A weak signal protection type enhancement method for low signal-to-noise ratio fluorescence images

Abstract

The application discloses a weak signal protection type enhancement method for low signal-to-noise ratio fluorescent images, and relates to the technical field of fluorescent image processing. The method is as follows: an original fluorescent image is acquired and basic correction is performed, local noise intensity, a plurality of scale background candidate layers, an initial background layer and an initial background residual image are estimated, and a noise-background feature set is formed; weak fluorescent signal significance scores are calculated in combination with multi-scale response, spatial continuity and background estimation reliability; a weak signal protection weight map with continuous values is generated; background deduction intensity and denoising intensity are adaptively adjusted to obtain a preliminary enhanced image; intensity fidelity verification and feedback correction are performed to obtain a final enhanced fluorescent image. The application can reduce the risk of real weak fluorescent signals being mis-deducted or excessively smoothed while suppressing background and noise without relying on hard segmentation of the target and the background, improve output signal-to-noise ratio, output signal-to-background ratio and weak signal detectability, and maintain relative intensity relationship.

Description

Technical Field

[0001] This invention relates to the field of fluorescence image processing technology, and more specifically to a weak signal protection enhancement method for low signal-to-noise ratio fluorescence images. Background Technology

[0002] Fluorescence imaging technology utilizes the fluorescence signals generated by fluorescent labels or substances under specific excitation conditions to observe and analyze target structures, molecules, tissue regions, or biological processes. It has been widely applied in fluorescence microscopy, fluorescence endoscopy, fluorescence molecular imaging, histopathological analysis, and intraoperative fluorescence navigation. Since fluorescence signals typically reflect the spatial distribution, concentration changes, or functional state of target substances, the quality of fluorescence images directly affects the accuracy of subsequent observation, identification, localization, and quantitative analysis.

[0003] In actual fluorescence imaging, the true fluorescence signal is often weak due to factors such as low-concentration labeling, low excitation light intensity, short exposure time, rapid dynamic acquisition, deep tissue imaging, or complex sample backgrounds. This weakness is easily masked by factors such as autofluorescence, scattered light, defocused fluorescence, camera readout noise, dark current noise, shot noise, and fixed pattern noise. In such cases, the difference between the true weak fluorescence signal and random noise peaks or gradually varying background fluctuations is small, making it difficult to distinguish weak fluorescence regions and thus affecting image interpretation and quantitative analysis.

[0004] To improve fluorescence image quality, existing technologies typically employ methods such as increasing excitation light intensity, extending exposure time, improving camera gain, using high-sensitivity detectors, or improving optical systems to enhance the imaging signal. However, increasing excitation light intensity or extending exposure time may cause sample photobleaching, phototoxicity, or dynamic imaging ghosting; increasing camera gain may simultaneously amplify noise; and using high-sensitivity detectors or complex optical systems increases equipment cost and system complexity. Therefore, improving the quality of low signal-to-noise ratio fluorescence images through image processing algorithms, without significantly increasing hardware burden or the risk of sample damage, has significant application value.

[0005] Existing software enhancement methods mainly include background subtraction, flat-field correction, spatial filtering, temporal averaging, multi-frame fusion, pixel merging, histogram equalization, local contrast enhancement, edge-preserving denoising, wavelet denoising, non-local mean denoising, and deep learning-based image enhancement methods. These methods can improve image brightness, reduce noise, or enhance visual contrast to some extent, but they still have shortcomings in low signal-to-noise ratio, weak fluorescence scenes. For example, traditional background subtraction methods may mistakenly subtract real weak fluorescence signals with near-background intensities; fixed-intensity filtering or denoising methods may over-smooth small, low-brightness, and poorly defined weak fluorescence structures; and ordinary contrast enhancement methods may simultaneously amplify background noise or artifacts, and may not necessarily maintain the relative intensity relationship of the fluorescence signals.

[0006] Furthermore, some existing methods employ the idea of ​​segmenting the target region and background region, enhancing or suppressing the region of interest and background region separately. However, in low signal-to-noise ratio fluorescence images, the true weak fluorescence signal often has not yet formed a clear boundary, and its gray value may be lower than the conventional segmentation threshold. If hard segmentation of the target region and background region is performed too early, it is easy to miss weak fluorescence targets, break the target boundary, or false detection due to background noise. If the segmentation threshold is lowered to avoid missed detection, it may introduce a large number of noise peaks and background fluctuations, causing the subsequent enhancement process to amplify the false signal.

[0007] Therefore, existing technologies still require an enhancement method suitable for low signal-to-noise ratio fluorescence images that can simultaneously consider local noise levels, multi-scale background response, positive background residuals, spatial continuity, and background estimation reliability after basic correction. This method should perform continuous significance analysis on potential weak fluorescence signals and generate a continuously valued weak signal protection weight map for adaptively adjusting background subtraction and denoising intensity. Furthermore, it is necessary to perform weak signal intensity fidelity verification during the enhancement output stage to reduce the risk of incorrect subtraction, over-smoothing, or over-enhancing of the true weak fluorescence signal while improving the output signal-to-noise ratio and output signal-to-background ratio. Summary of the Invention

[0008] The purpose of this invention is to provide a weak signal-protective enhancement method for low signal-to-noise ratio fluorescence images, thereby addressing the problems mentioned in the background section. The method includes the following steps:

[0009] S1. Obtain the original fluorescence image to be enhanced, and perform basic correction on the original fluorescence image to obtain the corrected fluorescence image;

[0010] S2. Based on the corrected fluorescence image, estimate the local noise intensity map, multiple scale background candidate layers, initial background layer and initial background residual map to form a noise-background feature set;

[0011] S3. Based on the noise-background feature set, noise normalization is performed on the positive background residual in the corrected fluorescence image, and combined with multi-scale response, spatial continuity and background estimation reliability, a weak fluorescence signal saliency score map is obtained.

[0012] S4. Generate a continuously valued weak signal protection weight map based on the weak fluorescence signal saliency score map. The weak signal protection weight map is used to characterize the degree to which each pixel position should be protected in the subsequent background subtraction and denoising process.

[0013] S5. Adjust the background subtraction intensity and denoising intensity according to the weak signal protection weight map, and perform protective background subtraction and protective adaptive denoising on the corrected fluorescence image to obtain a preliminary enhanced image;

[0014] S6. Based on the weak signal protection weight map, the weak signal residual components in the preliminary enhanced image are subjected to intensity fidelity verification and feedback correction to obtain the final enhanced fluorescence image.

[0015] Furthermore, S1 specifically refers to:

[0016] Acquire the raw fluorescence image;

[0017] The original fluorescence image is dark-field corrected based on the dark-field reference information to obtain a dark-field corrected image.

[0018] The dark-field corrected image is normalized and flat-field corrected based on the flat-field reference information to obtain the flat-field corrected fluorescence image;

[0019] The grayscale of the flat-field rectified image is normalized based on the exposure time and camera gain, and the grayscale range is constrained to obtain the rectified fluorescence image. .

[0020] Furthermore, S2 specifically refers to:

[0021] The corrected fluorescence image Calculate the local mean within the local neighborhood window , Indicates the neighborhood window scale;

[0022] Estimating the local noise intensity map based on the local intensity-related noise model:

[0023] ;

[0024] in, Indicates pixel position Local noise intensity at the location, This represents the intensity coefficient associated with photon shot noise. This indicates a bias term related to readout noise, dark current noise, or fixed noise. It is a preset, extremely small positive number;

[0025] Multiple background estimation scales are set, and multi-scale background estimation is performed on the corrected fluorescence image to obtain multiple scale background candidate layers. ,in, Indicates the scale number;

[0026] For the multiple scale background candidate layers Weighted fusion is performed to obtain the initial background layer. And calculate the initial background residual map. ;

[0027] The corrected fluorescence image, local noise intensity map, initial background layer, initial background residual map, and multiple scale background candidate layers are combined to form a noise-background feature set.

[0028] Furthermore, in step S3, the positive background residual is subjected to noise normalization processing, specifically as follows:

[0029] From the initial background residual map Extracting positive background residuals :

[0030] ;

[0031] According to the local noise intensity map The positive background residual is normalized to obtain the noise-normalized residual. ;

[0032] The noise-normalized residuals are mapped to residual significance scores:

[0033] ;

[0034] in, Indicates the significance score of the residuals. This represents the reference threshold for the significance of the residuals. This represents the slope of the mapping function.

[0035] Furthermore, in S3, a weak fluorescence signal significance score map is obtained by combining multi-scale response, spatial continuity, and background estimation reliability, specifically as follows:

[0036] Based on multiple background candidate layers at different scales, the positive background residual response and noise normalized response at the corresponding scales are calculated respectively.

[0037] Calculate the multi-scale stability significance score based on the noise normalized response at multiple scales. ;

[0038] The spatial continuity score is obtained by statistically analyzing the soft saliency response of neighboring pixels within the local support window. ;

[0039] The background estimation reliability score is calculated based on the differences between the candidate background layers at multiple scales and the initial background layer. ;

[0040] Based on the residual significance score Multiscale stability significance score Spatial continuity score and background estimation reliability score The fusion yielded a significance score map for weak fluorescence signals:

[0041] When the original fluorescence image is a multi-frame fluorescence image sequence and time-series consistency analysis is required:

[0042] Obtain the noise normalized residuals for each frame. Based on the current processing frame, and using the noise normalization residuals corresponding to each frame as a benchmark, the temporal consistency score of the current processing frame is calculated. ,in, Indicates the frame number;

[0043] ;

[0044] , , , These represent the fusion weights for residual significance, multi-scale stability, spatial continuity, and temporal consistency, respectively.

[0045] When timing consistency analysis is not used, let:

[0046] ;

[0047] at this time:

[0048] .

[0049] Furthermore, S4 specifically includes:

[0050] According to the significance score chart of the weak fluorescence signal The initial protection response diagram is generated through a continuous mapping function:

[0051] ;

[0052] in, Indicates the initial protection response. This represents the reference threshold for weak signal protection response. This represents the slope of the mapping function;

[0053] Reliability score based on background estimation and residual significance score The initial protection response diagram is conservatively modified to obtain the protection response after background reliability correction:

[0054] ;

[0055] in, This indicates the protection response after background reliability correction. Indicates the background uncertainty compensation coefficient;

[0056] Based on multi-scale stability significance score Spatial continuity score Constructing weak signal structure support factors:

[0057] ;

[0058] Based on the weak signal structural support factor, the protection response is modified to obtain the structural support-corrected protection response. :

[0059] ;

[0060] in, This indicates the protection response after structural support modification. To protect the lower limit coefficient;

[0061] Edge-preserving spatial smoothing is applied to the protection response after structural support modification, and the protection responses before and after smoothing are fused to obtain a continuously valued weak signal protection weight map. .

