Microscopic image intelligent processing system

Abstract

The application relates to the technical field of image processing, and discloses a microscopic image intelligent processing system, which comprises the following units: an image acquisition unit acquires microscopic original image data; a neighborhood space definition unit determines a local space neighborhood matrix of each pixel point; a gradient component decoupling unit calculates a local gradient amplitude and constructs a space weight mask, wherein a weight coefficient decreases with the increase of the amplitude, a low-frequency background gradient component is determined by using a filtering operator, and the local gradient amplitude is projected to a zero space of the component to determine an intrinsic structure gradient; and an image reconstruction unit completes anisotropic diffusion reconstruction according to the spatial vector direction of the intrinsic structure gradient, the application realizes orthogonal decoupling of a target edge and a non-uniform background field, eliminates diffraction halo interference caused by a point spread function limitation, and has obvious advantages in maintaining target boundary sharpness and inhibiting non-structural noise.

Description

Technical Field

[0001] This invention relates to an intelligent microscopic image processing system, belonging to the field of image processing technology. Background Technology

[0002] Current microscopic image processing solutions typically employ median filtering, histogram equalization, and other methods to preprocess the original image and improve target contrast. Their physical basis lies in the assumption that the background noise of the image is randomly distributed and the global light field is constant, thereby extracting target structural information using linear transformations in the spatial or frequency domains. The optical drift generated during microscopic digital imaging is not additive noise that is completely independent of the target structure, but rather a non-stationary field constrained by the sample refractive index, the flatness of the slide, and the point spread function of the optical system. When photons penetrate biological samples or semiconductor structures of non-uniform thickness, the resulting phase delay and diffraction effects induce low-frequency gradient deviations with definite directions in the discrete pixel array. These deviations are highly coupled with the target edge gradients in the spatial and frequency domains, constituting the fundamental constraints in the process of microscopic image enhancement.

[0003] To address non-stationary illumination interference, conventional methods employ large-scale low-pass filtering operators to predict the background field and use background subtraction to recover target features. However, due to the lack of linear separability between the background light field and the target entity in spatial distribution, large-scale convolution calculations, while smoothing the background, also introduce high-frequency energy from the target edges into the background gradient vector, causing feature leakage at structural boundaries. This leakage leads to biases in background gradient prediction and directly propagates to subsequent feature extraction stages. Simply increasing the resolution of sensing optical hardware cannot fundamentally resolve the feature coupling contradiction. Conventional software control methods also have inherent limitations. For example, Chinese invention patent CN114022384B discloses a method based on anisotropic... The anisotropic diffusion model adaptive edge-preserving denoising method introduces bilateral filtering and local variance into the anisotropic diffusion model in order to protect image texture details while filtering out noise. The underlying principle relies on the ideal premise that the background noise presents an isotropic and unstructured distribution. In actual microscopic observation, due to the non-uniform sample thickness, light field drift induces low-frequency pseudo-gradient components with clear physical directionality. The comparison file relies on the local variance and gradient scalar-based judgment mechanism, which cannot separate the real physical boundary of the target from the directional low-frequency interference in the spatial frequency domain. When facing highly overlapping gray-level ranges, the diffusion coefficient adjustment mechanism exacerbates the illegal penetration of edge energy into the background channel, resulting in feature aliasing and topological collapse at the weak contrast target boundary.

[0004] Therefore, how to achieve precise orthogonal decoupling between intrinsic gradient and optical drift gradient, and ensure the edge recall rate of the microscopic target topology without relying on customized optical hardware, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A microscopic image intelligent processing system, comprising: Image acquisition unit, used to acquire raw microscopic image data; The neighborhood space definition unit is used to determine the local spatial neighborhood matrix of each pixel in the original microscopic image data, providing topological support for subsequent tensor field modeling. The gradient component decoupling unit includes the following steps: Step S11, calculating the local gradient magnitude of each pixel; Step S12, constructing a spatial weight mask based on the local gradient magnitude, wherein the weight coefficient of each pixel in the spatial weight mask decreases as the local gradient magnitude increases, in order to suppress the weight contribution of high-frequency edge components in subsequent background estimation; Step S13, performing convolution operation on the original microscopic image data using a preset low-pass filter operator combined with the spatial weight mask, and determining the low-frequency background gradient component by shielding the spatial diffusion of structural edge energy into the low-frequency background path; Step S14, projecting the local gradient magnitude onto the null space of the low-frequency background gradient component, and determining the intrinsic structure gradient by stripping the gradient interference component in the same direction as the background field, wherein the dot product of the intrinsic structure gradient and the low-frequency background gradient component is 0. The image reconstruction unit is used to perform anisotropic diffusion reconstruction on the original microscopic image data based on the spatial vector direction of the intrinsic structure gradient and its local geometric characteristics. By controlling the diffusion tensor to perform grayscale smoothing in the tangent direction of the physical edge and implementing diffusion suppression in the normal direction perpendicular to the physical edge, the unit can maintain the boundary sharpness of the target object while filtering out unstructured noise in the background field and output the target microscopic image.

