obtain a pseudo-exposure image sequence E = {I 1 (x),I 2 (x),…,I K (x)};

[0008] Where E represents fuzzy pathological imaging I k (x), k∈[1,K] is the pseudo-exposure image sequence obtained by contrast adjustment;

[0009] x represents the pixel of the pseudo-exposure image;

[0010] K represents the number of random values ​​in the Gamma correction algorithm; S3, for the E'th k,k∈{1,…,K} Each pixel of the pseudo-exposure images The saturation S in the three channels of R, G, and B is calculated by the spatial linear saturation adjustment algorithm, and the saturation increment is normalized in the range of [0,100] to obtain the pseudo-exposure image after saturation adjustment. For each pseudo exposure image E' k , k∈{1,…,K} are decomposed into two scales by mean filter Z to obtain the base sequence E' B =(B 1 ,B 2 ,…,B K ) and detail layer sequence E' D =(D 1 ,D 2 ,…,D K );

[0011] In the formula, E k represents any pseudo-exposure image of sequence E after contrast adjustment;

[0012] E k (x) represents any pseudo exposure image E k The set of corresponding pixels x;

[0013] Represents any pseudo exposure image E k R channel;

[0014] Represents any pseudo exposure image E k G channel;

[0015] Represents any pseudo exposure image E k B channel;

[0016] E' k represents any pseudo-exposure image of sequence E after saturation adjustment;

[0017] Represents any pseudo exposure image E' k R channel;

[0018] Represents any pseudo exposure image E' k G channel;

[0019] Represents any pseudo exposure image E' k B channel;

[0020] E' k ,k∈{1,…,K} represents any pseudo-exposure image of sequence E after saturation adjustment;

[0021] B k It means grassroots;

[0022] D k Represents the detail layer;

[0023] S4, base sequence E' B =(B 1 ,B 2 ,…,B K ) The base layer fusion result F is obtained by energy saliency guided filtering algorithm B , detail layer sequence E' D =(D 1 ,D 2 ,…,D K ) The detail layer fusion result F is obtained by Gaussian saliency guide filtering algorithm D ;

[0024] S5. Basic fusion result F B And the detail layer fusion result F D Perform two-scale reconstruction to obtain single-layer deblurred pathological imaging I F .

[0025] Preferably, the gamma correction algorithm in step S2 comprises the following steps:

[0026] Adjust the contrast of the global image through power function transformation, that is, I(x)→α·I(x) γ ; Wherein, I(x) represents any pixel of a single-layer blurred pathology imaging; → represents a power function transformation process; α and γ represent positive constants, and when γ < 1, the brighter intensity is compressed and the darker intensity is expanded, and when γ > 1, the brighter intensity is distributed in a wider range and the darker intensity is mapped to the compressed interval, thereby obtaining a pseudo-exposure image sequence.

[0027] 3. The fluorescence microscopy image enhancement method based on multi-saliency guided filtering fusion according to claim 1, characterized in that:

[0028] The spatial linear saturation adjustment algorithm in step S3 comprises the following steps:

[0029] S301: Each pixel of the pseudo exposure image The maximum RGB component is rgb max =max(max(R,G),B), the minimum RGB component is rgb min =min(min(R,G),B);

[0030] When Δ=(rgb max -rgb min) / 255>0, the calculation formula of the saturation S of each pixel is:

[0031]

[0032] Where, value = (rgb max +rgb min ) / 255, L=value / 2;

[0033] S302. In the range of [0,100], the calculation formula for normalizing the saturation increment is:

[0034] E' k (x) = E k (x)+(E k (x)-L×255)×α,

[0035] in,

[0036] Percent represents a normal number. Represents the pseudo-exposed image after saturation adjustment.

[0037] Preferably, in step S3, for each pseudo exposure image E' k , k∈{1,…,K} are decomposed into two scales by the mean filter Z with a filter size of n×n to obtain the base layer B k and detail layer D k , the formula is as follows:

[0038]

[0039] Where n represents the size of the filter Z, (x, y) represents the pixel E' k (x) coordinates, i and j represent the filter Z relative to the center pixel E' k (x) offset, get the base layer B k Finally, the corresponding detail layer formula is obtained as follows:

[0040] D k =E' k -B k .

[0041] Preferably, the base sequence E' in step S4 B =(B 1 ,B 2 ,…,B K ) The base layer fusion result F is obtained by energy saliency guided filtering algorithm B , including the following steps:

[0042] S401, the base Bk The formula for obtaining Gabor coefficients by convolution with Gabor wavelet is:

[0043]

[0044] in, represents Gabor coefficient;

[0045] m 1 represents the scale parameter, which controls the scaling of the Gabor function g(x,y) in frequency;

[0046] n 1 Represents the rotation angle, which controls the direction of the Gabor function g(x,y);

[0047] represents the Gabor function g(x,y) through m 1 Scaling and n 1 Gabor wavelet obtained by direction rotation;

[0048] Thus, we can get the base layer B k The formula for the Gabor energy at the pixel point (x, y) is:

[0049]

[0050] S402, base sequence E' B =(B 1 ,B 2 ,…,B K ) is compared with the Gabor energy at the corresponding pixel point (x, y), and the initial weight map P of the base sequence fusion can be determined. k ,k∈{1,…,K}The formula is:

[0051]

[0052] S403, combined with the guide filter theory, the refined weight map of the base sequence fusion Defined as:

[0053]

[0054] In the formula, G r,ψ (P,I) represents the guide filter operation;

[0055] r represents the filter size of the guide filter;

[0056] ψ represents the blur level of the guide filter;

[0057] P and I represent the input image and guide image of the guide filter respectively;

[0058] S404, refine the weight map Normalize and get the base fusion result F by weighted average B The formula is:

[0059]

[0060] Furthermore, the Gabor wavelet in step S401 is obtained by using the Gabor function g(x,y) 1 Scaling and n 1 Direction rotation solution, Gabor function is defined as the modulation of Gaussian function, and its expression in two-dimensional space is:

[0061]

[0062] Where g represents the Gabor wavelet; σ is the standard deviation of the Gaussian envelope, which controls the spatial expansion of the Gabor function, and w 0 is the center frequency, which controls the oscillation rate of the Gabor function. The Gabor wavelet transform involves the m 1 Scaling and n 1 Direction rotation, the specific formula is:

[0063]

[0064] Among them, m 1 Represents the scale parameter, which controls the scaling of the Gabor function in frequency, n 1 Indicates the rotation angle, which controls the direction of the Gabor function.