[0062] Furthermore, in step S5, the background subtraction intensity is adjusted according to the weak signal protection weight map, specifically as follows:

[0063] According to the weak signal protection weight diagram and background estimation reliability score Generate a background subtraction factor map:

[0064]

[0065] in, This is a background subtraction factor map, representing pixel positions. Background subtraction factor value at the location, This represents the baseline value for the background deduction factor. This represents the maximum background subtraction factor. This represents the minimum retention factor when the background estimate is unreliable;

[0066] According to the background deduction coefficient diagram and the initial background layer For the corrected fluorescence image Perform protective background subtraction to obtain the fluorescence image after background subtraction:

[0067] .

[0068] Furthermore, in step S5, the noise reduction intensity is adjusted according to the weak signal protection weight map, specifically as follows:

[0069] Based on the local noise intensity map Positive background residual Weak signal protection weight diagram Generate a noise reduction intensity coefficient map:

[0070] ;

[0071] in, This is a noise reduction intensity coefficient map, representing pixel locations. The noise reduction intensity coefficient value at the location, Indicates the basic denoising intensity;

[0072] Fluorescence image after background subtraction Perform edge-preserving denoising to obtain the basic denoised image. ;

[0073] According to the noise reduction intensity coefficient diagram Fluorescent image after background subtraction Compared with the base denoised image The images are fused to obtain a protective adaptive denoised image:

[0074] .

[0075] Furthermore, the preliminary enhanced image is obtained in S5, specifically as follows:

[0076] According to the protective adaptive denoising image Weak signal protection weight diagram Significance score of weak fluorescence signal Background estimation reliability score and positive background residual Weak signal residual compensation is performed to obtain a preliminary enhanced image:

[0077] ;

[0078] in, This indicates the initial image enhancement. This represents the weak signal compensation coefficient.

[0079] Furthermore, in S6, intensity fidelity verification and feedback correction are performed on the initially enhanced image, specifically as follows:

[0080] According to the weak signal protection weight diagram Significance score of weak fluorescence signal Construct weak signal fidelity verification weights:

[0081] ;

[0082] With positive background residual As a reference residual for the weak signal before enhancement, and based on the preliminary enhanced image and the background components that may be retained after background subtraction, the estimated weak signal residual after enhancement is determined:

[0083] ;

[0084] Within the local verification window, based on the weak signal fidelity verification weights at each pixel position within the window, the weak signal reference residual before enhancement and the weak signal residual estimate after enhancement are weighted integrals to obtain the weighted weak signal integral strength before enhancement and the weighted weak signal integral strength after enhancement.

[0085] Calculate the weak signal intensity preservation ratio based on the weighted weak signal integral intensity before and after enhancement. ;

[0086] A feedback correction coefficient is constructed based on the weak signal strength preservation ratio. :

[0087] ;

[0088] in, This indicates that the expected intensity remains proportional. and These represent the lower and upper limits of the feedback correction coefficient, respectively.

[0089] The weak signal residual components in the initial enhanced image are corrected using the feedback correction coefficient to obtain the enhanced fluorescence image after intensity fidelity correction. :

[0090] ;

[0091] Enhanced fluorescence image after intensity fidelity correction By constraining the grayscale range, the final enhanced fluorescence image is obtained.

[0092] Compared with the prior art, the present invention has the following improvements and advantages:

[0093] 1. This invention eliminates or reduces the influence of non-sample factors such as camera bias, dark current, field of view unevenness, exposure differences, and abnormal gray values ​​through basic correction, avoiding mistaking camera fixation noise, uneven illumination, or gray value shifts for true weak fluorescence signals, and providing a stable, unified, and comparable image basis for subsequent weak fluorescence signal analysis.

[0094] 2. This invention, without hard segmentation of the target and background regions, simultaneously considers local noise levels, multi-scale background response, positive background residuals, spatial continuity, and background estimation reliability to perform continuous saliency analysis on potential weak fluorescence signals. This enables better differentiation of true weak fluorescence candidate signals, random noise peaks, and background fluctuations, reducing the risks of missed detection of weak fluorescence targets, false detection of background noise, and underestimation or overestimation of the background caused by a single-scale background model. Furthermore, the initial background layer obtained is further adjusted in subsequent steps using weak signal protection weights to adjust the background subtraction intensity, thereby reducing the risk of false subtraction of true weak fluorescence signals. The saliency analysis results also provide a more reliable basis for the subsequent generation of weak signal protection weight maps, allowing subsequent background subtraction and denoising processes to adaptively adjust based on the saliency of the weak fluorescence signal.

[0095] 3. This invention generates a continuously valued weak signal protection weight map and adaptively adjusts the background subtraction and denoising intensities based on this map. This provides strong protection for weak fluorescence candidate regions and strong suppression for background regions, thereby reducing the risk of false subtraction or over-smoothing of true weak fluorescence signals while suppressing background and noise. It also improves the output signal-to-noise ratio, output signal-to-background ratio, and weak signal detectability of low signal-to-noise ratio fluorescence images. Specifically, regions with high saliency, stable multi-scale response, and spatial continuity are assigned higher protection weights; isolated noise peaks, background regions, or regions with low saliency are assigned lower protection weights; and for regions with uncertain background estimation but positive background residuals, a conservative protection strategy is adopted to reduce the risk of subsequent excessive background subtraction. Edge-preserving spatial smoothing improves the spatial stability of the weak signal protection weight map while preventing excessive diffusion of protection weights across obvious fluorescence structure boundaries. In multi-frame scenarios, optional temporal consistency correction can further improve the temporal stability of the weak signal protection weights and reduce the impact of random noise peaks on single-frame results. The resulting weak signal protection weight map provides a control basis for subsequent protective background subtraction and adaptive denoising, enabling subsequent processing to adaptively adjust the background subtraction intensity and denoising intensity according to the credibility of weak signals at different locations.

[0096] 4. This invention further performs intensity fidelity verification and feedback correction based on weak signal residual components, which can reduce the interference of background components on fidelity judgment. While improving the output signal-to-noise ratio and output signal-to-background ratio, it better maintains the relative intensity relationship of the real weak fluorescence signal, reduces the risk of over-enhancement, and thus facilitates subsequent observation, identification, localization and quantitative analysis. Attached Figure Description

[0097] Figure 1 This is a schematic diagram illustrating an application scenario of the present invention;

[0098] Figure 2 This is a flowchart of the algorithm of the present invention;

[0099] Figure 3 Comparison chart showing the effects of different enhancement methods;

[0100] Figure 4 Comparison of ablation experiments for key modules;

[0101] Figure 5 A schematic diagram illustrating the consistency between the weight map for weak signal protection and the actual location of the weak signal. Detailed Implementation

[0102] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0103] like Figure 1 As shown, the object to be processed is a low signal-to-noise ratio fluorescence image acquired by an imaging system. During fluorescence imaging, excitation light irradiates the sample, generating a fluorescence emission signal, which is then acquired by a camera or detector. Due to the low intensity of weak fluorescence signals, the acquired images typically contain interference factors such as scattered light, defocused fluorescence, background autofluorescence, and imaging noise, causing the true weak fluorescence signal to be easily submerged by the background or masked by noise, affecting subsequent analysis and identification. To address these issues, this invention first performs basic correction on the input image and estimates local noise information and multi-scale background information. Then, combining local noise information, multi-scale background response, spatial continuity, and background estimation reliability, a saliency analysis of the weak fluorescence signal is performed to generate a weak signal protection weight map, which characterizes the degree of protection of weak fluorescence candidate regions corresponding to each pixel. Finally, using the weak signal protection weight map as a constraint, the background subtraction intensity and adaptive denoising intensity are synergistically adjusted to reduce background noise while preserving the true weak fluorescence signal as much as possible. The enhanced fluorescence image is then output through intensity fidelity verification, achieving a simultaneous improvement in fluorescence image sensitivity and output signal-to-noise ratio. The weak signal protection weight map is not the result of target segmentation, but a continuous weight mapping used to adjust the background subtraction intensity and adaptive denoising intensity, which can effectively reduce the risk of real weak fluorescence signals being mistakenly subtracted or over-smoothed during the enhancement process.

[0104] like Figure 2 As shown, this invention proposes a weak signal protection enhancement method for low signal-to-noise ratio fluorescence images, comprising the following steps:

[0105] S1: Fluorescence image acquisition and basic correction:

[0106] This step is used to acquire the low signal-to-noise ratio fluorescence image to be enhanced, and to perform basic corrections on camera bias, dark current, field of view inhomogeneity, exposure differences, and grayscale differences in the original image to obtain the corrected fluorescence image required for subsequent processing.

[0107] This step reduces the impact of non-sample factors introduced by the imaging equipment itself on subsequent local noise estimation, background modeling, and weak fluorescence signal significance analysis, avoiding misinterpreting camera fixation noise, uneven illumination, or grayscale shifts as true weak fluorescence signals.

[0108] S1.1: Obtain the raw fluorescence image:

[0109] Obtain the original fluorescence image to be enhanced, denoted as:

[0110] ;

[0111] in, Indicates the pixel position in the image. This indicates the original grayscale value or fluorescence intensity value at that pixel location.

[0112] The original fluorescence image can be a single frame or a sequence of multiple continuously acquired fluorescence images. When the input is a sequence of multiple frames, it can be denoted as:

[0113] ;

[0114] in, Indicates the frame number. This indicates the number of fluorescence image frames that were processed.

[0115] The fluorescence images can be obtained from fluorescence microscopes, fluorescence endoscopes, fluorescence molecular imaging devices, fluorescence detectors, fluorescence navigation devices, or other imaging systems with fluorescence image acquisition capabilities.

[0116] S1.2: Obtain the basic input parameters corresponding to the original fluorescence image:

[0117] To complete the basic correction of the original fluorescence image and provide a unified data foundation for subsequent local noise estimation, background layer modeling, and weak fluorescence signal significance analysis, this step obtains the basic input parameters corresponding to the original fluorescence image.