[0006] Preferably, when constructing the spatial weight mask, the gradient component decoupling unit maps the local gradient magnitude to a normalized suppression factor and limits the contribution ratio of high gradient pixels in the convolution operation through the spatial weight mask, so as to block the spatial leakage of the edge energy of the target object to the low-frequency background gradient component, thereby eliminating the pseudo background component generated at the structural boundary due to the point spread function limitation, and making the spatial vector direction of the intrinsic structural gradient consistent with the physical edge contour of the target object.

[0007] Preferably, when the gradient component decoupling unit completes step S14, it calculates the unit projection vector of the low-frequency background gradient component in the local gradient magnitude direction, and subtracts the unit projection vector from the local gradient magnitude to achieve nonlinear separation of the target intrinsic features and background field features, ensuring that the intrinsic structure gradient only characterizes the intrinsic geometric features of the target object.

[0008] Preferably, when the image reconstruction unit completes anisotropic diffusion reconstruction, it constructs a diffusion damping function that is positively correlated with the intrinsic structure gradient intensity, controls the diffusion flow to complete grayscale smoothing in the edge tangent direction of the target object, and stops the diffusion motion in the direction perpendicular to the edge, so as to repair the boundary blur caused by imaging system noise.

[0009] Preferably, when determining the local spatial neighborhood matrix, the neighborhood definition unit calculates the gray-level variance within the window region of the current pixel point; when the gray-level variance is greater than the preset structure description threshold, the sampling radius of the local spatial neighborhood matrix is ​​reduced; when the gray-level variance is less than the structure description threshold, the sampling radius of the local spatial neighborhood matrix is ​​increased, so as to achieve adaptive coverage of targets at different scales.

[0010] Preferably, the image reconstruction unit is also used to identify unstructured interference in the original microscopic image data. By scanning the intrinsic structure gradient through connected components, it identifies and removes isolated gradient spots with a spatial continuity length of less than 5px, thereby achieving automatic filtering of interference from coverslip reflection spots or sample bubbles, ensuring that the output target microscopic image eliminates artifacts while maintaining edge recall.

[0011] Preferably, the gradient component decoupling unit adopts a parallel computing architecture to synchronously process the discrete row vectors of the original microscopic image data, so that the processing latency of a single frame image is no more than 20ms, realizing real-time high-frequency dynamic monitoring of pathological samples, thereby supporting continuous image tracking in micron-level part detection or biological live observation scenarios.

[0012] Preferably, the intelligent microscopic image processing system further includes a calibration feedback unit, which is used to monitor the edge sharpness of the target microscopic image in real time, and linearly adjust the transformation step size of the weight coefficient in step S12 according to the deviation value of the edge sharpness relative to the preset target, so as to achieve adaptive gain for sample images with different staining depths.

[0013] Preferably, the intelligent microscopic image processing system further includes a storage unit for recording the historical distribution characteristics of the low-frequency background field; in step S13, the gradient component decoupling unit uses the historical distribution characteristics and the pixel-level mean of the current frame extraction result to perform weighted correction on the low-frequency background gradient component, so as to eliminate transient fluctuation interference of the light source system at frequencies above 100Hz.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In intelligent microscopic image processing, by constructing a structure-suppressed spatial mask and an orthogonal decoupling mechanism for intrinsic gradients, a deep separation between the non-stationary background light field gradient and the intrinsic gradient of the target is achieved. By making the convolution weights inversely proportional to the initial local gradient magnitude, the physical leakage of high-frequency structural energy in the image to the low-frequency background field calculation loop is blocked, thereby ensuring that the extracted background light field gradient vector has high physical purity. Even in the extreme low-contrast condition where the target edge and background grayscale highly overlap, the system can still accurately eliminate the pseudo-gradient components generated by optical drift through orthogonal projection operation, and obtain the intrinsic first derivative that only represents the physical boundary of the sample, thus solving the problem of feature vector deflection caused by the inability of traditional operators to distinguish unstructured optical noise.

[0015] 2. By utilizing the decoupled second-order tensor field to guide the anisotropic logic transfer, a strong constraint protection mechanism for the physical topological boundary of microscopic targets is constructed. Based on the directional coherence function, it precisely controls the diffusion path of energy in the pixel neighborhood, causing the pixel response during image processing to converge towards the real physical edge. At the same time, it has a directional suppression effect on unstructured background spots. This topological reconstruction driven by intrinsic gradients effectively avoids the diffraction halo interference that traditional smoothing algorithms inevitably produce when processing small-sized targets close to the point spread function limit. This allows the processed image data stream to maintain edge sharpness while eliminating boundary collapse and hole phenomena, enhancing the logical determinism of subsequent image segmentation actions.