[0065] Furthermore, the refined weight map of the base sequence fusion in step S403 The solution is obtained through the guide filter theory. The specific process is as follows: The guide filter theory assumes that the filter output O of the guide filter is the guide image I in the local window w centered at pixel k. k The linear transformation formula on is:

[0066]

[0067] Among them, O i represents the filtered output of the wizard filtering operation;

[0068] I represents the guide image of the guide filtering operation;

[0069] i represents any pixel in the guide image I;

[0070] k represents a pixel in the guide image I;

[0071] w krepresents the local window of the guide image I centered at pixel k;

[0072] a k and b k They represent the linear coefficients of the guide filter linear transformation operation respectively;

[0073] w k is a window of size (2r+1)×(2r+1), where r represents the window radius and a is the linear coefficient. k and b k is the local window w k The constant in , considering that the filtering principle is affected by the multi-valued output, that is, if the central pixel k is just in the local window w centered on pixel i and with a size of (2r+1)×(2r+1) i On the other hand, different local windows w k contains pixel i, and due to the local window w k changes with pixel k, resulting in coefficient a k and b k The corresponding changes also occur. At this time, the filter output O i It may not be unique. In order to solve the impact of output multi-value, the linear coefficient a k and b k Take the average, and the estimated value formula of the wizard filter output is:

[0074]

[0075] In the formula, and Represents the linear coefficient a k and b k The average value in the local window w;

[0076] in, Adopt G r,ψ (P, I) represents the guide filter operation, where the parameters r and ψ respectively determine the filter size and blur level of the guide filter, P and I respectively represent the input image and the guide image. At this time, the base layer B in step S401 is k As a guide image, the initial weight map P in step S402 is k , k∈{1,…,K} as the input image, the refined weight map of the base sequence fusion can be obtained by the guided filtering operation

[0077] Among them, w i represents the local window centered at pixel i;

[0078] |w| represents the size of the local window w.

[0079] Preferably, the detail layer sequence E' in step S4 D =(D 1 ,D 2 ,…,D K ) The detail layer fusion result F is obtained by Gaussian saliency guide filtering algorithm D , including the following steps:

[0080] The detail layer D k The high-pass image H is obtained by Laplace filtering k The formula is:

[0081] H k =D k *L,

[0082] Among them, L represents a 3×3 Laplace filter, and then the high-pass image H k The absolute value of is filtered through a Gaussian low-pass filter g to construct a Gaussian saliency map The formula is:

[0083]

[0084] Among them, g represents a 5×5 Gaussian low-pass filter, which is used for the detail layer sequence E' D =(D 1 ,D 2 ,…,D K ) is compared with the Gaussian saliency map at the corresponding pixel point (x, y) to determine the initial weight map for the fusion of the detail layer sequence The formula is:

[0085]

[0086] Combined with the guide filter theory, the refined weight map W of detail layer sequence fusion k D Defined as:

[0087]

[0088] Refine the weight map Normalize and obtain the detail layer fusion result F through weighted average D The formula is:

[0089]

[0090] A fluorescence microscopic image enhancement device based on multi-saliency guided filter fusion comprises: a processor; a memory on which a computer program that can be run on the processor is stored; wherein the computer program, when executed by the processor, implements the steps of the above-mentioned fluorescence microscopic image enhancement method based on multi-saliency guided filter fusion.

[0091] A computer-readable storage medium stores a data processing program, and when the data processing program is executed by a processor, the steps of the fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion are implemented.

[0092] Compared with the prior art, the present invention has the following beneficial effects:

[0093] The present invention provides a fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion, which adopts the fusion rule of energy-saliency guided filtering to retain the non-significant texture and brightness information of the local weak exposure degree in the pseudo-exposure image, and adopts the Gaussian saliency guided filtering of local window averaging to maintain the significant features in the pseudo-exposure image. The above measures can effectively improve the color fidelity of the sample, and make the cell texture structure clearer while effectively eliminating noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 This is a schematic diagram of the guided filter fusion enhancement based on multiple saliency representations;

[0095] Figure 2 It is a schematic diagram of the process of an embodiment;

[0096] Figure 3 Schematic diagram of the influence of guide filtering on feature preservation in multi-layer pathology imaging fusion process;

[0097] Figure 4 Schematic diagram of the impact of local average-based saliency detection on weight accuracy;

[0098] Figure 5 Schematic diagram of the impact of Gabor filter-based saliency detection on weight accuracy;

[0099] Figure 6 Schematic diagram of subjective evaluation of multi-saliency GIF fusion strategy in ablation experiment;

[0100] Figure 7 The first set of comparison diagrams of the performance of different image enhancement methods are shown;

[0101] Figure 8 A second set of comparison diagrams showing the performance of different image enhancement methods;

[0102] Fig. 9 The third set of comparison diagrams showing the performance of different image enhancement methods. DETAILED DESCRIPTION

[0103] The present invention will be further described with reference to the accompanying drawings.