[0118] The basic input parameters include at least one of the following: original fluorescence image grayscale value, image bit depth or effective grayscale range, dark field reference information, flat field reference information, exposure time, and camera gain. Specifically, the basic input parameters may include the following:

[0119] First, the grayscale values ​​of the original fluorescence image. The original fluorescence image is denoted as... .in, Indicates the pixel position in the image. This represents the original grayscale value or fluorescence intensity value at the pixel location. The original fluorescence image is used for subsequent dark-field correction, flat-field correction, local noise estimation, background layer estimation, and weak fluorescence signal saliency analysis. When the input is a multi-frame fluorescence image sequence, the original fluorescence image can be denoted as... .

[0120] Second, the image bit depth or effective grayscale range. Image bit depth determines the effective range of grayscale values ​​in an image, such as 8-bit, 12-bit, 14-bit, or 16-bit images. The maximum effective grayscale value of the current image can be determined based on the image bit depth. This is used for subsequent grayscale range constraints, saturation grayscale judgment, and abnormal grayscale value processing. When the image bit depth cannot be directly obtained, the effective grayscale range can also be determined based on the image data format, the maximum grayscale value of the image, or the preset range of the imaging system.

[0121] Third, dark field reference information. Dark field reference information is used to subtract camera bias, dark current, and fixed readout offset. If a pre-acquired dark field image exists, it is recorded as... ,in, Indicates pixel position The dark field grayscale value at the location. If there is no pre-acquired dark field image, the dark field bias can also be estimated based on the non-fluorescent area, the low quantile grayscale value of the image, the background area at the edge of the image, or the historical dark field statistics, and the estimation result can be used as dark field reference information.

[0122] Fourth, flattening reference information. Flattening reference information is used to correct for uneven illumination, uneven lens response, and differences in sensor pixel response. If a pre-acquired flattening image exists, it is recorded as... ,in, Indicates pixel position The flat field response value at the location. If there is no pre-acquired flat field image, an approximate field response distribution can be obtained by large-scale low-pass filtering, statistical analysis of blank areas, or morphological background estimation, and used as flat field reference information.

[0123] Fifth, exposure time. Exposure time is recorded as... This method is used to normalize the grayscale of fluorescence images acquired under different exposure conditions. When multiple frames or batches of images have different exposure times, or when it is necessary to unify fluorescence images acquired under different conditions to the same grayscale, exposure time can be used to normalize the image grayscale. If the exposure time remains constant during the current image processing, or if exposure normalization is not required, then... The exposure time term can be taken as a fixed constant or omitted in the subsequent normalization step.

[0124] Sixth, camera gain. Camera gain is denoted as... It is used to normalize the grayscale of fluorescence images acquired under different gain conditions. When the camera gain changes, the image grayscale values ​​can be corrected according to the gain parameter, so that subsequent local noise estimation, background modeling, and weak signal saliency analysis are on a consistent scale. If the camera gain remains constant during the current image processing, or if gain normalization is not required, then... The gain term can be set as a fixed constant or omitted in subsequent normalization steps.

[0125] S1.3: Dark Field and Bias Correction:

[0126] To eliminate the effects of camera bias, dark current, and fixed readout offset, dark field correction is performed on the original fluorescence image.

[0127] If a dark field image exists The dark-field corrected image can then be obtained according to the following formula:

[0128] ;

[0129] in, This represents the image after dark field correction.

[0130] When the input is a multi-frame fluorescence image, dark-field correction is performed on each frame separately:

[0131] ;

[0132] If there is no pre-acquired dark field image, the dark field bias value can be estimated based on the non-fluorescent area, the low quantile gray value of the image, the background area at the edge of the image, or the historical dark field statistics, and then subtracted.

[0133] To avoid invalid negative values ​​after deduction, non-negativity truncation can be performed further:

[0134] ;

[0135] S1.4: Flat field correction:

[0136] Because fluorescence imaging systems may suffer from uneven illumination, lens aberrations, and inconsistent sensor responses, the same fluorescence intensity may exhibit different grayscale values ​​at different field-of-view locations. Therefore, this step further performs flat-field correction on the image after dark-field correction.

[0137] If a flat field image exists First, normalize the flat field image as follows:

[0138] ;

[0139] in, This represents the normalized flat-field reference image. This represents the average grayscale value of a flat field image.

[0140] Then perform flat field correction according to the following formula:

[0141] ;

[0142] in, This represents the fluorescence image after flat-field correction. This is a preset, extremely small positive number, used as a stabilizing term in division or square root operations to prevent the denominator from being zero or non-positive values ​​from appearing within the square root.

[0143] If a flat-field image is not available, large-scale low-pass filtering, morphological background estimation, or blank area statistics can be used to estimate the field response distribution and apply it to approximate flat-field correction.

[0144] S1.5: Exposure and gain normalization:

[0145] When there are differences in exposure time or camera gain between input images, exposure and gain normalization can be performed on the flat-field corrected image to ensure that the subsequent local noise estimation, background modeling and weak signal saliency analysis have a consistent gray scale.

[0146] Let the exposure time be Camera gain is Then the normalized image can be represented as:

[0147] ;

[0148] in, This represents the fluorescence image after exposure and gain normalization. It is a preset minimal positive number.

[0149] When the exposure time and camera gain remain constant during the current processing, or when there is no need to unify the grayscale of different images, it is possible to let , Or directly:

[0150] ;

[0151] This step can reduce the impact of different exposure times and different gain settings on the image grayscale, making the subsequent weak fluorescence signal saliency analysis more stable.

[0152] S1.6: Gray-scale range constraint and correction image output:

[0153] The normalized image is constrained by grayscale range, and abnormal saturation values ​​and invalid values ​​are removed to obtain the corrected fluorescence image:

[0154] ;

[0155] In one implementation, grayscale range constraint can be performed in the following manner:

[0156] ;

[0157] in, This represents the maximum grayscale value corresponding to the current image bit depth, or the effective upper limit obtained based on the actual imaging system settings.

[0158] In the case of multi-frame input, the corrected fluorescence image sequence is obtained:

[0159] ;

[0160] The corrected fluorescence image Or corrected fluorescence image sequence It serves as input for subsequent local noise estimation, background layer modeling, and weak fluorescence signal significance analysis.

[0161] After processing with S1, the corrected fluorescence image is obtained:

[0162] ;

[0163] When the input is a sequence of multiple image frames, the following is obtained:

[0164] ;

[0165] Optionally, dark field correction results, flat field correction results, saturated pixel or invalid pixel masks, and basic grayscale statistics of the current image can also be obtained.

[0166] S2: Local noise and background feature estimation:

[0167] This step is used to estimate the local noise intensity and slowly varying background distribution in the corrected fluorescence image obtained based on S1, and to obtain the local noise intensity map and the initial background layer.

[0168] This step provides a noise and background benchmark for subsequent weak fluorescence signal significance analysis, enabling subsequent steps to determine whether a low-brightness residual signal is a random noise fluctuation, background undulation, or a statistically significant true weak fluorescence signal.

[0169] S2.1: Obtain the corrected fluorescence image:

[0170] Obtain the corrected fluorescence image output by S1, denoted as:

[0171] ;

[0172] in, Indicates the pixel position in the image. This represents the fluorescence intensity value at the pixel location after dark field correction, flat field correction, exposure gain normalization, and grayscale range constraint.

[0173] When the input is a multi-frame fluorescence image sequence, obtain the corrected fluorescence image sequence:

[0174] ;

[0175] in, Indicates the frame number. This indicates the number of fluorescence image frames that were processed.

[0176] In single-frame processing scenarios, subsequent steps are directly based on Perform local noise and background estimation; in multi-frame processing scenarios, it can perform local noise and background estimation for each frame. Performing this step separately, or updating the noise and background estimation results by combining multi-frame statistical results in subsequent implementations.

[0177] S2.2: Calculate local gray-level statistical characteristics:

[0178] With the corrected fluorescence image Based on this, at each pixel location Calculate local gray-level statistical features within the neighborhood window.

[0179] Let's assume The local neighborhood window centered on is:

[0180] ;

[0181] in, This indicates the scale of the neighboring window. The window can be a square window, a circular window, or a local area of ​​other preset shapes.

[0182] Within the neighborhood window, calculate the local mean:

[0183] ;

[0184] in, Indicates pixel position The local average fluorescence intensity at that location This indicates the number of pixels in the neighborhood window.

[0185] In a preferred embodiment, to reduce the impact of obvious bright spots, abnormal noise, or strong fluorescent areas on local statistical results, the pixels with the highest gray values ​​in the neighborhood window can be removed when calculating the local mean, or robust statistical methods such as local median or truncated mean can be used instead of ordinary mean.

[0186] S2.3: Estimated local noise intensity map:

[0187] Noise in fluorescence images is typically related to local fluorescence intensity. Under low-light conditions, image noise is mainly affected by photon shot noise, camera readout noise, dark current noise, and background noise. Therefore, this step uses a local intensity-dependent noise model to estimate the noise level at different locations in the image.

[0188] In one implementation, the local noise variance of the corrected fluorescence image is expressed as:

[0189] ;

[0190] in, Indicates pixel position The local noise variance at that location Indicates the standard deviation of local noise. This indicates the local average fluorescence intensity at that location. This represents the intensity coefficient associated with photon shot noise. This represents the bias term associated with readout noise, dark current noise, and other approximately fixed noise.

[0191] Further, a local noise intensity map was obtained:

[0192] ;

[0193] in, Indicates pixel position Local noise intensity at the location, It is a preset minimal positive number.

[0194] Optionally, local variance can also be calculated. ;

[0195] in, Indicates pixel position The degree of local grayscale fluctuation at a given location.

[0196] parameter and It can be determined in the following ways:

[0197] First, it can be pre-defined based on camera calibration results, historical acquisition data, or imaging system parameters;

[0198] Second, background candidate regions with low local gradients and gradual grayscale changes can be selected in the current image, based on the local mean. and local variance Fit the relationship between them;

[0199] Third, when calibration data or reliable background candidate regions cannot be obtained, preset empirical values ​​or estimated values ​​based on overall image statistical results can be used.

[0200] In a preferred embodiment, the background candidate region is not obtained through hard segmentation of the target region, but is selected based on conditions such as low local gradient, stable local variance, and grayscale values ​​not reaching a significantly strong fluorescence range. This avoids prematurely simplifying the subsequent weak signal protection problem into a simple target region and background region segmentation problem.

[0201] S2.4: Estimating the multi-scale initial background layer:

[0202] In fluorescence images, the background typically consists of slowly changing low-frequency components, such as tissue autofluorescence, scattered light, defocused fluorescence, residual illumination unevenness, or sample substrate background. To obtain the background baseline required for subsequent weak signal significance analysis, this step performs multi-scale background estimation on the corrected fluorescence image to obtain the initial background layer.