[0016] 3. By using the nonlinear constraints of the local spatial neighborhood matrix and tensor operations in analytical form, a highly efficient physical optical error hedging mode is achieved. It compensates for the inherent imaging defects of the microscope objective and light source system through pure digital signal processing logic. It can automatically filter unstructured artifacts such as coverslip reflections and sample bubbles without relying on customized large-aperture objectives or complex closed-loop light source driving hardware. This mechanism reflects the adaptability of the algorithm's underlying logic to physical environment constraints, enabling the system to maintain a high edge recall rate and target recognition stability without increasing hardware complexity. Attached Figure Description

[0017] Figure 1 This is a flowchart of gradient decoupling and intelligent reconstruction processing of microscopic images according to the present invention. Figure 2 This is a diagram of the hardware and software collaborative microscopic imaging perception and computing architecture of the present invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A microscopic image intelligent processing system, comprising: Image acquisition unit, used to acquire raw microscopic image data; The neighborhood space definition unit is used to determine the local spatial neighborhood matrix of each pixel in the original microscopic image data, providing topological support for subsequent tensor field modeling. The gradient component decoupling unit includes the following steps: Step S11, calculating the local gradient magnitude of each pixel; Step S12, constructing a spatial weight mask based on the local gradient magnitude, wherein the weight coefficient of each pixel in the spatial weight mask decreases as the local gradient magnitude increases, in order to suppress the weight contribution of high-frequency edge components in subsequent background estimation; Step S13, performing convolution operation on the original microscopic image data using a preset low-pass filter operator combined with the spatial weight mask, and determining the low-frequency background gradient component by shielding the spatial diffusion of structural edge energy into the low-frequency background path; Step S14, projecting the local gradient magnitude onto the null space of the low-frequency background gradient component, and determining the intrinsic structure gradient by stripping the gradient interference component in the same direction as the background field, wherein the dot product of the intrinsic structure gradient and the low-frequency background gradient component is 0. The image reconstruction unit is used to perform anisotropic diffusion reconstruction on the original microscopic image data based on the spatial vector direction of the intrinsic structure gradient and its local geometric characteristics. By controlling the diffusion tensor to perform grayscale smoothing in the tangent direction of the physical edge and implementing diffusion suppression in the normal direction perpendicular to the physical edge, the unit can maintain the boundary sharpness of the target object while filtering out unstructured noise in the background field and output the target microscopic image.

[0021] Preferably, when constructing the spatial weight mask, the gradient component decoupling unit maps the local gradient magnitude to a normalized suppression factor and limits the contribution ratio of high gradient pixels in the convolution operation through the spatial weight mask, so as to block the spatial leakage of the edge energy of the target object to the low-frequency background gradient component, thereby eliminating the pseudo background component generated at the structural boundary due to the point spread function limitation, and making the spatial vector direction of the intrinsic structural gradient consistent with the physical edge contour of the target object.

[0022] Preferably, the gradient component decoupling unit determines the weight of pixel (i,j) in the spatial weight mask based on the local gradient magnitude. Its computation path satisfies: ,in, α is the local gradient magnitude of the corresponding pixel, and α is the preset edge suppression coefficient, and α is greater than 0.

[0023] Preferably, when the gradient component decoupling unit completes step S14, it calculates the unit projection vector of the low-frequency background gradient component in the local gradient magnitude direction, and subtracts the unit projection vector from the local gradient magnitude to achieve nonlinear separation of the target intrinsic features and background field features, ensuring that the intrinsic structure gradient only characterizes the intrinsic geometric features of the target object.

[0024] Preferably, when the image reconstruction unit completes anisotropic diffusion reconstruction, it constructs a diffusion damping function that is positively correlated with the intrinsic structure gradient intensity, controls the diffusion flow to complete grayscale smoothing in the edge tangent direction of the target object, and stops the diffusion motion in the direction perpendicular to the edge, so as to repair the boundary blur caused by imaging system noise.

[0025] Preferably, when determining the local spatial neighborhood matrix, the neighborhood definition unit calculates the gray-level variance within the window region of the current pixel point; when the gray-level variance is greater than the preset structure description threshold, the sampling radius of the local spatial neighborhood matrix is ​​reduced; when the gray-level variance is less than the structure description threshold, the sampling radius of the local spatial neighborhood matrix is ​​increased, so as to achieve adaptive coverage of targets at different scales.

[0026] Preferably, the image reconstruction unit is also used to identify unstructured interference in the original microscopic image data. By scanning the intrinsic structure gradient through connected components, it identifies and removes isolated gradient spots with a spatial continuity length of less than 5px, thereby achieving automatic filtering of interference from coverslip reflection spots or sample bubbles, ensuring that the output target microscopic image eliminates artifacts while maintaining edge recall.

[0027] Preferably, the gradient component decoupling unit adopts a parallel computing architecture to synchronously process the discrete row vectors of the original microscopic image data, so that the processing latency of a single frame image is no more than 20ms, realizing real-time high-frequency dynamic monitoring of pathological samples, thereby supporting continuous image tracking in micron-level part detection or biological live observation scenarios.

[0028] Preferably, the intelligent microscopic image processing system further includes a calibration feedback unit, which is used to monitor the edge sharpness of the target microscopic image in real time, and linearly adjust the transformation step size of the weight coefficient in step S12 according to the deviation value of the edge sharpness relative to the preset target, so as to achieve adaptive gain for sample images with different staining depths.