[0104] It should be noted that the following embodiments assume that strict registration has been completed between the pseudo-exposure images. This assumption is a typical case of this type of fusion enhancement scheme and is not a special requirement of the present invention. The following embodiments refer to the guided filtering method as the GIF algorithm.

[0105] A fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion is proposed. This method overcomes the traditional enhancement algorithm's prior estimation of scene depth and complex depth mapping process. It can not only improve the color fidelity of the sample, but also make the cell texture structure clearer while effectively eliminating noise interference, thereby obtaining fluorescence microscopy images that meet the requirements of pathological diagnostic analysis.

[0106] Explanation of each formula:

[0107] Supplementary explanation of each parameter in the formula

[0108] E={I 1 (x),I 2 (x),…,I K (x)}

[0109] E stands for fuzzy pathological imaging k (x), k∈[1,K] is the pseudo-exposure image sequence obtained by contrast adjustment;

[0110] x represents the pixel of the pseudo-exposure image;

[0111] K represents the number of pseudo-exposure image sequences obtained by randomly taking values ​​of γ in the Gamma correction algorithm.

[0112]

[0113] E k represents any pseudo-exposure image of sequence E after contrast adjustment;

[0114] x represents the pixel of the pseudo-exposure image;

[0115] E k (x) represents any pseudo exposure image E k The set of corresponding pixels x;

[0116] Represents any pseudo exposure image E k R channel;

[0117] Represents any pseudo exposure image E k G channel;

[0118] Represents any pseudo exposure image E k B channel.

[0119] According to the above requirements, the following formula is supplemented

[0120]

[0121] E' k represents any pseudo-exposure image of sequence E after saturation adjustment;

[0122] Represents any pseudo exposure image E' k R channel;

[0123] Represents any pseudo exposure image E' k G channel;

[0124] Represents any pseudo exposure image E' k B channel.

[0125] E' k , k∈{1,…,K} represents any pseudo-exposure image of sequence E after saturation adjustment; K represents the number of pseudo-exposure image sequences obtained by randomly taking values ​​of γ in the Gamma correction algorithm.

[0126] E' B =(B 1 ,B 2 ,…,B K ) represents the base sequence obtained by two-scale decomposition of sequence E;

[0127] K represents the number of pseudo-exposure image sequences obtained by randomly taking values ​​of γ in the Gamma correction algorithm.

[0128] E' D =(D 1 ,D 2 ,…,D K ) represents the corresponding detail layer sequence obtained by two-scale decomposition of sequence E; K represents the number of pseudo-exposure image sequences obtained by randomly taking values ​​of γ in the Gamma correction algorithm.

[0129] RGB max =max(max(R,G),B) represents the maximum pixel value of the extracted pixel in the three channels of R, G, and B. min =min(min(R,G),B) represents the minimum pixel value of the extracted pixel in the three channels of R, G, and B.

[0130]

[0131] B k It means grassroots;

[0132] (x,y) represents pixel E' k(x) coordinate;

[0133] i and j represent the filter Z relative to the center pixel E' k (x) offset;

[0134] n represents the size of the filter Z.

[0135]

[0136] represents Gabor coefficient;

[0137] m 1 represents the scale parameter, which controls the scaling of the Gabor function g(x,y) in frequency;

[0138] n 1 Represents the rotation angle, which controls the direction of the Gabor function g(x,y);

[0139] B k It means grassroots;

[0140] represents the Gabor function g(x,y) through m 1 Scaling and n 1 The Gabor wavelet obtained by rotating the direction.

[0141]

[0142] Represents the base sequence E' B Gabor energy;

[0143] P k Represents Gabor energy The initial weight map obtained after comparison;

[0144] (x,y) represents the base layer B k Pixels of

[0145] Indicates the base layer B k Gabor energy.

[0146]

[0147] G r,ψ (P,I) represents the guide filter operation;

[0148] r represents the filter size of the guide filter;

[0149] ψ represents the blur level of the guide filter;

[0150] P and I represent the input image and guide image of the guide filter respectively;

[0151] Represents the refined weight map obtained by the guided filtering operation.

[0152]

[0153] g represents Gabor wavelet;

[0154] σ is the standard deviation of the Gaussian envelope, which controls the spatial expansion of the Gabor function g(x,y); w 0 is the center frequency, which controls the oscillation rate of the Gabor function g(x,y).

[0155]

[0156] represents the Gabor function g(x,y) through m 1 Scaling and n 1 The Gabor wavelet is obtained by rotating the Gabor function in the same direction; σ is the standard deviation of the Gaussian envelope, which controls the spatial expansion of the Gabor function g(x,y);

[0157] m 1 represents the scale parameter, which controls the scaling of the Gabor function g(x,y) in frequency;

[0158] n 1 Represents the rotation angle, which controls the direction of the Gabor function g(x,y).

[0159]

[0160] O i represents the filtered output of the wizard filtering operation;

[0161] I represents the guide image of the guide filtering operation;

[0162] i represents any pixel in the guide image I;

[0163] k represents a pixel in the guide image I;

[0164] w k represents the local window of the guide image I centered at pixel k;

[0165] a k and b k They respectively represent the linear coefficients of the guide filter linear transformation operation.

[0166]

[0167] I represents the guide image of the guide filtering operation;

[0168] and Represents the linear coefficient a k and b k The average value in the local window w;

[0169] O i Represents the filtered output of the wizard filter operation.

[0170]

[0171] a k Represents the linear coefficients of the guide filter linear transform operation;

[0172] w i represents the local window centered at pixel i;

[0173] w| represents the size of the local window w;

[0174] k represents a pixel in the guide image I;

[0175] Indicates the linear coefficient a k The average value in the local window w.

[0176]

[0177] b k Represents the linear coefficients of the guide filter linear transform operation;

[0178] w i represents the local window centered at pixel i;

[0179] w| represents the size of the local window w;

[0180] k represents a pixel in the guide image I;

[0181] Indicates the linear coefficient b k The average value in the local window w.