[0203] Let the set of background estimation scales be:

[0204] ;

[0205] in, Indicates the first A background estimation scale, This indicates the number of scales. Different scales are used to accommodate background components of different sizes and rates of change. Specifically, the background estimation scale is determined based on the range of background variation in the fluorescence image to be processed, the typical size of weakly fluorescent targets, and the background smoothing requirements, to ensure that different scales can cover the changing features of both local and global backgrounds.

[0206] For each scale By employing low-pass filtering, morphological opening, rolling ball background estimation, median filtering, or other methods capable of extracting slowly varying backgrounds, a background candidate layer at the corresponding scale is obtained:

[0207] ;

[0208] in, Representing scale The background candidate layer below, The scale is represented as Background estimation operator.

[0209] Then, the background candidate layers at multiple scales are fused to obtain the initial background layer:

[0210] ;

[0211] in, Indicates pixel position Initial background estimate at the location, Representing scale The background blending weights corresponding to the pixel position, These are preset minimal positive numbers used for numerical stability protection.

[0212] In one implementation, background fusion weights The background candidate layer can be determined based on local smoothness, local gradient intensity, local noise intensity, or consistency between background candidate layers at different scales. For regions with slow background changes and weak local structures, a higher weight can be assigned to the background candidate layer; for regions with significant local gray-level changes or potential weak fluorescence structures, the dependence of the background estimation in that region on the local brightness abrupt change in the original image can be reduced to avoid the initial background layer from over-fitting the real fluorescence signal.

[0213] It should be noted that the result obtained in this step This layer serves only as the initial background layer and is not directly used for image enhancement or final background subtraction in this step. Subsequent steps will further adaptively adjust the background subtraction intensity by combining the saliency of weak fluorescence signals and the weak signal protection weights.

[0214] S2.5: Calculate the initial background residual map:

[0215] Based on the corrected fluorescence image and the initial background layer Calculate the initial background residual map:

[0216] ;

[0217] in, Indicates pixel position The residual signal relative to the initial background layer.

[0218] when A larger value indicates that the grayscale value at that location is higher than the estimated background, which may correspond to the true fluorescence signal, noise peaks, or background estimation bias; when... When the value is close to zero or negative, it indicates that the location is close to or below the background layer estimate.

[0219] It should be noted that this step does not directly... The region exceeding a certain threshold is identified as the target region, but it is not directly defined as such. Smaller areas are identified as background areas. The initial background residual map is only used as one of the input features for subsequent weak fluorescence signal saliency analysis.

[0220] S2.6: Forming a noise-background feature set:

[0221] The local noise intensity map, initial background layer, initial background residual map, and multiple scale background candidate layers obtained in the above steps, along with the corrected fluorescence image obtained in S1, are used in combination. Combine them to form a noise-background feature set:

[0222] ;

[0223] in, Indicates pixel position The noise-background feature set at the location.

[0224] This feature set is used for the subsequent saliency analysis of weak fluorescence signals in S3. Based on the aforementioned local statistical features participating in noise estimation and parameter determination, by simultaneously considering local noise intensity, background estimate, background residual, and multiple scale background candidate layers, we can avoid relying solely on grayscale thresholds to judge weak fluorescence signals, thereby improving the stability of weak signal recognition in low signal-to-noise ratio fluorescence images.

[0225] Noise-background feature set As input for subsequent weak fluorescence signal significance analysis, multiple background candidate layers at various scales are included. As input for S3 multiscale response and background reliability calculations.

[0226] S3: Significance analysis of weak fluorescence signals:

[0227] This step is used to perform significance analysis on the low-brightness residual signal in the corrected fluorescence image based on the local noise intensity map, initial background layer, and initial background residual map obtained from S2, to determine whether it has statistical significance relative to local noise fluctuations and background undulations, and to obtain a weak fluorescence signal significance score map.

[0228] This step avoids relying solely on global grayscale thresholds or simple background residuals to identify weak fluorescent targets, thereby reducing the risk of misjudging random noise peaks as fluorescent signals or misjudging true weak fluorescent signals as background noise.

[0229] S3.1: Obtain noise-background feature input:

[0230] Obtain the noise-background feature set output by S2. The noise-background feature set includes the corrected fluorescence image, local noise intensity map, initial background layer, initial background residual map, and multiple scale background candidate layers, denoted as:

[0231] ;

[0232] in, This represents the corrected fluorescence image. This represents a local noise intensity map. Indicates the initial background layer. Represents the initial background residual map. This represents candidate background layers at multiple scales.

[0233] The initial background residual map is obtained by the following formula:

[0234] ;

[0235] in, Background candidate layers at multiple scales in S2 It was obtained through fusion.

[0236] It should be noted that the dark field reference information in S1 Primarily used to eliminate camera bias, dark current, and fix readout offset; while the background candidate layer in S2 It is mainly used to estimate sample background, autofluorescence, scattered light, defocused fluorescence, and residual slowly varying background in the corrected fluorescence image. The two processes different objects; even after dark-field correction, it is still possible to estimate background candidate layers at multiple scales based on the corrected fluorescence image.

[0237] S3.2: Calculate the positive background residual:

[0238] Since fluorescence signals typically exhibit a positive grayscale increment relative to the background, this step extracts the positive background residual signal from the initial background residual map:

[0239] ;

[0240] in, Indicates pixel position The positive background residual is higher than the initial background layer.

[0241] when A larger value indicates that the grayscale value at that location is higher than the initial background estimate, which may correspond to the true fluorescence signal, local noise peaks, or background estimation errors; when... When the value is close to zero, it indicates that the location does not exhibit a significant positive background residual.

[0242] It should be noted that this step does not directly identify pixels with large positive background residuals as fluorescent targets, but rather uses them as the basis for subsequent noise normalization significance analysis.

[0243] S3.3: Calculate the significance of noise-normalized residuals:

[0244] To determine whether the positive background residual signal exceeds the current local noise fluctuation range, this step normalizes the positive background residual using the local noise intensity map, obtaining the noise-normalized residual:

[0245] ;

[0246] in, Indicates pixel position The noise normalized residual at that point It is a preset minimal positive number.

[0247] when When the value is small, it indicates that the positive background residual at that location may be within the range of local noise fluctuations; when... A larger value indicates that the positive background residual at that location is more significant relative to local noise and is more likely to contain a true fluorescence signal.

[0248] Furthermore, the noise-normalized residuals can be mapped to residual significance scores:

[0249] ;

[0250] in, This represents the significance score of the residuals, with a value ranging from 0 to 1; This represents the reference threshold for the significance of residuals, typically, The methods for determining the value include, but are not limited to: the truncated mean (the average value after removing a certain percentage of the highest and lowest values) based on the noise-normalized residual data of the whole map or a local area, the fixed high quantile (such as the 90th or 95th quantile), or the mean plus k times the standard deviation (k is usually 2 or 3), etc. In this specific embodiment, the truncated mean is preferred to enhance the ability to resist interference from extreme outliers. This represents the slope of the mapping function. It controls the sharpness of the transition from insignificant to significant; the larger the value, the closer the decision is to the hard threshold, and vice versa, the stronger the ability to preserve weak signals and the smoother the noise fluctuations. In practical implementation, A fixed empirical value method can be adopted, directly setting the typical value range as follows: The fixed value, preferably, can be set to or In a preferred embodiment, Alternatively, an adaptive value selection method based on global data distribution can be adopted, letting... ,in For all of the map standard deviation This is the proportionality coefficient, and its typical range of values ​​is... This allows the slope to automatically decrease when the overall noise of the current image fluctuates greatly, and to automatically increase when the noise is low, thereby adaptively adjusting the sensitivity of the score.

[0251] when Higher than hour, Increase; when Below hour, This method avoids the use of hard thresholds to directly determine weak fluorescent targets, thus making the analysis of weak signal significance continuous.

[0252] S3.4: Calculate the response of a multi-scale weak signal:

[0253] This step is used to determine whether the weak fluorescence residual signal has a stable response under different background estimation scales. Among them, Used to describe relative to the initial background layer after fusion. The significance of the overall residuals; while in this step Used to describe relative to the first Background candidate layer at each scale The scale response. By comparing responses at multiple scales, the risk of misjudgment caused by a single background estimation scale can be reduced.

[0254] True weak fluorescence signals typically exhibit a continuous distribution or structural response within a certain spatial range, while random noise peaks often appear as isolated and scale-unstable local bright spots. Therefore, this step further calculates the multi-scale weak signal response to determine whether the positive background residual signal exhibits stable behavior across multiple scales.

[0255] Let the multi-scale background candidate layer in S2 be:

[0256] ;

[0257] in, Indicates the scale number. Indicates the number of scales.

[0258] Based on the background candidate layer at each scale, calculate the positive background residual at the corresponding scale:

[0259] ;

[0260] Further calculations are performed on the noise normalized response at the corresponding scale:

[0261] ;

[0262] in, Representing scale The noise-normalized weak signal response.

[0263] Map the response at each scale to a scale significance score:

[0264] ;

[0265] in, Representing scale The significance score of the weak signal is as follows. This represents the reference threshold at this scale. This represents the corresponding mapping slope.

[0266] Weigh and fuse the saliency scores of weak signals at multiple scales to obtain a multi-scale weak signal response score:

[0267] ;

[0268] in, This represents the multi-scale weak signal response score. Indicates the first Fusion weights for each scale.

[0269] In one implementation, The stability of the fluorescence target can be pre-set based on the expected size range, imaging resolution, or background estimation at different scales. For smaller point-like fluorescence signals, the weight of small-scale responses can be increased; for larger diffuse fluorescence regions, the weight of medium- and large-scale responses can be increased.

[0270] S3.5: Calculate the multi-scale consistency score:

[0271] To further distinguish between true weak fluorescence signals and random noise peaks, this step calculates a multi-scale consistency score. True weak fluorescence signals typically do not increase abruptly at a single scale, but rather exhibit a relatively consistent positive response across adjacent or multiple scales; random noise peaks, on the other hand, are often only significant at local small scales, with the response rapidly diminishing at larger scales.

[0272] Noise normalization response at multiple scales Calculate its mean:

[0273] ;

[0274] And calculate its scale fluctuations:

[0275] ;

[0276] Further, a multi-scale consistency score was obtained:

[0277] ;

[0278] in, This represents the multi-scale consistency score. The larger the value, the more stable the response at that location across multiple scales; the smaller the value, the more likely the response at that location is to originate from isolated noise peaks or unstable background fluctuations.