[0029] Preferably, the intelligent microscopic image processing system further includes a storage unit for recording the historical distribution characteristics of the low-frequency background field; in step S13, the gradient component decoupling unit uses the historical distribution characteristics and the pixel-level mean of the current frame extraction result to perform weighted correction on the low-frequency background gradient component, so as to eliminate transient fluctuation interference of the light source system at frequencies above 100Hz.

[0030] Example 1: In high-frequency dynamic observation scenarios of low-contrast micron-scale cell slices, the grayscale of the target object overlaps with the background and the contrast is lower than a preset threshold. Due to the limitations of the microscope objective point spread function and the phase delay caused by the fluctuation of slice thickness, a low-frequency optical drift gradient exists in a specific direction within the field of view of the original microscopic image data. When using the structure tensor extraction operator to calculate the original microscopic image data containing a non-stationary light field background, the pseudo-gradient of the background generated by the non-uniformity of the physical structure boundary and the background light field overlaps, causing target boundary collapse and diffraction halo interference. The intelligent microscopic image processing system provided by this invention reconstructs the boundary constraints of the background extraction logic, and uses the gradient component decoupling unit and the image reconstruction unit to initiate the conversion of multi-dimensional pixel space physical quantities and special... The orthogonal decoupling process is as follows: The image acquisition unit acquires the original microscopic image data, and the neighborhood space definition unit determines the local spatial neighborhood matrix of each pixel in the original microscopic image data. The local gray-level variance fluctuation amplitude is directly positively correlated with the image spatial frequency band energy distribution. High-frequency regions such as physical structure edges correspond to larger gray-level variances under discrete distribution. The neighborhood space definition unit executes the spatial sampling window closed-loop update procedure according to objective physical laws, acquires the objective measurement value of gray-level variance in the initial state of the current pixel, and calls the preset structure description threshold stored in the internal register. When the gray-level variance measurement value is greater than the preset structure description threshold, it is determined that the current calculation node is located at the boundary of a high-frequency entity, and the sampling radius of the local spatial neighborhood matrix is ​​reduced by one integer pixel unit based on the current value.

[0031] At the bottom layer, boundary constraints are set, limiting the minimum usable sampling radius to a constant 1 pixel. When the grayscale variance measurement value is less than or equal to a preset structure description threshold, the current field of view is determined to be in a smooth optical background region. The sampling radius of the local spatial neighborhood matrix is ​​increased to an integer pixel unit, and a corresponding maximum sampling radius safety cutoff upper limit of 7 pixels is set. After outputting the explicit sampling radius parameters, the gradient component decoupling unit calculates the local gradient magnitude of each pixel and constructs a spatial weight mask. The weight coefficients within the spatial weight mask are... The calculation formula satisfies ,in, Here, α represents the weight coefficient for the corresponding pixel, and α is a preset edge suppression coefficient greater than 0. The local gradient magnitude corresponds to the pixel. The gradient component decoupling unit maps the local gradient magnitude to a normalized suppression factor using the aforementioned formula, limiting the weight ratio of high-gradient pixels in the convolution operation and blocking the spatial diffusion of high-frequency energy from the edge of the target object to the low-frequency path. To establish the specific numerical path for extracting the low-pass component, the preset low-pass filter operator uses an isotropic two-dimensional Gaussian kernel function for matrix modeling, and its mathematical model follows a two-dimensional normal distribution. The variance parameter of the filter kernel is calibrated according to the cutoff frequency of the microscopic light source, and the kernel order is set to be more than twice the radius of the point spread function based on the system's optical diffraction limit. During spatial domain operations, each discrete weight value of the Gaussian kernel matrix is ​​multiplied element-wise by the suppression factor at the corresponding position of the spatial weight mask, and then global normalized before being used as the dynamic convolution kernel parameter. A low-pass filter operator, combined with a spatial weight mask, calculates a weighted convolution on the original microscopic image data to determine the low-frequency background gradient component. The gradient component decoupling unit projects the local gradient magnitude onto the null space of the low-frequency background gradient component. The true projection of a one-dimensional vector onto the orthogonal complement space of the reference vector is equivalent to the original vector minus the parallel projection component in the direction of the reference vector. Based on this, a deterministic orthogonal decoupling procedure is executed. The original local gradient vector containing physical direction and magnitude information and the low-frequency background gradient component vector are received as dual-path input data. The normalized direction scalar of the low-frequency background gradient component vector is calculated. The inner product value of the original local gradient vector on the normalized direction scalar is extracted. The inner product value is multiplied by the normalized direction scalar to generate the background basis interference vector. The vector subtraction operation is performed to completely remove and subtract the background basis interference vector from the original local gradient vector.