[0182] P k ,k∈{1,…,K}

[0183] P k Represents the base sequence E' B Gabor energy The initial weight map obtained after comparison;

[0184] K represents the number of pseudo-exposure image sequences obtained by randomly taking values ​​of γ in the Gamma correction algorithm.

[0185] like Figure 1 and Figure 2 As shown, a fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion includes the following steps:

[0186] Input single layer fuzzy pathology imaging I;

[0187] The single-layer fuzzy pathological image I is converted into a pseudo-exposure image sequence E = {I 1 (x),I 2 (x),…,I K (x)}, K represents the number of random values ​​of parameter γ in the Gamma correction algorithm;

[0188] For each pixel of the kth, k∈{1,L,K}th pseudo-exposure image in the pseudo-exposure image sequence The saturation S in the three channels of R, G, and B is calculated by the spatial linear saturation adjustment algorithm, and the saturation increment is normalized in the range of [0,100] to obtain the pseudo-exposure image after saturation adjustment. For each pseudo exposure image E' k , k∈{1,…,K} are decomposed into two scales by mean filter Z to obtain the base sequence E' B =(B 1 ,B 2 ,…,B K ) and detail layer sequence E' D =(D 1 ,D 2 ,…,D K );

[0189] Base sequence E' B =(B 1 ,B 2 ,…,B K ) The base layer fusion result F is obtained by energy saliency guided filtering algorithm B , detail layer sequence E' D =(D 1 ,D 2 ,…,D K ) The detail layer fusion result F is obtained by Gaussian saliency guide filtering algorithm D ;

[0190] The base fusion result F B And the detail layer fusion result F D Perform two-scale reconstruction to obtain single-layer deblurred pathological imaging I F .

[0191] The specific implementation methods of the above algorithms are as follows:

[0192] Gamma correction algorithm:

[0193] Adjust the contrast of the global image through power function transformation, that is, I(x)→α·I(x) γ , where I(x) represents any pixel of the single-layer blurred pathological imaging, → represents the power function transformation process, α and γ represent positive constants, and when γ < 1, the brighter intensity is compressed and the darker intensity is expanded, and when γ > 1, the brighter intensity is distributed in a wider range and the darker intensity is mapped to the compressed interval, thereby obtaining the pseudo-exposure image sequence E = {I 1 (x),I 2 (x),…,I K (x)}, where K represents the number of random values ​​of the parameter γ in the Gamma correction algorithm.

[0194] The change of I(x) is adjusted with γ (γ<1 or γ>1) to obtain pseudo exposure image sequences with different exposure levels. Therefore, the number of random values ​​of γ determines the number K of pseudo exposure image sequences.

[0195] Spatial linear saturation adjustment algorithm:

[0196] Each pixel of the pseudo-exposure image The maximum RGB component is rgb max =max(max(R,G),B), and the minimum RGB component is rgb min =min(min(R,G),B).

[0197] When Δ=(rgb max -rgb min ) / 255>0, the saturation S of each pixel is:

[0198]

[0199] Where, value = (rgb max +rgb min ) / 255, L=value / 2.

[0200] In the range [0,100], the saturation increment is normalized as:

[0201] E' k (x) = E k (x)+(E k (x)-L×255)×α,

[0202] in, Percent represents a normal number. Represents the pseudo-exposed image after saturation adjustment.

[0203] Two-scale decomposition of pseudo-exposure image:

[0204] Each pseudo exposure image E' k , k∈{1,…,K} are decomposed into two scales by the mean filter Z with a filter size of n×n to obtain the base layer B k and detail layer D k :

[0205]

[0206] Where (x, y) represents pixel E' k (x) coordinates, i and j represent the filter Z relative to the center pixel E' k (x) offset. Get the base layer B k After that, the corresponding detail layer is finally obtained:

[0207] D k =E' k -B k ,

[0208] Energy saliency guided filtering algorithm:

[0209] The base B k The Gabor coefficients obtained by convolution with Gabor wavelet are:

[0210]

[0211] Among them, Gabor wavelet is composed of Gabor function g(x,y) 1 Scaling and n 1 Direction rotation is obtained, thus the base layer B can be obtained k The Gabor energy at the pixel point (x, y) is:

[0212]

[0213] For the base sequence E' B =(B 1 ,B 2 ,…,B K ) is compared with the Gabor energy at the corresponding pixel point (x, y), and the initial weight map P of the base sequence fusion can be determined. k ,k∈{1,…,K} is:

[0214]

[0215] Combined with the guide filter theory, the refined weight graph W of the base sequence fusion k B for:

[0216]

[0217] Refine the weight map Normalize and get the base fusion result F by weighted average B for:

[0218]

[0219] Gaussian saliency guided filtering algorithm:

[0220] The detail layer D k The high-pass image H is obtained by Laplace filtering k for:

[0221] H k =D k *L,

[0222] Among them, L represents a 3×3 Laplace filter, and then the high-pass image H k The absolute value of is filtered through a Gaussian low-pass filter g to construct a Gaussian saliency map for:

[0223]

[0224] Among them, g represents a 5×5 Gaussian low-pass filter, which is used for the detail layer sequence E' D =(D 1 ,D 2 ,…,D K ) is compared with the Gaussian saliency map at the corresponding pixel point (x, y) to determine the initial weight map for the fusion of the detail layer sequence for:

[0225]

[0226] Combined with the guided filter theory, the refined weight map of detail layer sequence fusion for:

[0227]

[0228] Refine the weight map Normalize and get the base fusion result F by weighted average D for:

[0229]

[0230] Two-scale reconstruction:

[0231] Get the base fusion result F B And the detail layer fusion result F D After that, the final enhanced image can be obtained through two-scale reconstruction:

[0232] I F =F B +F D .