[0279] In one implementation, the multi-scale weak signal response score can be combined with the multi-scale consistency score to obtain a multi-scale stable significance score:

[0280] ;

[0281] in, This represents a weak signal score that considers both multi-scale response intensity and multi-scale response stability.

[0282] S3.6: Computational space continuity score:

[0283] In addition to multi-scale stability, true weak fluorescence signals usually also exhibit a certain degree of spatial continuity or local support relationships, while random noise often manifests as isolated pixels or scattered bright spots. Therefore, this step further calculates the spatial continuity score.

[0284] Let the pixel position The central local support window is:

[0285] ;

[0286] in, This indicates the neighborhood scale used to determine spatial continuity.

[0287] Within this local support window, the soft saliency response of neighboring pixels is first calculated:

[0288] ;

[0289] Then, the soft saliency responses within the local support window are averaged to obtain the spatial continuity score:

[0290] ;

[0291] in, Indicates pixel position Spatial continuity score, Represents neighboring pixels The soft significant response, Indicates the neighborhood support reference threshold. Represents the slope of the mapping function. This indicates the number of pixels within the local support window.

[0292] Specifically, It should be slightly lower than or equal to the residual significance reference threshold in S3.3. This allows for the inclusion of neighborhood support in a lenient manner, avoiding the omission of edge pixels in weak signal clusters. In this specific embodiment, The methods for determining the value of include, but are not limited to: the fixed proportion method, that is, letting ,in It is a positive number less than or equal to 1, preferably Optimal selection Alternatively, a robust statistical method based on the distribution of local neighborhood data can be used, that is, for each local support window, the data within the window is taken. The median value is used as the adaptive value for that pixel. This enhances the ability to adapt to non-uniform noise. A fixed ratio is used by default. If the background noise distribution of the image is extremely uneven, then switch to the local median adaptive method. Control the sharpness of the transition from no support to support provided by neighboring pixels. Since the neighboring support determination should be lenient, Typically, a relatively small value is chosen. In practical implementation, A fixed empirical value method can be used, directly setting it as a constant, preferably. To achieve a balance between the smoothness and distinguishability of the support fusion, an adaptive value selection method based on the distribution of local neighborhood data can also be used to calculate the value within the window for each local support window. Standard deviation of values ,make ,in The proportionality coefficient (preferably) This allows the slope to be dynamically adjusted based on local data fluctuations. In this implementation, a fixed slope is used by default. If the baseline scheme is found to have a low spatial continuity score leading to missed detection in the edge region of weak signal clusters, the scheme will switch to a local adaptive scheme.

[0293] When a relatively stable noise-normalized residual response exists in a certain pixel and its neighborhood. High; when a pixel experiences a sudden increase in brightness only on its own while its neighboring area lacks support. The impact of isolated noise points on the significance analysis of weak fluorescence signals can be reduced in this way.

[0294] S3.7: Calculate the reliability score of the background estimation:

[0295] Initial background layer It is obtained by fusing background candidate layers at multiple scales. If the background candidate layers at different scales differ significantly, it indicates uncertainty in the background estimation at that location. In this case, judging the weak fluorescence signal solely based on the background residual may not be reliable enough. Therefore, this step further calculates the background estimation reliability score.

[0296] Background candidate layers at multiple scales and the initial background layer Calculate the uncertainty of the background estimation:

[0297] ;

[0298] Further, the normalized background uncertainty is obtained:

[0299] ;

[0300] And calculate the background estimation reliability score:

[0301] ;

[0302] in, This indicates the reliability score for the background estimation. The larger the value, the more consistent the background estimation results across multiple scales, and the higher the reliability of the background residual signal. The smaller the value, the greater the uncertainty in the background estimation at that location. Therefore, when generating weak signal protection weights, we should avoid over-reliance on the results of a single residual judgment.

[0303] It should be noted that the background estimation reliability score is not used to directly exclude weak fluorescence signals, but rather to adjust the confidence level of the significance analysis. For regions with uncertain backgrounds, subsequent steps can employ more conservative background subtraction and denoising strategies to avoid the erroneous deletion of weak fluorescence signals due to background estimation bias.

[0304] S3.8: Optional timing consistency analysis:

[0305] When the input is a sequence of consecutive fluorescence images, the reliability of weak fluorescence signals can be further determined by utilizing the temporal consistency between adjacent frames. Real fluorescence signals typically exhibit a certain degree of continuity over a short period, while random noise peaks tend to be unstable between adjacent frames.

[0306] For the From the frame image, the corresponding noise-normalized residual can be obtained:

[0307] ;

[0308] If there is motion or system jitter between adjacent frames, the adjacent frames can be mapped to the coordinate system of the current frame based on the image registration results. If the adjacent frames are approximately aligned, temporal statistics can be performed directly at the same pixel position. Optionally, the above registration process can adopt a direct registration method based on grayscale information, a matching method based on feature points, or a registration method based on the transform domain, including but not limited to these.

[0309] Let the set of frames participating in the timing consistency analysis be:

[0310] ;

[0311] First, calculate the soft saliency response of each frame participating in the temporal consistency analysis:

[0312] ;

[0313] The time series consistency score can then be expressed as:

[0314] ;

[0315] in, Indicates the pixel position of the current frame. Temporal consistency score at the location, Indicates the first Temporal soft saliency response at the corresponding frame position, This represents a reference threshold for the significance of time series data, typically. The value should be equal to or slightly lower than the residual significance reference threshold in S3.3. This allows for the loose inclusion of temporal support information from adjacent frames. In one specific embodiment, The methods for determining the value of include, but are not limited to: the fixed proportion method, that is, letting ,in It is a positive number less than or equal to 1, preferably Alternatively, a robust statistical method based on temporal data distribution can be used, which involves collecting data from each frame participating in the temporal analysis at each pixel location. The value is used to calculate the median or truncated mean as the adaptive value for that pixel. The default setting uses a fixed ratio. If the registration error leads to serious misjudgment of the timing support, then switch to the timing median adaptive method. This represents the corresponding mapping slope, which controls the sharpness of the transition from temporally unstable to stable. Since temporal fusion should be smooth, Typically, a moderate or slightly smaller value is chosen. In practical implementation, A fixed empirical value method can be used, directly setting it as a constant, preferably. Optimal selection Alternatively, an adaptive value selection method based on temporal data distribution can be used, which calculates the value of each pixel location in each frame involved in the analysis. Standard deviation of values ,make ,in The proportionality coefficient (preferably) In this embodiment, a fixed default value is used. If the baseline scheme is found to be overly sensitive or overly smooth in the experiment, the scheme is switched to a local adaptive scheme.

[0316] This step can be omitted when the input is a single-frame fluorescence image or when temporal consistency analysis is not performed.

[0317] S3.9: Significance score of weak fluorescence signal obtained by fusion:

[0318] The residual significance score, multi-scale stability significance score, spatial continuity score, and optional temporal consistency score are fused to obtain a weak fluorescence signal significance score map.

[0319] In one implementation, the fusion response term can be calculated first:

[0320] ;

[0321] Then, by combining the background estimation reliability score with weighted normalization, the significance score of the weak fluorescence signal is obtained:

[0322] ;

[0323] in, Indicates pixel position The significance score of the weak fluorescence signal at the location; The fusion response term represents residual significance, multi-scale stability, spatial continuity, and temporal consistency. , , , , representing the fusion weights for residual significance, multi-scale stability, spatial continuity, and temporal consistency, respectively, are all non-negative real numbers; This indicates the reliability score for the background estimation.

[0324] When not using time-series consistency analysis, we can set:

[0325] .

[0326] To ensure the stability of the scoring range, range constraints can be imposed on the fusion results:

[0327] ;

[0328] in, The closer to 1, the more likely a real weak fluorescence signal is to exist at that location; the closer to 0, the more likely the location is a background region, random noise fluctuations, or insignificant residuals.

[0329] S3.10: Generate a set of weak fluorescence candidate features:

[0330] To facilitate the generation of the weak signal protection weight map in subsequent S4, this step combines the significance score of the weak fluorescence signal with relevant auxiliary features to form a set of weak fluorescence candidate features:

[0331] ;

[0332] in, Indicates pixel position The set of weak fluorescence candidate features at the location.

[0333] When using temporal consistency analysis, the weak fluorescence candidate feature set may also include ,Right now:

[0334] ;

[0335] This feature set is used to subsequently generate a continuous weak signal protection weight map based on saliency strength, background reliability, and spatial / temporal stability.

[0336] Weak fluorescence candidate feature set As input to the subsequent generation of the weak signal protection weight map by S4, when using multi-frame input, a timing consistency score map can also be output. .

[0337] S4: Generate a weak signal protection weight map:

[0338] This step is used to generate a continuously valued weak signal protection weight map based on the weak fluorescence signal significance score map, residual significance score map, multi-scale stability significance score map, spatial continuity score map, and background estimation reliability score map obtained in S3.

[0339] The weak signal protection weight map is used to characterize the degree to which each pixel location in the image should be protected during subsequent background subtraction and denoising processes. The higher the protection weight, the more likely the location is to contain a real weak fluorescence signal, and the background subtraction and denoising intensity should be reduced subsequently; the lower the protection weight, the more likely the location is to belong to the background region or random noise region, and stronger background subtraction and noise suppression can be used subsequently.

[0340] It should be noted that this step does not hard segment the image into fluorescent target regions and background regions. Instead, it generates a continuous weak signal protection weight map, so that weak fluorescence candidate regions, strong fluorescence regions, background regions and uncertain regions can all participate in subsequent processing in the form of continuous weights.

[0341] S4.1: Input for obtaining weak fluorescence candidate features:

[0342] Obtain the set of weak fluorescence candidate features output by S3, denoted as:

[0343] ;

[0344] The weak fluorescence candidate feature set includes noise-normalized residuals, residual significance scores, multi-scale stable significance scores, spatial continuity scores, background estimation reliability scores, and weak fluorescence signal significance scores, which can be expressed as:

[0345] ;

[0346] in, This represents the noise-normalized residual. Indicates the significance score of the residuals. Indicates the multi-scale stability significance score. Indicates spatial continuity score, This indicates the reliability score of the background estimation. The score indicates the significance of the weak fluorescence signal.