[0032] In this physical feature stripping process, because the spatial weight mask constructed in the aforementioned steps has preemptively blocked the leakage of high-frequency structural edge energy to the background field, the extracted low-frequency background gradient components only represent pure non-stationary optical environment drift. Therefore, even if the true physical boundary orientation of the microscopic target is exactly parallel to the background gradient direction of optical drift, the subtraction operation during orthogonal projection will only deduct the low-frequency baseline drift amplitude superimposed on the edge, while the high-frequency intrinsic physical gradient caused by the abrupt change in refractive index is completely preserved because it does not participate in background modeling, thus avoiding the leakage of high-frequency intrinsic physical gradient caused by the abrupt change in refractive index from the underlying mathematical mechanism. To mitigate the risk of target edge energy being truncated to zero in a specific physical direction, a unique transformation mapping path for extracting intrinsic parameters is established through closed-loop algebraic operations. Based on this, the following operation is defined: The unit projection vector of the low-frequency background gradient component in the local gradient magnitude direction is calculated, and this unit projection vector is subtracted from the local gradient magnitude to determine the intrinsic structure gradient whose dot product with the low-frequency background gradient component is zero, thus achieving orthogonal stripping of the structural topological boundary from the background light field drift. The image reconstruction unit receives the intrinsic structure gradient and constructs a gradient with the intrinsic structure gradient based on the spatial vector direction and local geometric characteristics. A diffusion damping function positively correlated with the intensity controls the diffusion tensor to smooth grayscale along the tangent direction of the physical edge and suppress diffusion along the normal direction perpendicular to the physical edge, thus completing anisotropic diffusion reconstruction. Simultaneously, by scanning the connected components using intrinsic structure gradients, isolated gradient spots with a spatial continuous length less than 5px are eliminated. To meet the requirements of high-frequency dynamic observation, the gradient component decoupling unit adopts a parallel computing architecture based on a Field-Programmable Gate Array (FPGA) as the underlying hardware medium. The system utilizes on-chip block random access memory to construct a row buffer queue, synchronously processing discrete rows in a pipelined manner. Vector; To address data dependency conflicts in the calculation of local spatial neighborhood matrices, a multi-row overlapping caching mechanism is adopted, enabling adjacent processing cores to share boundary pixel registers without repeated memory access. In terms of single-frame, single-threaded computational overhead distribution, local gradient magnitude calculation and mask construction take about 5ms, weighted convolution and orthogonal projection decoupling take about 8ms, and anisotropic diffusion and connected component scanning take about 6ms. Through hard-wired logic, millisecond-level deterministic output is achieved. Within a computation cycle with a processing latency of no more than 20ms, the system outputs a target microscopic image that filters out background noise and maintains the sharpness of topological boundaries.

[0033] Example 2: In a high-frequency dynamic observation scenario of low-contrast micron-sized cell slices, the gray levels of the target object's edge overlap with the background, and the contrast is lower than a preset threshold. Due to the limitations of the microscope objective's point spread function and the phase delay caused by slice thickness fluctuations, a low-frequency optical drift gradient exists in a specific direction within the field of view of the original microscopic image data. Using an optical microscopy imaging device as a test platform, with a magnification of 40 and a numerical aperture of 0.65, images of micron-sized cell slice samples are collected to construct an original test dataset. Gaussian white noise with a signal-to-noise ratio of 15dB is injected into this original test dataset, and a low-frequency sinusoidal light field drift with a spatial fluctuation amplitude accounting for a preset percentage of the pixel peak gray level is superimposed as an environmental disturbance term. The edge suppression coefficient α is determined based on the engineering requirements of edge high-frequency energy isolation and low-frequency background fidelity. This coefficient is determined according to the distribution characteristics of the average gradient amplitude of pixels in the central region of the original test dataset. When the background light field fluctuation frequency is lower than the target edge span and the local gradient amplitude is lower than the target edge span, the edge suppression coefficient α is determined. As the value increases, the edge suppression coefficient α tends towards the upper limit of the set interval. Based on the aforementioned conditions, the edge suppression coefficient α of the experimental group is set to 0.8. A partially missing control group is established, and the background is extracted directly using a low-pass filter operator in the bypass spatial weight mask construction step of the control group. An out-of-range control group one is established, and its edge suppression coefficient α is set to 0.05. An out-of-range control group two is established, and its edge suppression coefficient α is set to 5.0.

[0034] The image acquisition unit takes as input raw microscopic image data containing environmental perturbation terms, and the gradient component decoupling unit calculates the local gradient magnitude of each pixel. For the local gradient magnitude at the edge of the target object For 120 high-frequency pixels, the experimental group followed the formula Calculate the weighting coefficients ,in, Here, α represents the weight coefficient for the corresponding pixel, and α is the edge suppression coefficient. The local gradient magnitude; the calculated weighting coefficients. When the value is reduced to 0.01, the energy transmission path of the edge pixel participating in the low-frequency path weighted convolution calculation is cut off. In the partial missing control group, the weight of the edge pixel remains at 1.0, resulting in a pseudo gradient vector with an amplitude of 45.2 remaining at the same pixel coordinates in the extracted low-frequency background gradient component. The gradient component decoupling unit projects the local gradient amplitude to the null space of the low-frequency background gradient component and calculates the intrinsic structure gradient. The vector direction of the intrinsic structure gradient output by the experimental group is not deflected at the physical boundary. The partial missing control group is contaminated by the pseudo gradient vector, and a diffraction halo with a width of 8.5px is generated at its physical boundary.