[0233] In order to verify the effectiveness of the guide filter in the proposed method for preserving the features of the multi-layer pathology imaging fusion process, the experiment was designed to use a color natural image as a guide image and compare the edge preservation performance of different filtering algorithms. The experimental setting conditions are as follows:

[0234] (1) The spatial resolution of the color natural images used in the experiment is 256×256;

[0235] (2) In the present invention, the comparison of different filtering algorithms includes: bilateral filtering, rotation filtering and the guide filtering adopted in this embodiment;

[0236] (3) All experiments were conducted in an experimental environment with a computer configuration of 3.3 GHz CPU and 16.0 GHz RAM, MATLAB version R2017a and Win7 64-bit operating system.

[0237] The influence of guide filtering on feature preservation in multi-layer pathology imaging fusion process is shown in the experiment. Figure 3 As shown. Figure 3 It can be seen that the filter input Figure 3 (b) has missing edges and is affected by speckle noise. After bilateral filtering, the filtered output Figure 3 The speckle noise in (c) is almost eliminated, but there is still edge blur. After rotation filtering, the filter output Figure 3 (d) Some noise remains and the filter edge is still missing. After the guide filter, the filter output Figure 3 The noise in (e) is almost completely eliminated and the integrity of the filter edge is maintained. In summary, the guide filter algorithm used in this embodiment has better edge preservation performance and better noise suppression performance.

[0238] In order to verify that the Gaussian saliency GIF in the proposed method is more suitable for the saliency representation of the detail layer, while the energy saliency GIF is more suitable for the saliency representation of the base layer, the experimental design takes two registered multimodal medical images as examples to compare and analyze the effects of different saliency representations on the weight accuracy in the guided filtering process. The experimental setting conditions are as follows:

[0239] (1) Multimodal brain images were obtained from the Whole Brain Atlas Medical Image Database (WBAMI) created by Harvard Medical School, and the spatial resolution of the images was 256 × 256;

[0240] (2) All experiments were conducted in an experimental environment with a computer configuration of 3.3 GHz CPU and 16.0 GHz RAM, MATLAB version R2017a and Win7 64-bit operating system.

[0241] The experiment on the influence of different saliency representations on weight accuracy in guided filtering is as follows Figure 4 and Figure 5 As shown by Figure 4 and Figure 5 It can be seen that the weight of the significant features in the initial weight map obtained by Gaussian saliency GIF is more prominent, while the weight of the non-significant texture is missing. The weight distribution in the initial weight map obtained by energy saliency GIF is more uniform, and the proportion of non-significant texture is significantly increased.

[0242] In summary, Gaussian saliency GIF is more conducive to maintaining the salient features in the detail layer, while energy saliency GIF can effectively improve the accuracy of the refined weight map used for base layer image fusion, and is more conducive to maintaining the uniformity and reliability of local texture and brightness changes in enhanced pathological images.

[0243] Figure 4 The figure is a schematic diagram of the impact of the existing local average-based saliency detection on weight accuracy, where the initial weight map and the refined weight map correspond to the weight distribution and characterize the weight accuracy. The existing method comes from the paper "Image Fusion with Guided Filtering" - Shutao Li, Xudong Kang and Jianwen Hu. Formula (13) in the paper uses the local average of high-pass filtering to construct a saliency map and is applied in medical image fusion.

[0244] In order to verify the rationality of the multi-significant GIF fusion strategy in the method proposed in the present invention, the experimental design designs an ablation experiment by controlling variables, and compares and analyzes the effects of the Gaussian significant GIF fusion strategy and the multi-significant GIF fusion strategy proposed in the present invention on the image enhancement performance during the guide filtering process. The experimental setting conditions are as follows:

[0245] (1) The experiment used a fluorescent digital pathology section scanning device with a system model of KF-FL-400 to obtain cytological fluorescence microscopic imaging samples for tissue sampling, and the spatial resolution of pathological imaging was 2048×1024;

[0246] (2) In order to comprehensively evaluate the enhancement effect of the algorithm, objective evaluation indicators include: peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and structural similarity (SSIM);

[0247] (3) All experiments were conducted in an experimental environment with a computer configuration of 3.3 GHz CPU and 16.0 GHz RAM, MATLAB version R2017a and Win7 64-bit operating system.

[0248] In addition, the enhancement performance analysis is carried out by combining subjective evaluation and objective evaluation. The enhancement results of subjective evaluation and the local magnification are shown in Figure 6 As shown, Figure 6 (a1), Figure 6 (a2) Figure 6 (a) Local magnification of Gaussian saliency GIF fusion enhanced imaging; Figure 6 (b1), Figure 6 (b2) Figure 6 (b) Local magnification of multi-saliency GIF fusion enhanced imaging. The objective evaluation performance indicators are shown in the objective evaluation table of multi-saliency GIF fusion strategy in the ablation experiment, where the underlined font represents the optimal indicator. Figure 6 (a) It can be seen that the brightness and contrast of pathological imaging are improved after Gaussian saliency GIF fusion, but there is obvious uneven imaging brightness, such as Figure 6 As shown in (al), there is obvious occlusion in the local area where multiple layers of cells overlap, making it difficult to accurately distinguish the cytoplasm and the cell nucleus, as shown in Figure 6 As shown in (a2), it is not conducive to the observation of cell morphology and the counting of cell numbers in the subsequent diagnosis process, and will increase the risk of missed tumor detection. Figure 6 (b) is the pathological imaging obtained by the method proposed in this embodiment, which has better brightness uniformity and noise suppression performance. Figure 6 (b1) shows that the number of cell nuclei can still be accurately and clearly distinguished in complex areas where multiple layers of cells overlap, such as Figure 6 As shown in (b2).