[0347] When the input is a multi-frame fluorescence image sequence, and temporal consistency analysis is used in S3, the weak fluorescence candidate feature set may also include a temporal consistency score:

[0348] ;

[0349] Right now:

[0350] ;

[0351] S4.2: Generate the initial protection response based on the significance score:

[0352] The significance score of weak fluorescence signal obtained based on S3 Generate the initial protection response diagram:

[0353] ;

[0354] in, Indicates pixel position The initial protection response at the location has a value ranging from 0 to 1; The reference threshold representing the weak signal protection response. The typical range of values ​​is In one specific embodiment, The methods for determining the value include, but are not limited to: fixed empirical threshold method, direct setting Alternatively, an adaptive thresholding method based on a significance score histogram (such as the Otsu method) can be used to automatically determine the boundary point according to the actual distribution. This represents the slope of the mapping function, which controls the sharpness of the transition of the protection response from weak to strong. The larger the threshold, the closer the decision is to the hard threshold. The smaller the protection, the smoother the protection. In practical implementation, The methods for determining the value include, but are not limited to: fixed empirical methods, directly setting it as a constant, and typical value range. ; or an adaptive value selection method based on the dynamic range of significance scoring, let ,in For the full map standard deviation The proportionality coefficient (preferably) Alternatively, a reverse derivation method based on the desired protection strength can be used to derive the protection strength from the desired transition zone width. .

[0355] when Higher than hour, An increase indicates that the location is more likely to contain a true weak fluorescence signal and should be given stronger protection in subsequent processing; when Below hour, A decrease indicates that the location is more likely to be part of the background or a region of insignificant noise.

[0356] By using the above continuous mapping method, we can avoid directly using hard thresholds to divide the image into target and background regions, so that pixels in weak signal boundaries, low brightness candidate regions or uncertain regions can still obtain a degree of protection commensurate with their saliency.

[0357] S4.3: Combine background estimation with reliability correction of protection response:

[0358] In S3, the background estimation reliability score Used to reflect the consistency of the background candidate layer. When A higher value indicates a more reliable initial background layer; when... A lower value indicates that there is significant uncertainty in the background estimation at that location.

[0359] If the background estimation is unreliable, directly subtracting strong background signals may result in the incorrect subtraction of true weak fluorescence signals. Therefore, this step further incorporates a background estimation reliability score to conservatively correct the initial protection response.

[0360] In one implementation, the protection response after background reliability correction can be obtained according to the following formula:

[0361] ;

[0362] in, This indicates the protection response after background reliability correction; This represents the background uncertainty compensation coefficient, which controls the strength of the improvement in protection response when the background estimate is unreliable. The larger the value, the stronger the compensation. In practical implementation, The methods for determining the value include, but are not limited to: fixed empirical methods, directly setting it as a constant, with a typical value range of [missing information]. If the false negative rate of weak signals is too high, increase it appropriately; if the background area is overprotected, decrease it appropriately. Indicates the reliability score of the background estimation; This indicates the significance score of the residuals.

[0363] This formula indicates that when a positive background residual signal exists at a certain location and the reliability of the background estimation is low, the protection response at that location should be appropriately increased to make the subsequent background subtraction process more conservative. This can avoid incorrectly subtracting the true weak fluorescence signal as background due to an overestimation of the background layer or an inappropriate selection of the background scale.

[0364] If the background estimation at a certain location is uncertain, but the residual significance is low, then The smaller size limits the improvement in protection response, thus avoiding overprotection of large background areas due to background uncertainty.

[0365] S4.4: Combining multi-scale stability and spatial continuity to modify the protection response:

[0366] True weak fluorescence signals typically exhibit relatively stable responses across multiple background estimation scales and demonstrate a degree of spatial continuity within their local neighborhoods; in contrast, random noise peaks often appear isolated, scale-instable, or lack neighborhood support. Therefore, this step further combines multi-scale stability significance scoring and spatial continuity scoring to correct the conservation response.

[0367] In one implementation, a weak signal structure support factor can be constructed:

[0368] ;

[0369] in, Indicates the structural support factor for weak signals; Indicates the multi-scale stability significance score; Indicates spatial continuity score; and These represent the fusion weights for multi-scale stability and spatial continuity, respectively; It is a preset minimal positive number.

[0370] Further refine the protection response based on the structural support factor:

[0371] ;

[0372] in, This indicates the protection response after structural support modification. To protect the lower limit coefficient, the value range is from 0 to 1.

[0373] When a certain location has high multi-scale stability and spatial continuity, When the value is large, the protection response is preserved; when a location lacks multi-scale stability and spatial neighborhood support, The lower limit coefficient results in a moderately reduced protection response. This can be achieved by setting the lower protection limit coefficient. This can avoid the complete suppression of real but small-area weak fluorescence signals.

[0374] S4.5: Perform edge-preserving spatial smoothing:

[0375] because The results, calculated from pixel-by-pixel saliency scores, may contain local discontinuities or isolated noise responses. To improve the spatial stability of the weak signal protection weight map and avoid excessive diffusion of protection weights across fluorescent structure edges, this step performs edge-preserving spatial smoothing on the protection response.

[0376] Let's assume The local smoothing window centered on is:

[0377] ;

[0378] in, This indicates the scale of the protection weight smoothing window.

[0379] In one implementation, edge-preserving weights related to spatial distance and gray-level similarity are employed:

[0380] ;

[0381] in, Indicates the location of neighboring pixels within a local window. Represents pixels For the center pixel Smoothing weights, Indicates spatial distance control parameters. This represents the grayscale similarity control parameter.

[0382] Based on the aforementioned edge-preserving weights, the protection response is smoothed:

[0383] ;

[0384] in, This indicates the protection response after spatial smoothing.

[0385] This process enables adjacent weakly fluorescent candidate regions with similar grayscale structures to obtain more continuous protection weights, while avoiding unreasonable diffusion of protection weights across obvious edges or strong fluorescent structure boundaries.

[0386] S4.6: Combine the original response and the smoothed response to obtain continuous protection weights:

[0387] To balance the local sensitivity and spatial stability of the weak signal protection weight map, this step modifies the protection response after structural support adjustment. Protection response after spatial smoothing By fusing the data, a continuous weak signal protection weight map is obtained:

[0388] ;

[0389] in, This represents the final generated weak signal protection weight; This represents the fusion coefficient between the original response and the smoothed response.

[0390] when When the value is large, the final protection weight places more emphasis on local saliency responses; when When the value is small, the final protection weight places more emphasis on spatial continuity and stability.

[0391] Generated in the above manner Instead of a binary mask, this is a continuously weighted map. Therefore, the weight map can continuously adjust the background subtraction and denoising intensity in subsequent steps, thereby avoiding the boundary abruptness and false exclusion of weak signals caused by hard segmentation.

[0392] S4.7: Optional timing protection weight consistency correction:

[0393] When the input is a sequence of multiple consecutive fluorescence images, the consistency of protection weights between adjacent frames can be further utilized to perform time-series correction on the weak signal protection weight map of the current frame.

[0394] Let the first The initial protection weight corresponding to the frame is:

[0395] ;

[0396] If there is motion or system jitter between adjacent frames, the weak signal protection weight map of the adjacent frames can be mapped to the coordinate system of the current frame based on the registration relationship. Let the mapped adjacent frame protection weights be:

[0397] ;

[0398] in, Indicates the sequence number of adjacent frames.

[0399] Let the set of frames participating in timing correction be... Then the timing-corrected protection weight can be expressed as:

[0400] ;

[0401] in, Indicates the first Protection weights after frame timing correction; Indicates the weight of the current frame; This represents the fusion weight of the protection weights of adjacent frames.

[0402] In one implementation, Scoring can be based on the registration error and temporal consistency of adjacent frames. Alternatively, the noise level of adjacent frames can be determined. The larger the registration error, the lower the weight of adjacent frames; the higher the temporal consistency, the higher the weight of adjacent frames.

[0403] If timing correction is used, then in subsequent steps, let:

[0404] ;

[0405] S4.8: Forming a set of weak signal protection features:

[0406] To facilitate the protective background subtraction and adaptive denoising in subsequent S5 steps, this step combines the final weak signal protection weight map with relevant auxiliary features to form a weak signal protection feature set:

[0407] ;

[0408] in, Indicates pixel position The set of weak signal protection features at the location; This represents a weighted graph for weak signal protection. Indicates the significance score of weak fluorescence signal; Indicates local noise intensity; Indicates positive background residual; Indicates the initial background layer; This indicates the reliability score for the background estimation.

[0409] When using multi-frame timing correction, the final protection weight can be written as:

[0410] ;

[0411] And take it as the first The protection weight input for subsequent frame processing.

[0412] Weak signal protection feature set As input for subsequent S5 protective background subtraction and adaptive denoising, when multi-frame timing correction is used, a timing-corrected weak signal protection weight map can also be output. .

[0413] S5: Protective Background Removal and Adaptive Denoising:

[0414] This step is used to perform protective background subtraction and adaptive denoising on the corrected fluorescence image based on the noise-background feature set and weak signal protection feature set obtained in the previous steps. Its core principle is to apply different processing intensities to different regions according to the weak signal protection weight, so that weak fluorescence candidate regions are not over-subtracted or over-smoothed, and background regions achieve more sufficient background and noise suppression.

[0415] S5.1: Obtaining Protective Enhanced Input:

[0416] Obtain the noise-background feature set combining the outputs of S1 and S2:

[0417] ;

[0418] Obtain the set of weak signal protection features output by S4:

[0419] ;

[0420] in, Indicates local noise intensity. Indicates the initial background layer. This represents the positive background residual obtained from S3. This represents the weighting diagram for weak signal protection. Indicates the significance score of weak fluorescence signal. This indicates the reliability score for the background estimation.

[0421] It should be noted that, Used to provide a background and noise basis, Used to provide a basis for protection and control in subsequent background subtraction and noise reduction.

[0422] S5.2: Determine the protective background subtraction factor:

[0423] Based on weak signal protection weights and background estimation reliability score Generate a background subtraction factor map:

[0424]

[0425] in, This is a background subtraction factor map, representing pixel positions. Background subtraction factor value at the location, This represents the baseline value for the background deduction factor. This represents the maximum background subtraction factor. This represents the minimum retention factor that is deducted when the background estimate is unreliable, and it satisfies the following:

[0426] ;

[0427] ;

[0428] This expression means: when A larger value indicates that the region is more likely to contain true weak fluorescence signals, and the background subtraction factor decreases; when When the value is low, it indicates that the background estimation is unstable, and the background subtraction becomes more conservative in order to avoid the weak signal being mistakenly subtracted due to the bias in the background estimation.

[0429] S5.3: Perform protective background subtraction:

[0430] According to the background deduction factor diagram and the initial background layer This yields a protective background subtraction image:

[0431] ;

[0432] in, This represents the fluorescence image after background subtraction.