[0035] The fluctuation amplitude of the low-frequency sinusoidal light field drift was adjusted and set to 10%, 30%, and 50% of the pixel peak grayscale, respectively. The output results showed that when the fluctuation amplitude was 10%, the structural similarity index of the target microscopic image output by the experimental group was 0.92, and the edge localization deviation was 0.8px; when the fluctuation amplitude increased to 30%, the structural similarity index of the experimental group was 0.89, and the edge localization deviation was 1.1px; when the fluctuation amplitude reached 50%, the structural similarity index of the experimental group was 0.85, and the edge localization deviation was 1.5px. The image reconstruction unit constructed a diffusion damping function based on the spatial vector direction of the intrinsic structure gradient to maintain the topological boundary sharpness of the target microscopic image under different intensities of background drift constraints. The experimental data were compared among the groups with a low-frequency sinusoidal light field drift fluctuation amplitude of 30%. According to the data, the edge localization deviation of the experimental group was 1.1px, and the structural similarity index was 0.89. The out-of-range control group 1 was limited by insufficient high-frequency energy isolation caused by the low edge suppression coefficient α, and its structural similarity index dropped to 0.68, while the edge localization deviation increased to 4.2px. The out-of-range control group 2 had an excessively smooth low-frequency background estimation due to the high edge suppression coefficient α, and its structural similarity index was 0.71, with an edge localization deviation of 3.5px. The partially missing control group had a structural similarity index of 0.62 and an edge localization deviation of 5.8px, and its output image experienced topological boundary collapse. The experimental data showed that the spatial weight mask and the null space orthogonal projection operation constituted a physical synergy relationship, and the edge suppression coefficient α was set to 0.8 to meet the working window requirements of feature orthogonal stripping and diffusion damping reconstruction.

[0036] Example 3: In the high-frequency dynamic observation scenario of low-contrast micron-scale cell slices, there is a technical contradiction between the high-frequency noise smoothing requirement and the target topological boundary fidelity constraint after the original microscopic image data is stripped of its intrinsic structure gradient. The scalar diffusion model causes the physical boundary of the target object to be blurred when filtering out unstructured noise in the background field. The image reconstruction unit receives the intrinsic structure gradient output by the gradient component decoupling unit and triggers anisotropic diffusion reconstruction processing. The image reconstruction unit calculates the Gaussian smoothed local structure tensor in the local spatial neighborhood matrix based on the intrinsic structure gradient.

[0037] The image reconstruction unit calculates the first eigenvalue, the second eigenvalue, and a set of corresponding orthogonal eigenvectors of the local structure tensor. The first eigenvalue is set to be greater than the second eigenvalue. The image reconstruction unit determines the normal direction of the first eigenvector corresponding to the first eigenvalue, pointing towards the physical edge, and the tangent direction of the second eigenvector corresponding to the second eigenvalue, pointing towards the physical edge. Based on the aforementioned local geometric characteristics, the image reconstruction unit constructs a diffusion damping function. Using a monotonically decreasing mapping rule, it sets the first diffusion coefficient in the direction of the first eigenvector to be inversely proportional to the first eigenvalue, thus restricting the transmission of grayscale energy across the physical boundary. Simultaneously, it sets the second diffusion coefficient in the direction of the second eigenvector to a constant reference. The image reconstruction unit uses a first diffusion coefficient, a second diffusion coefficient, and a set of orthogonal eigenvectors to assemble and generate a diffusion tensor. The image reconstruction unit calculates the inner product divergence of the diffusion tensor with the current pixel gray-level gradient, updates the gray-level value of the corresponding pixel in the original microscopic image data according to a preset time step, and transforms the spatial vector direction of the intrinsic structure gradient into a quantized spatial smoothing constraint term. This controls the diffusion tensor to smooth the gray-level in the tangent direction of the physical edge and suppresses diffusion in the normal direction perpendicular to the physical edge, outputting a target microscopic image that maintains the sharpness of the target object boundary and filters out unstructured noise.

[0038] Example 4: In the deployment scenario of a dynamic observation system for a new batch of micron-level cell slices, assembly tolerances of optical imaging components cause a non-stationary background light field baseline shift. The system connects to a calibration target with a known spatial frequency distribution. The image acquisition unit acquires the response image of the calibration target under the current physical path. The neighborhood space definition unit extracts the gray-level fluctuation variance of the flat region in the response image and the pixel expansion span of the target edge. The system determines the basic variance threshold based on the gray-level fluctuation variance, determines the actual convolution radius of the point spread function of the current optical path based on the pixel expansion span, and sets the spatial receptive field scale of the low-pass filter operator to a positive integer multiple of the actual convolution radius to establish an initial physical benchmark for extracting low-frequency background gradient components. The gradient component decoupling unit calculates the theoretical gradient magnitude and the measured local gradient magnitude of the calibration target edge region, and sets the ratio of the two as the attenuation benchmark factor. The system sets the edge suppression coefficient α as the product of the basic variance threshold and the attenuation benchmark factor. When processing the original microscopic image data, the gradient component decoupling unit determines the edge suppression coefficient α and the local gradient magnitude of each pixel. Construct a spatial weight mask based on the formula Constraint weight coefficient The attenuation slope blocks the spatial leakage of high-frequency energy to the low-frequency path at the physical boundary of the target; the timing update of the gain parameter depends on the linear product of the current physical output error and the fixed response coefficient, and the built-in online adaptive calibration procedure smooths out the light source attenuation fluctuations caused by long-term continuous microscopic observation.