[0249] It can be seen from the objective evaluation table of the multi-significant GIF fusion strategy in the ablation experiment that the source pathological imaging is A, the enhanced image obtained by the Gaussian significant GIF fusion strategy after local window averaging is B, and the enhanced image obtained by the multi-significant GIF fusion strategy of this embodiment is C. Compared with the image enhanced by Gaussian significant GIF fusion (AB), after the multi-significant GIF fusion enhancement (AC) of this embodiment (comparison between the third row and the second row), the PSNR, RMSE and SSIM indicators are improved by 4.35%, 43.25% and 0.12% respectively.

[0250] Objective evaluation table of multiple saliency GIF fusion strategies in ablation experiments:

[0251] Source Pathology Imaging Fusion Enhanced Imaging PSNR RMSE SSIM A AB 56.98 0.1304 0.9952 A AC <![CDATA[ 59.46 ]]> <![CDATA[ 0.0740 ]]> <![CDATA[ 0.9964 ]]>

[0252] In summary, the design of the multi-significance GIF fusion strategy in the method proposed in the present invention is more reasonable and can effectively avoid the risk of missed detection in tumor detection.

[0253] In order to verify the advantages of the enhancement method based on multi-significance GIF fusion in the present invention for clarifying fluorescence microscopy images, a variety of the latest image enhancement methods were used to compare and analyze with the method proposed in the present invention. The experimental setting conditions are as follows:

[0254] (1) The experiment used a fluorescent digital pathology section scanning device with a system model of KF-FL-400 to obtain cytological fluorescence microscopic imaging samples for tissue sampling, and the spatial resolution of pathological imaging was 2048×1024;

[0255] (2) In order to comprehensively evaluate the enhancement effect of the algorithm, three indicators, peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and structural similarity (SSIM), are used for quantitative analysis;

[0256] (3) Image enhancement and contrast methods include: MSR-based enhancement method, MSRCP multi-scale Retinex enhancement method with color protection, HLRP enhancement method based on super Laplace reflectance prior, and DET-MDH enhancement method based on double exponential transform and multi-scale detail highlighting;

[0257] (4) Algorithm parameter settings: The noise random variable e in the HLRP algorithm model is set to 0.9, and the parameter settings of other algorithm models are consistent with the original settings;

[0258] (5) All experiments were conducted in an experimental environment with a computer configuration of 3.3 GHz CPU and 16.0 GHz RAM, MATLAB version R2017a and Win7 64-bit operating system.

[0259] In addition, the enhancement performance analysis is carried out by combining subjective evaluation and objective evaluation. The enhancement results of subjective evaluation and the local magnification are shown in Figures 7 to 9 As shown, the first group of experiments Figure 7 (a1) is Figure 7 (a) Local magnification of source pathology imaging, Figure 7 (b2) Figure 7 (b) Local magnification of MSR enhancement method. Figure 7 (c1) is Figure 7 (c) Local enlarged image of the MSRCP enhancement method. Figure 7 (d1) is Figure 7 (d) Local magnification of HLRP enhancement method, Figure 7 (e1) is Figure 7 (e) Local magnification of the DET-MDH enhancement method. Figure 7 (f1) is Figure 7 (f) A partial enlarged view of the method proposed by the present invention; Figure 8 (a1) is Figure 8 (a) Local magnification of source pathology imaging, Figure 8 (b2) Figure 8 (b) Local magnification of MSR enhancement method. Figure 8 (c1) is Figure 8 (c) Local enlarged image of the MSRCP enhancement method. Figure 8 (d1) is Figure 8 (d) Local magnification of HLRP enhancement method, Figure 8 (e1) is Figure 8 (e) Local magnification of the DET-MDH enhancement method. Figure 8 (f1) is Figure 8 (f) A partial enlarged view of the method proposed by the present invention; Fig. 9 (a1) is Fig. 9 (a) Local magnification of source pathology imaging, Fig. 9 (b2) Fig. 9 (b) Local magnification of MSR enhancement method. Fig. 9 (c1) is Fig. 9 (c) Local enlarged image of the MSRCP enhancement method. Fig. 9 (d1) is Fig. 9 (d) Local magnification of HLRP enhancement method, Fig. 9 (e1) is Fig. 9 (e) Local magnification of the DET-MDH enhancement method. Fig. 9 (f1) is Fig. 9 (f) A partial enlarged image of the method proposed by the present invention. The performance indicators of objective evaluation are shown in the overall objective statistical result table of different image enhancement methods. It can be seen from the comparison result table of the three groups of experiments that compared Figure 7 (a1), Figure 8 (a1) and Fig. 9 For the source pathology imaging shown in (a1), the color fidelity of the MSR algorithm result is good, but the blur is high and the recognition of cell features is low, such as Figure 7 (b1), Figure 8 (b1) and Fig. 9 (b1) The MSRCP algorithm results have obvious color distortion and no significant improvement in pixel contrast, as shown in Figure 7 (c1), Figure 8 (c1) and Fig. 9 (c1) As shown in the figure. Although the pixel contrast of the results of HLRP and DET-MDH algorithms is significantly improved, the low color distinction between different types of cells reduces the recognition of cell classification. At the same time, the processing results of both algorithms are greatly affected by noise, such as Figure 7(d1)(e1), Figure 8 (d1)(e1) and Fig. 9 (d1)(e1) In comparison, the processing result of the algorithm proposed in the present invention not only has higher color fidelity, but also effectively eliminates noise interference and makes the cell texture structure clearer, as shown in Figure 2. Figure 7 (f1), Figure 8 (f1) and Fig. 9 As shown in (f1).