[0433] Unlike uniform background subtraction, this step applies different subtraction intensities to different locations based on weak signal protection weights and background reliability, resulting in stronger suppression of pure background regions and stronger protection of weak fluorescence candidate regions.

[0434] S5.4: Determine the protective noise reduction intensity:

[0435] Based on local noise intensity Positive background residual And weak signal protection weight Generate a noise reduction intensity coefficient map:

[0436] ;

[0437] in, This is a noise reduction intensity coefficient map, representing pixel locations. The noise reduction intensity coefficient value at the location, Indicates the basic denoising intensity. It is a preset, extremely small positive number;

[0438] This formula indicates that when the local noise is stronger than the positive background residual and the weak signal protection weight is lower, the denoising intensity is greater; when the positive background residual is more obvious or the protection weight is higher, the denoising intensity is reduced to reduce the risk of weak fluorescence signals being over-smoothed.

[0439] S5.5: Perform protective adaptive denoising:

[0440] Image after background subtraction Perform edge-preserving denoising to obtain the basic denoised image:

[0441] ;

[0442] in, The edge-preserving denoising operator can be represented by bilateral filtering, guided filtering, nonlocal means, wavelet denoising, total variational denoising, or other edge-preserving denoising methods.

[0443] Then, based on the denoising intensity coefficient map, the background subtracted image is fused with the base denoised image:

[0444] ;

[0445] in, This represents a protective adaptive denoising image.

[0446] This process makes the background and low-protection areas closer to the denoising result, while the weak fluorescence candidate areas are closer to the original response after background subtraction, thus preserving weak signal details.

[0447] S5.6: Generate preliminary enhanced image:

[0448] To improve the visibility of true weak fluorescence candidate signals, appropriate residual compensation can be performed based on weak signal protection weights, weak fluorescence signal significance scores, and positive background residuals.

[0449] ;

[0450] in, This indicates the initial image enhancement. This represents the weak signal compensation coefficient, typically... or .

[0451] This compensation term only plays a significant role when the protection weight is high, the weak signal is highly significant, and the background estimation is relatively reliable, thereby avoiding excessive enhancement of background noise or uncertain regions of background estimation.

[0452] If residual compensation is not required, we can set:

[0453] ;

[0454] at this time:

[0455] .

[0456] After processing with S5, a protective background subtraction image is obtained. Protective adaptive denoising image Preliminary image enhancement Background deduction factor chart Denoising intensity coefficient diagram ;

[0457] The above results serve as inputs for the output of S6 strength fidelity verification and enhancement results.

[0458] S6: Strength fidelity verification and enhancement result output:

[0459] This step is used to perform a weak fluorescence signal intensity fidelity verification on the preliminary enhanced image obtained in S5, and to correct the enhancement result based on the verification result, so as to obtain the final enhanced fluorescence image.

[0460] The purpose of this step is to avoid the enhancement process, while improving visual contrast, disrupting the relative intensity relationship of weak fluorescence signals, thereby affecting subsequent observation and quantitative analysis.

[0461] S6.1: Constructing Weak Signal Fidelity Verification Weights:

[0462] To avoid hard segmentation of the target region, this step constructs soft verification weights based on weak signal protection weights and weak fluorescence signal significance scores:

[0463] ;

[0464] in, Indicates pixel position The weight for weak signal fidelity verification at a given location. The larger the weight, the more necessary it is for that location to participate in weak fluorescence signal intensity fidelity verification.

[0465] Let's assume The local verification window centered on is:

[0466] ;

[0467] in, This indicates the size of the local verification window.

[0468] S6.2: Calculate the residual intensity of the weak signal before and after enhancement:

[0469] To avoid background interference caused by directly comparing the brightness of the original image and the enhanced image, this step performs fidelity verification based on the weak signal residual components.

[0470] The positive background residual obtained from S3 is used as a weak signal reference before enhancement:

[0471] ;

[0472] Since the background subtraction image of S5 may retain some background components, the weak signal residual estimate in the preliminary enhanced image is defined as:

[0473] ;

[0474] in, This indicates the background components that may be retained after background subtraction. This represents the weak signal residual estimate in the initially enhanced image relative to the preserved background.

[0475] Within the local verification window, calculate the weighted weak signal integral strength before enhancement:

[0476] ;

[0477] Calculate the enhanced weighted weak signal integral strength:

[0478] ;

[0479] Further calculation of the weak signal strength preservation ratio:

[0480] ;

[0481] in, This indicates the proportion of local intensity change of weak fluorescence candidate signals before and after enhancement.

[0482] S6.3: Feedback correction based on strength fidelity results:

[0483] Further construct the feedback correction coefficient:

[0484] ;

[0485] in, This represents the feedback correction coefficient. and These represent the lower and upper limits of the feedback correction coefficient, respectively. This indicates the desired intensity ratio; this value can be set as needed, but empirically it can generally be set to a certain value. Considering the background estimation error, noise disturbance, and the over-enhancement effect that the residual compensation term may introduce, the weak signal strength is allowed to be moderately attenuated within a reasonable range in exchange for improved background suppression and noise stability.

[0486] Based on this coefficient, feedback correction is performed on the weak signal residual components in the preliminary enhanced image:

[0487] ;

[0488] in, This represents the enhanced fluorescence image after intensity fidelity correction.

[0489] This formula indicates that for regions with high fidelity verification weights for weak signals, the residual signal is locally corrected according to the intensity preservation ratio; for background regions or low-protection regions... The value is relatively small, and the corrective effect is weak.

[0490] S6.4: Gray-scale range constraints and result output:

[0491] Enhanced fluorescence image after intensity fidelity correction By constraining the grayscale range, the final enhanced fluorescence image is obtained:

[0492] ;

[0493] in, This indicates the final enhanced fluorescence image. This represents the maximum effective grayscale value corresponding to the current image bit depth.

[0494] Simultaneously, it can output a weak signal protection weight map, a weak fluorescence signal significance score map, a local noise intensity map, and an initial background layer:

[0495] ;

[0496] Optionally, an enhanced local signal-to-noise ratio map can also be output:

[0497] ;

[0498] And local signal-to-background ratio plot:

[0499] ;

[0500] in, and Used to assist in evaluating the detectability of weak fluorescence signals after enhancement.

[0501] The invention will be further explained below with reference to experimental scenarios:

[0502] Figure 3This paper presents simulation comparisons of the method of this invention and different enhancement methods on low signal-to-noise ratio (SNR) fluorescence images. The experiments used constructed low SNR fluorescence images as input data. Real fluorescent targets were simulated using two-dimensional Gaussian fluorescent spots of varying intensities and sizes. Gradually varying backgrounds were constructed using two-dimensional low-order polynomial surfaces, low-frequency Gaussian blurred random fields, or low-frequency backgrounds generated by bicubic interpolation to simulate changes in sample thickness, uneven illumination, and slow grayscale changes caused by scattered light. The autofluorescence background was generated by superimposing multiple randomly distributed low-frequency fluorescent regions, with their spatial distribution and intensity statistically simulated according to a random Gaussian field or Perlin noise field, used to simulate tissue autofluorescence and non-specific background luminescence. The noise component was generated using a hybrid model of Poisson noise and Gaussian readout noise. Poisson noise was randomly generated based on the statistical characteristics of fluorescence photon counts for each pixel, while Gaussian noise was randomly superimposed based on the standard deviation of camera readout noise, thus simulating the actual fluorescence imaging environment. Subsequently, the original low SNR image, ordinary background subtraction, ordinary denoising, unprotected weighting methods, and the method of this invention were used for processing. The figure shows the output signal-to-noise ratio (SNR), output signal-to-background ratio (SNR), and weak signal retention rate. The output SNR is calculated based on the average signal intensity after enhancement within the real fluorescent target area and the standard deviation of noise in the background area, used to evaluate the degree of signal enhancement relative to random noise. The output SNR is calculated based on the ratio of the average gray level of the real fluorescent target area to the average gray level of the neighboring background area, used to evaluate the ability to distinguish the target signal from the background. The weak signal retention rate is calculated based on the proportion of real weak fluorescent target pixels that maintain an effective response after enhancement to the total number of real weak fluorescent target pixels, used to evaluate the degree of retention of the real weak fluorescent signal during the enhancement process. An effective response is defined as a pixel gray value after enhancement that is higher than the local background mean plus three times the noise standard deviation, or a residual retention rate before and after enhancement that is not less than 0.3.

[0503] Depend on Figure 3 It can be seen that the method of this invention achieves superior results in all three indicators. Specifically, the output signal-to-noise ratio and output signal-to-background ratio are transformed by a logarithm of 20 times the amplitude ratio (…). The grayscale signal was converted to dB units (processed by amplitude), improving the output signal-to-noise ratio from approximately 6.5 dB to approximately 33.6 dB, and the output signal-to-background ratio from approximately 3.2 dB to approximately 25.7 dB. The weak signal retention rate reached approximately 92%, all significantly better than ordinary background subtraction, ordinary denoising, and methods without protected weights. The results show that this invention, by jointly adjusting the background subtraction and adaptive denoising processes through weak signal protection weights, effectively suppresses background noise while better preserving the true weak fluorescence signal, achieving a simultaneous improvement in sensitivity and output signal-to-noise ratio.

[0504] Figure 4The ablation experimental results of the key module of the method of this invention are presented. The experiment is based on... Figure 3 In the same low signal-to-noise ratio (SNR) fluorescence image simulation environment, based on the complete method, key modules such as multi-scale response, spatial continuity, background reliability correction, protection weight control, and intensity fidelity feedback were removed, and the output SNR, weak signal retention rate, and intensity fidelity index were statistically analyzed. The output SNR was calculated based on the standard deviation of the average signal of the real target region and the background noise, and was used to evaluate the overall enhancement quality. The weak signal retention rate was calculated by statistically analyzing the proportion of real weak fluorescence target pixels that still maintain an effective response after enhancement, and was used to evaluate the preservation ability of real weak fluorescence targets. The intensity fidelity index was statistically analyzed based on the normalized intensity consistency between the enhanced real fluorescence target region and the original real fluorescence target region. In this embodiment, the Pearson correlation coefficient was used as the intensity fidelity index to evaluate the relative intensity retention between the enhanced fluorescence signal and the original real signal.