[0039] To address the common power supply ripple phenomenon in AC grid-driven light sources, the system pre-samples and performs discrete Fourier transform on the light intensity of the light source using a high-frequency photodetector. Measured spectral data confirms the existence of transient alternating light field fluctuations in the background field, with a fundamental frequency of 100Hz and its integer multiples of higher harmonics. To satisfy the Nyquist-Shannon sampling theorem and avoid spectral aliasing, the hardware sampling frame rate of the image acquisition unit is configured to be at least 200fps (oversampling). The gradient component decoupling unit utilizes a circular buffer containing multiple frames of historical extraction results to record low-frequency background frequencies. The historical distribution characteristics of the field are used as the basis for weighted correction by superimposing the historical characteristics with the current extraction results and taking the average. This effectively filters out the high-frequency alternating physical environment interference in the digital domain. A preset edge sharpness target value is defined in the storage medium, representing the expected gradient amplitude when the standard reference target is in the ideal focal plane. Using grayscale per pixel physical dimension, the measured edge sharpness value of the current frame output target microscopic image is synchronously acquired and calculated, with the dimension consistent with the expected value. The preset edge sharpness target value is calculated by subtracting the algebraic difference of the measured edge sharpness value, and the algebraic difference is multiplied by the dimensionless feedback. The proportional gain constant, the output product result is defined as the indicator adjustment amplitude transformation step size. The transformation step size is directly accumulated into the original value of the edge suppression coefficient α extracted in the current frame to generate the updated edge suppression coefficient for the next frame. The value range of the updated edge suppression coefficient is forcibly truncated in the software layer so that it always falls within the floating-point range of greater than 0.1 and less than 5.0. This closed-loop feedback link is constructed based on the temporal coherence assumption in high-frequency continuous microscopic observation, that is, the relative displacement or morphological change of physical samples between adjacent frames is extremely small, far lower than the frame rate sampling period of the sensor. The system does not use the reshaping result caused by diffusion damping distortion in the current frame to correct the initial parameters of the current frame. Instead, it uses the global structural features after reconstruction as temporal prior knowledge and feeds them forward to the background mask construction stage of the input end of the next calculation cycle through a ring delay feedback network to achieve temporal isolation between the slow drift of the optical system and the high-frequency reshaping features. After completing the parameter iteration through the calculation and assignment path, the image reconstruction unit constructs the diffusion damping function according to the intrinsic structure gradient and outputs the target microscopic image that filters out background noise and maintains the sharpness of the target object boundary.

[0040] Example 5: In a multi-batch micrometer-scale cell slice observation system deployment scenario, the physical condition of cross-domain drift of local structure tensor eigenvalue scale caused by differences in optical sensor response sensitivity and fluctuations in sample staining baseline; the system introduces a basic test map set containing step grayscale boundaries and flat background areas to establish a standardized parameter baseline model, and the image reconstruction unit calculates the first eigenvalue of the local structure tensor corresponding to each pixel in the basic test map set. Extract the first feature value of the flat background area in the basic test image set. The upper limit of variance is set as the singular value truncation threshold ϵ, and the first feature value at the step grayscale boundary is extracted. The distribution mean, combined with the dynamic response range of the optical sensor, is set as a scale normalization constant K; the image reconstruction unit constructs the first diffusion coefficient. The quantization mapping rules, and the specific calculation formulas satisfy... ,in, The first diffusion coefficient is the first eigenvector direction, and K is the scale normalization constant. Let K be the first eigenvalue, and ϵ be the singular value truncation threshold. The dimensions of K and ϵ are the same as those of the first eigenvalue. To maintain consistency, the image reconstruction unit uses quantization mapping rules to limit the diffusion damping attenuation path in high gradient regions and establishes a spatial smoothing constraint matrix that matches the current hardware response baseline.

[0041] The image reconstruction unit addresses sensor saturation and edge abrupt changes based on the spatial smoothing constraint matrix, when it detects the first eigenvalue of a local pixel in the original microscopic image data. When the value is below the singularity truncation threshold ϵ, the image reconstruction unit determines that the corresponding physical space is a flat feature or a sensor blind spot, and sets the first diffusion coefficient. The baseline value is constant, maintaining the isotropic grayscale flow state within the region; when the first characteristic value is detected... When the value exceeds a preset multiple of the scale normalization constant K, the preset multiple is set to 3 based on the full-well capacity characteristics of the optical detector. The system determines that the corresponding optical sensing node has experienced local saturation overexposure, and the image reconstruction unit adjusts the first diffusion coefficient. The system is reset to zero, blocking the data diffusion path from the overexposed pixel to the neighborhood space; the system converts the discrete optical hardware response parameters into a closed-loop quantization constraint model, and outputs a target microscopic image that meets the physical boundary sharpness conditions.