[0260] The overall objective statistical results of different image enhancement methods are shown in the table below:

[0261] Comparison Method MSR MSRCP HLRP DET-MDH This embodiment PSNR 55.51 58.90 53.45 53.38 <![CDATA[ 59.46 ]]> RMSE 0.1831 0.0853 0.2943 0.2989 <![CDATA[ 0.0740 ]]> SSIM <![CDATA[ 0.9967 ]]> 0.9966 0.9883 0.9966 0.9964

[0262] As can be seen from the above table, the PSNR and RMSE indicators of the method proposed in this embodiment can still maintain the best, and the evaluation results can be improved by up to 11.39% and 75.24% compared with the relevant algorithms. Although the evaluation result of the SSIM indicator is comparable to that of the relevant algorithms, the method proposed in this embodiment can still maintain a high color fidelity and recognition of cell feature information based on source pathology imaging.

[0263] In summary, the method proposed in this embodiment can effectively suppress the influence of color deviation and clarity reduction on the characterization of pathological characteristics, and has high application value for improving the accuracy and resolution of subsequent digital pathology imaging systems.

[0264] A fluorescence microscopic image enhancement device based on multi-saliency guided filter fusion comprises: a processor; a memory on which a computer program that can be run on the processor is stored; wherein the computer program, when executed by the processor, implements the steps of the above-mentioned fluorescence microscopic image enhancement method based on multi-saliency guided filter fusion.

[0265] The memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines, such as through a bus connection. The memory stores computer-executable instructions for implementing the data access control method, including at least one software function module that can be stored in the memory in the form of software or firmware. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory.

[0266] The memory may be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc. The memory is used to store programs, and the processor executes the programs after receiving the execution instruction.

[0267] The processor may be an integrated circuit chip having the ability to process signals. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. The methods, steps, and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0268] A computer-readable storage medium stores a data processing program, and when the data processing program is executed by a processor, the steps of the fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion are implemented.

[0269] The embodiments of the present invention are described with reference to flowcharts of methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate means for implementing the functions specified in the flowchart and / or.

[0270] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in the flowchart.

[0271] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operation steps are executed on the computer or other programmable terminal device to produce computer-implemented processing, whereby the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in the flowchart.

[0272] The technical principle of the present invention is described above in conjunction with specific embodiments. These descriptions are only for explaining the principle of the present invention and cannot be interpreted as limiting the protection scope of the present invention in any way. Based on the explanations herein, those skilled in the art can associate other specific implementations of the present invention without paying creative labor, and these methods will fall within the protection scope of the claims of the present invention.

Claims

1. A fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion, characterized in that: The following steps are involved: S1, input single-layer fuzzy pathological imaging I; S2, the single-layer fuzzy pathological image I is subjected to the Gamma correction algorithm to obtain the pseudo-exposure image sequence E = {I1(x), I2(x), ..., I K (x)}; Where E represents fuzzy pathological imaging I k (x), k∈[1,K] is the pseudo-exposure image sequence obtained by contrast adjustment; x represents the pixel of the pseudo-exposure image; K represents the number of random values ​​in the Gamma correction algorithm; S3, for the E'th k ,k∈{1,…,K} Each pixel of the pseudo-exposure images The saturation S in the three channels of R, G, and B is calculated by the spatial linear saturation adjustment algorithm, and the saturation increment is normalized in the range of [0,100] to obtain the pseudo-exposure image after saturation adjustment. For each pseudo exposure image E' k , k∈{1,…,K} are decomposed into two scales by mean filtering Z to obtain the base sequence E' B =(B1,B2,…,B K ) and detail layer sequence E' D =(D1,D2,…,D K ); In the formula, E k represents any pseudo-exposure image of sequence E after contrast adjustment; E k (x) represents any pseudo exposure image E k The set of corresponding pixels x; Represents any pseudo exposure image E k R channel; Represents any pseudo exposure image E k G channel; Represents any pseudo exposure image E k B channel; E' k represents any pseudo-exposure image of sequence E after saturation adjustment; Represents any pseudo exposure image E' k R channel; Represents any pseudo exposure image E' k G channel; Represents any pseudo exposure image E' k B channel; B k It means grassroots; D k Represents the detail layer; S4, base sequence E' B =(B1,B2,…,B K ) The base layer fusion result F is obtained by energy saliency guided filtering algorithm B , detail layer sequence E' D =(D1,D2,…,D K ) The detail layer fusion result F is obtained by Gaussian saliency guide filtering algorithm D ; S5. Basic fusion result F B And the detail layer fusion result F D Perform two-scale reconstruction to obtain single-layer deblurred pathological imaging I F .

2. The fluorescence microscopy image enhancement method based on multi-saliency guided filtering fusion according to claim 1, characterized in that: The gamma correction algorithm in step S2 comprises the following steps: Adjust the contrast of the global image through power function transformation, that is, I(x)→α·I(x) γ ; Wherein, I(x) represents any pixel of a single-layer blurred pathology imaging; → represents a power function transformation process; α and γ represent positive constants, and when γ < 1, the brighter intensity is compressed and the darker intensity is expanded, and when γ > 1, the brighter intensity is distributed in a wider range and the darker intensity is mapped to the compressed interval, thereby obtaining a pseudo-exposure image sequence.

3. The fluorescence microscopy image enhancement method based on multi-saliency guided filtering fusion according to claim 1, characterized in that: The spatial linear saturation adjustment algorithm in step S3 comprises the following steps: S301: Each pixel of the pseudo exposure image The maximum RGB component is rgb max =max(max(R,G),B), the minimum RGB component is rgb min =min(min(R,G),B); When Δ=(rgb max -rgb min ) / 255>0, the calculation formula of the saturation S of each pixel is: Where, value = (rgb max +rgb min ) / 255, L=value / 2; S302. In the range of [0,100], the calculation formula for normalizing the saturation increment is: AND' k (x)=And k (x)+(E k (x)-L×255)×α, in, Percent represents a normal number. Represents the pseudo-exposed image after saturation adjustment.