[0505] Figure 4 The complete method achieved optimal results across all three metrics. The most significant decrease was observed when the protection weight control module was removed, reducing the weak signal retention rate from approximately 91.2% to approximately 36.5%, indicating that the weak signal protection weight is crucial for preventing the incorrect subtraction of genuine weak fluorescence targets. Removing the intensity fidelity feedback reduced the intensity fidelity index from approximately 0.92 to approximately 0.48, demonstrating that intensity fidelity verification effectively maintains the relative intensity relationship of weak fluorescence targets before and after enhancement. Removing the remaining modules resulted in varying degrees of decrease in both the output signal-to-noise ratio and the weak signal retention rate, indicating that multi-scale background response, spatial continuity, and background reliability correction all positively impact the enhancement effect.

[0506] Figure 5This paper demonstrates the verification results of the correspondence between the weak signal protection weight map of this invention and real weak fluorescence targets. The experiment, based on simulation data of known real weak fluorescence locations, statistically analyzes the generated weak signal protection weight map and evaluates it using four indicators: position consistency, high weight coverage, peak matching rate, and positioning accuracy score. Position consistency is calculated based on the overlap ratio between the peak position of the protection weight and the center position of the real fluorescence target, or the average positioning error, and is used to evaluate the correspondence between the peak value of the protection weight and the real weak fluorescence location. High weight coverage is calculated based on the proportion of pixels with protection weights greater than a preset threshold within the real weak fluorescence target area, and is used to evaluate the degree to which the real weak fluorescence area is covered by high protection weights. Peak matching rate is calculated based on the one-to-one correspondence ratio between the local peak value of the protection weight and the peak value of the real fluorescence intensity, and is used to evaluate the consistency between the peak value of the protection weight and the peak value of the real fluorescence. The positioning accuracy score is a normalized score that comprehensively considers position deviation, target coverage, and peak matching, and is used to comprehensively evaluate the accuracy of the protection weight in locating the real weak fluorescence target.

[0507] Depend on Figure 5 As can be seen, the weak signal protection weight map generated by this invention has a high degree of consistency with the real weak fluorescence target, with a positional consistency of approximately 93%, a high-weight coverage of approximately 91%, a peak matching rate of approximately 95%, and a positioning accuracy score of approximately 88%. These results demonstrate that the weak signal saliency analysis in this invention can accurately identify potential weak fluorescence regions, and the generated weak signal protection weight map can effectively characterize the positional distribution of the real weak fluorescence target, providing a reliable basis for subsequent background subtraction intensity adjustment and adaptive denoising, thereby improving the protection capability of weak fluorescence signals.

[0508] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A weak signal-protective enhancement method for low signal-to-noise ratio fluorescence images, characterized in that, Includes the following steps: S1. Obtain the original fluorescence image to be enhanced, and perform basic correction on the original fluorescence image to obtain the corrected fluorescence image; S2. Based on the corrected fluorescence image, estimate the local noise intensity map, multiple scale background candidate layers, initial background layer and initial background residual map to form a noise-background feature set; S3. Based on the noise-background feature set, noise normalization is performed on the positive background residual in the corrected fluorescence image, and combined with multi-scale response, spatial continuity and background estimation reliability, a weak fluorescence signal saliency score map is obtained. S4. Generate a continuously valued weak signal protection weight map based on the weak fluorescence signal saliency score map. The weak signal protection weight map is used to characterize the degree to which each pixel position should be protected in the subsequent background subtraction and denoising process. S5. Adjust the background subtraction intensity and denoising intensity according to the weak signal protection weight map, and perform protective background subtraction and protective adaptive denoising on the corrected fluorescence image to obtain a preliminary enhanced image; S6. Based on the weak signal protection weight map, the weak signal residual components in the preliminary enhanced image are subjected to intensity fidelity verification and feedback correction to obtain the final enhanced fluorescence image.

2. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 1, characterized in that, Specifically, S1 is: Acquire the raw fluorescence image; The original fluorescence image is dark-field corrected based on the dark-field reference information to obtain a dark-field corrected image. The dark field corrected image is normalized and flat-field corrected based on the flat-field reference information to obtain the flat-field corrected fluorescence image; The grayscale of the flat-field rectified image is normalized based on the exposure time and camera gain, and the grayscale range is constrained to obtain the rectified fluorescence image. .

3. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 1, characterized in that, Specifically, S2 is: The corrected fluorescence image Calculate the local mean within the local neighborhood window , Indicates the neighborhood window scale; Estimating the local noise intensity map based on the local intensity-related noise model: ； in, Indicates pixel position Local noise intensity at the location, This represents the intensity coefficient associated with photon shot noise. This indicates a bias term related to readout noise, dark current noise, or fixed noise. It is a preset, extremely small positive number; Multiple background estimation scales are set, and multi-scale background estimation is performed on the corrected fluorescence image to obtain multiple scale background candidate layers. ,in, Indicates the scale number; For the multiple scale background candidate layers Weighted fusion is performed to obtain the initial background layer. ; Based on the corrected fluorescence image and the initial background layer Calculate the initial background residual map : ； The corrected fluorescence image, local noise intensity map, initial background layer, initial background residual map, and multiple scale background candidate layers are combined to form a noise-background feature set.

4. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 3, characterized in that, In step S3, noise normalization processing is performed on the positive background residual, specifically as follows: From the initial background residual map Extracting positive background residuals : ； According to the local noise intensity map The positive background residual is normalized to obtain the noise-normalized residual. ; The noise-normalized residuals are mapped to residual significance scores: ； in, Indicates the significance score of the residuals. This represents the reference threshold for the significance of the residuals. This represents the slope of the mapping function.

5. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 4, characterized in that, The S3 section combines multi-scale response, spatial continuity, and background estimation reliability to obtain a weak fluorescence signal significance score map, specifically: Based on multiple background candidate layers at different scales, the positive background residual response and noise normalized response at the corresponding scales are calculated respectively. Calculate the multi-scale stability significance score based on the noise normalized response at multiple scales. ; The spatial continuity score is obtained by statistically analyzing the soft saliency response of neighboring pixels within the local support window. ; The background estimation reliability score is calculated based on the differences between the candidate background layers at multiple scales and the initial background layer. ; Based on the residual significance score Multiscale stability significance score Spatial continuity score and background estimation reliability score The fusion yielded a significance score map for weak fluorescence signals: When the original fluorescence image is a multi-frame fluorescence image sequence and time-series consistency analysis is required: Obtain the noise normalized residuals for each frame. Based on the current processing frame, and using the noise normalization residuals corresponding to each frame as a benchmark, the temporal consistency score of the current processing frame is calculated. ,in, Indicates the frame number; ； , , , These represent the fusion weights for residual significance, multi-scale stability, spatial continuity, and temporal consistency, respectively. When not using time-series consistency analysis, let .

6. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 5, characterized in that, Specifically, S4 is: According to the significance score chart of the weak fluorescence signal The initial protection response diagram is generated through a continuous mapping function: ； in, Indicates the initial protection response. This represents the reference threshold for weak signal protection response. Represents the slope of the mapping function; Reliability score based on background estimation and residual significance score The initial protection response diagram is conservatively modified to obtain the protection response after background reliability correction: ； in, This indicates the protection response after background reliability correction. Indicates the background uncertainty compensation coefficient; Based on multi-scale stability significance score Spatial continuity score Constructing weak signal structure support factors: ； Based on the weak signal structural support factor, the protection response is modified to obtain the structural support-corrected protection response. : ； in, This indicates the protection response after structural support modification. To protect the lower limit coefficient; Edge-preserving spatial smoothing is applied to the protection response after structural support modification, and the protection responses before and after smoothing are fused to obtain a continuously valued weak signal protection weight map. .

7. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 6, characterized in that, In step S5, the background subtraction intensity is adjusted according to the weak signal protection weight map, specifically as follows: According to the weak signal protection weight diagram and background estimation reliability score Generate a background subtraction factor map: in, This is a background subtraction factor map, representing pixel positions. Background subtraction factor value at the location, This represents the baseline value for the background deduction factor. This represents the maximum background subtraction factor. This represents the minimum retention factor when the background estimate is unreliable; According to the background deduction coefficient diagram and the initial background layer For the corrected fluorescence image Perform protective background subtraction to obtain the fluorescence image after background subtraction: 。 8. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 7, characterized in that, In step S5, the noise reduction intensity is adjusted according to the weak signal protection weight map, specifically as follows: Based on the local noise intensity map Positive background residual Weak signal protection weight diagram Generate a noise reduction intensity coefficient map: ； in, This is a noise reduction intensity coefficient map, representing pixel locations. The noise reduction intensity coefficient value at the location, Indicates the basic denoising intensity; Fluorescence image after background subtraction Perform edge-preserving denoising to obtain the basic denoised image. ; According to the noise reduction intensity coefficient diagram Fluorescent image after background subtraction Compared with the base denoised image The images are fused to obtain a protective adaptive denoised image: 。 9. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 8, characterized in that, The preliminary enhanced image obtained in S5 is as follows: According to the protective adaptive denoising image Weak signal protection weight diagram Significance score of weak fluorescence signal Background estimation reliability score and positive background residual Weak signal residual compensation is performed to obtain a preliminary enhanced image: ； in, This indicates the initial image enhancement. This represents the weak signal compensation coefficient.

10. The weak signal protection enhancement method for low signal-to-noise ratio fluorescence images according to claim 9, characterized in that, In step S6, intensity fidelity verification and feedback correction are performed on the preliminary enhanced image, specifically as follows: According to the weak signal protection weight diagram Significance score of weak fluorescence signal Construct weak signal fidelity verification weights: ； With positive background residual As a reference residual for the weak signal before enhancement, and based on the preliminary enhanced image and the background components that may be retained after background subtraction, the estimated weak signal residual after enhancement is determined: ； Within the local verification window, based on the weak signal fidelity verification weights at each pixel position within the window, the weak signal reference residual before enhancement and the weak signal residual estimate after enhancement are weighted integrals to obtain the weighted weak signal integral strength before enhancement and the weighted weak signal integral strength after enhancement. Calculate the weak signal intensity preservation ratio based on the weighted weak signal integral intensity before and after enhancement. ; A feedback correction coefficient is constructed based on the weak signal strength preservation ratio. : ； in, This indicates that the expected intensity remains proportional. and These represent the lower and upper limits of the feedback correction coefficient, respectively. The weak signal residual components in the initial enhanced image are corrected using the feedback correction coefficient to obtain the enhanced fluorescence image after intensity fidelity correction. : ； Enhanced fluorescence image after intensity fidelity correction By constraining the grayscale range, the final enhanced fluorescence image is obtained.

Images