[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A microscopic image intelligent processing system, characterized in that, include: Image acquisition unit, used to acquire raw microscopic image data; The neighborhood space definition unit is used to determine the local spatial neighborhood matrix of each pixel in the original microscopic image data, providing topological support for subsequent tensor field modeling. The gradient component decoupling unit includes the following steps: Step S11, calculating the local gradient magnitude of each pixel; Step S12, constructing a spatial weight mask based on the local gradient magnitude, wherein the weight coefficient of each pixel in the spatial weight mask decreases as the local gradient magnitude increases, in order to suppress the weight contribution of high-frequency edge components in subsequent background estimation; Step S13, performing convolution operation on the original microscopic image data using a preset low-pass filter operator combined with the spatial weight mask, and determining the low-frequency background gradient component by shielding the spatial diffusion of structural edge energy into the low-frequency background path; Step S14, projecting the local gradient magnitude onto the null space of the low-frequency background gradient component, and determining the intrinsic structure gradient by stripping the gradient interference component in the same direction as the background field, wherein the dot product of the intrinsic structure gradient and the low-frequency background gradient component is 0. The image reconstruction unit is used to perform anisotropic diffusion reconstruction on the original microscopic image data based on the spatial vector direction of the intrinsic structure gradient and its local geometric characteristics. By controlling the diffusion tensor in the tangent direction of the physical edge for grayscale smoothing and implementing diffusion suppression in the normal direction perpendicular to the physical edge, the unit maintains the boundary sharpness of the target object while filtering out unstructured noise in the background field and outputs the target microscopic image.

2. The intelligent microscopic image processing system according to claim 1, characterized in that, When constructing the spatial weight mask, the gradient component decoupling unit maps the local gradient magnitude to a normalized suppression factor and limits the contribution ratio of high gradient pixels in the convolution operation through the spatial weight mask. This prevents the edge energy of the target object from leaking into the low-frequency background gradient component, thereby eliminating the pseudo-background component generated at the structural boundary due to the point spread function limitation. This ensures that the spatial vector direction of the intrinsic structural gradient is consistent with the physical edge contour of the target object.

3. The intelligent microscopic image processing system according to claim 1, characterized in that, When the gradient component decoupling unit completes step S14, it calculates the unit projection vector of the low-frequency background gradient component in the direction of the local gradient magnitude, and subtracts the unit projection vector from the local gradient magnitude to achieve nonlinear separation of the target intrinsic features and background field features, ensuring that the intrinsic structure gradient only characterizes the intrinsic geometric features of the target object.

4. The intelligent microscopic image processing system according to claim 1, characterized in that, When the image reconstruction unit completes anisotropic diffusion reconstruction, it constructs a diffusion damping function that is positively correlated with the intrinsic structure gradient intensity. This function controls the diffusion flow to complete grayscale smoothing in the tangential direction of the target object's edge and stops the diffusion motion in the direction perpendicular to the edge, thereby repairing the boundary blur caused by imaging system noise.

5. The intelligent microscopic image processing system according to claim 1, characterized in that, When determining the local neighborhood matrix, the neighborhood definition unit calculates the gray-level variance within the current pixel window area. When the gray-level variance is greater than the preset structure description threshold, the sampling radius of the local neighborhood matrix is ​​reduced. When the gray-level variance is less than the structure description threshold, the sampling radius of the local neighborhood matrix is ​​increased to achieve adaptive coverage of targets at different scales.

6. The intelligent microscopic image processing system according to claim 1, characterized in that, The image reconstruction unit is also used to identify unstructured interference in the original microscopic image data. By scanning the intrinsic structure gradient, it identifies and removes isolated gradient spots with a spatial continuity length of less than 5px, thereby automatically filtering interference from coverslip reflection spots or sample bubbles, ensuring that the output target microscopic image eliminates artifacts while maintaining edge recall.

7. The intelligent microscopic image processing system according to claim 1, characterized in that, The gradient component decoupling unit adopts a parallel computing architecture to synchronously process the discrete row vectors of the original microscopic image data, so that the processing delay of a single frame image is no more than 20ms, thereby realizing real-time high-frequency dynamic monitoring of pathological samples.

8. The intelligent microscopic image processing system according to claim 1, characterized in that, The intelligent microscopic image processing system also includes a calibration feedback unit, which monitors the edge sharpness of the target microscopic image in real time and linearly adjusts the transformation step size of the weight coefficient in step S12 according to the deviation value of the edge sharpness relative to the preset target, so as to achieve adaptive gain for sample images with different staining depths.

9. The intelligent microscopic image processing system according to claim 1, characterized in that, The intelligent microscopic image processing system also includes a storage unit for recording the historical distribution characteristics of the low-frequency background field; in step S13, the gradient component decoupling unit uses the historical distribution characteristics and the pixel-level mean of the current frame extraction result to perform weighted correction on the low-frequency background gradient component in order to eliminate transient fluctuation interference of the light source system at frequencies above 100Hz.

Images