4. The fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion according to claim 1, characterized in that: In step S3, for each pseudo exposure image E' k , k∈{1,…,K} are decomposed into two scales by the mean filter Z with a filter size of n×n to obtain the base layer B k and detail layer D k , the formula is as follows: Where n represents the size of the filter Z, (x, y) represents the pixel E' k (x) coordinates, i and j represent the filter Z relative to the center pixel E' k (x) offset, get the base layer B k Finally, the corresponding detail layer formula is obtained as follows: D k =E′ k -B k 。 5. The fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion according to claim 1, characterized in that: The base sequence E' in step S4 B =(B1,B2,…,B K ) The base layer fusion result F is obtained by energy saliency guided filtering algorithm B , including the following steps: S401, the base B k The formula for obtaining Gabor coefficients by convolution with Gabor wavelet is: in, represents Gabor coefficient; m1 represents the scale parameter, which controls the frequency scaling of the Gabor function g(x,y); n1 represents the rotation angle, which controls the direction of the Gabor function g(x,y); represents the Gabor wavelet obtained by scaling the Gabor function g(x,y) by m1 and rotating it by n1; Thus, we can get the base layer B k The formula for the Gabor energy at the pixel point (x, y) is: S402, base sequence E' B =(B1,B2,…,B K ) is compared with the Gabor energy at the corresponding pixel point (x, y), and the initial weight map P of the base sequence fusion can be determined. k ,k∈{1,…,K}The formula is: S403, combined with the guide filter theory, the refined weight map of the base sequence fusion Defined as: In the formula, G r,ψ (P,I) represents the guided filtering operation; r represents the filter size of the guide filter; ψ represents the blur level of the guide filter; P and I represent the input image and guide image of the guide filter respectively; S404, refine the weight map Normalize and get the base fusion result F by weighted average B The formula is:

6. The fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion according to claim 5, characterized in that: The Gabor wavelet in step S401 is solved by scaling the Gabor function g(x,y) in m1 scale and rotating it in n1 direction. The Gabor function is defined as the modulation of the Gaussian function, and its expression in two-dimensional space is: Among them, g represents Gabor wavelet; σ is the standard deviation of Gaussian envelope, which controls the expansion degree of Gabor function in space; w0 is the center frequency, which controls the oscillation rate of Gabor function. Gabor wavelet transform involves m1 scale scaling and n1 direction rotation of Gabor function. The specific formula is: Among them, m1 represents the scale parameter, which controls the scaling of the Gabor function in frequency, and n1 represents the rotation angle, which controls the direction of the Gabor function.

7. The fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion according to claim 5, characterized in that: The refined weight map of the base sequence fusion in step S403 The solution is obtained through the guide filter theory. The specific process is as follows: The guide filter theory assumes that the filter output O of the guide filter is the guide image I in the local window w centered at pixel k. k The linear transformation formula on is: Among them, O i represents the filtered output of the wizard filtering operation; I represents the guide image of the guide filtering operation; i represents any pixel in the guide image I; k represents a pixel in the guide image I; w k represents the local window of the guide image I centered at pixel k; a k and b k They represent the linear coefficients of the guide filter linear transformation operation respectively; w k is a window of size (2r+1)×(2r+1), where r represents the window radius and a is the linear coefficient. k and b k is the local window w k The constant in , considering that the filtering principle is affected by the multi-valued output, that is, if the central pixel k is just in the local window w centered on pixel i and with a size of (2r+1)×(2r+1) i On the other hand, different local windows w k contains pixel i, and due to the local window w k changes with pixel k, resulting in coefficient a k and b k The corresponding changes also occur. At this time, the filter output O i It may not be unique. In order to solve the impact of output multi-value, the linear coefficient a k and b k Take the average, and the estimated value formula of the wizard filter output is: In the formula, and Represents the linear coefficient a k and b k The average value in the local window w; in, Adopt G r,ψ (P, I) represents the guide filter operation, where the parameters r and ψ respectively determine the filter size and blur level of the guide filter, P and I respectively represent the input image and the guide image. At this time, the base layer B in step S401 is k As a guide image, the initial weight map P in step S402 is k , k∈{1,…,K} as the input image, the refined weight map of the base sequence fusion can be obtained by the guided filtering operation Among them, w i represents the local window centered at pixel i; w| represents the size of the local window w.

8. The fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion according to claim 1, characterized in that: The detail layer sequence E' in step S4 D =(D1,D2,…,D K ) The detail layer fusion result F is obtained by Gaussian saliency guide filtering algorithm D , including the following steps: The detail layer D k The high-pass image H is obtained by Laplace filtering k The formula is: H k =D k *L, Among them, L represents the 3×3 Laplace filter, and then the high-pass image H k The absolute value of is filtered through a Gaussian low-pass filter g to construct a Gaussian saliency map The formula is: Among them, g represents a 5×5 Gaussian low-pass filter, which is used for the detail layer sequence E' D =(D1,D2,…,D K ) is compared with the Gaussian saliency map at the corresponding pixel point (x, y) to determine the initial weight map for the fusion of the detail layer sequence The formula is: Combined with the guided filter theory, the refined weight map of detail layer sequence fusion Defined as: Refine the weight map Normalize and obtain the detail layer fusion result F through weighted average D The formula is:

9. A fluorescence microscopic image enhancement device based on multi-saliency guided filter fusion, characterized in that: include: processor; a memory having stored thereon a computer program executable on the processor; Wherein, when the computer program is executed by the processor, the steps of a fluorescence microscopy image enhancement method based on multi-saliency guided filter fusion as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a data processing program, and when the data processing program is executed by the processor, the steps of a fluorescence microscopy image enhancement method based on multi-saliency guided filtering fusion as described in any one of claims 1 to 8 are implemented.

Images