An adaptive non-uniform depth of field extension microscopic imaging method and system

Abstract

The application provides a self-adaptive non-uniform depth-of-field extension microscopic imaging method and system, and belongs to the technical field of optical microscopic imaging and digital image processing. A low numerical aperture illumination is used to shoot an out-of-focus original bottom map at an extremely sparse fixed height, and a high-precision full-view focusing surface map is quickly reconstructed. Then, based on the gradient fluctuation characteristics of the topographic map, the microscopic field of view is dynamically divided into layers, a small longitudinal depth adaptive acquisition and a single frame direct output are configured in the non-fluctuation area, and a high-density large longitudinal depth fusion is configured in the significant fluctuation area. The method not only eliminates the axial redundancy of full scanning, but also introduces a structure tensor and a phase-preserving complex wavelet into the fusion mechanism, so that the imaging space-time efficiency and high-frequency information reconstruction are greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of optical microscopy and digital image processing, specifically to an adaptive non-uniform depth-of-field extended microscopy imaging method and system. Background Technology

[0002] Microscopic imaging systems play an indispensable role in modern life sciences, precision semiconductor testing, and materials science. However, due to the diffraction limit, traditional optical microscopes have an irreconcilable physical trade-off between lateral resolution and axial depth of field (DOF). When a system is equipped with high numerical aperture (High NA) objectives to obtain high-resolution images at the sub-micron or even nanometer scale, its DOF ​​is typically drastically compressed to the level of a few micrometers or even hundreds of nanometers. When dealing with thick samples with significant surface undulations, three-dimensional microstructures, or multilayered cells, a single image can only clearly image a very narrow section, while the remaining areas appear extremely blurry due to severe defocusing.

[0003] To overcome this physical limitation, researchers have proposed Extended Depth of Field (EDoF) or all-focus image fusion techniques. Wavefront coding, in particular, introduces a specialized phase mask (such as a cubic phase plate) on the objective lens pupil surface, enabling the system's point spread function (PSF) to exhibit depth invariance over a wide axial range. However, due to the extremely low Strehl ratio of wavefront coding systems, the original acquired images are extremely blurry, necessitating complex inverse filtering or regularized deconvolution algorithms for image restoration. This digital post-processing is highly sensitive to noise, easily loses high-frequency microscopic details, and consumes enormous computational resources, making it difficult to meet the demands of high-throughput, rapid imaging.

[0004] Another mainstream approach is optical focus stacking or Z-stack focusing fusion technology. This method controls the microscope stage or objective lens to perform high-density scanning along the axis with extremely small, equidistant steps (typically less than or equal to the physical depth of field at high NA), acquiring a complete sequence of high-resolution images. Subsequently, specific image fusion algorithms (such as wavelet transform, Laplacian pyramid, etc.) are used to evaluate the focus measures of each region of the image at each axial height, extracting the sharpest pixels, patches, or regions, and finally stitching them together to form a full-focus image. Although this method can retain extremely high lateral resolution, it exposes serious efficiency bottlenecks in practical applications: 1. Extremely high spatiotemporal data redundancy: In most flat areas with small surface undulations, high-density axial full-axis scanning has acquired a large number of completely overlapping blurry frames, which not only does not help improve the clarity, but also causes serious waste of computation and storage. 2. Photobleaching and phototoxicity are serious: long-term, high-frequency repeated excitation irradiation can easily accelerate the quenching of fluorescent samples and cause irreversible physiological damage to living biological samples. 3. Mechanical return error and springback artifact: The mechanical displacement stage is prone to step hysteresis and return error during frequent start and stop of micro-steps, which will introduce edge ghosting and asymmetric distortion during splicing and fusion.

[0005] Chinese invention patent CN115375550B proposes a multi-layer microscopic imaging method combining low-NA and high-NA light sources. This patent points out that due to the low depth of field of high-NA light sources, when two layers overlap in the transition region, the focus measurement curve exhibits multi-peak characteristics, leading to random jumps and distortions in the focus plane determination result at the boundary. To address this, the patent proposes utilizing the flat, single-peak characteristic of the focus measurement curve under large depth of field with low-NA light sources to detect the parfocal plane, and then using the large depth-of-field image formed by the low-NA light source at the boundary for fusion, thereby accurately recording the layer boundaries in the image quality.

[0006] Although the patent has achieved significant technical results in suppressing focus jumps and edge artifacts, its system architecture and execution logic are still based on "full-axis equal-step scanning". The patent fails to take advantage of the macroscopic and rapid perception of spatial depth representation by low-NA light sources, and it is also unable to perform adaptive region division for the spatial undulation morphology of the sample. Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive non-uniform depth-of-field extended microscopic imaging method, which divides the imaging process into three core stages: macroscopic low-aperture morphology perception, hierarchical adaptive resampling region division, and multi-scale phase-preserving seamless fusion. This not only eliminates the axial redundancy of the full scan, but also introduces structural tensors and phase-preserving complex wavelets into the fusion mechanism, achieving a leapfrog improvement in imaging spatiotemporal efficiency and high-frequency information reconstruction.

[0008] To achieve the above objectives, the present invention employs the following technical solutions.

[0009] An adaptive non-uniform depth-of-field extended microscopic imaging method includes: S1. Aperture-encoded sparse topography pre-acquisition: Switch the microscope system to low numerical aperture illumination mode, acquire multiple sets of sparse height defocused images along the axis, and obtain a low NA image sequence. S2. Global Focus Topography Reconstruction: Based on low NA image sequences, the focus measurement of each region is calculated, and the focus height of the entire field of view is solved through an interpolation model to generate a focus surface topography map. S3. Field of view undulation region division: Based on the focal plane topography map, calculate the spatial undulation feature index, and combine it with a preset threshold to divide the imaging field of view into flat feature region and undulating significant region. S4, High NA Adaptive Differential Sampling: Switch the microscope system to high numerical aperture imaging mode and perform axial non-uniform image acquisition in different regions; in flat areas, directly locate to the height corresponding to the topography map to acquire clear images, and in undulating areas, perform Z-stack sequence acquisition with the topography height as the center. S5. Construction of low-NA global reference map: Synthesize a low-NA full-focus reference image with continuous depth-of-field features using a low-NA off-focus image sequence; S6. Image multi-scale decomposition and focus weight extraction: Decompose all high NA acquired images to obtain a base layer that retains low-frequency topology and a detail layer that retains high-frequency texture, and generate an initial detail weight map based on local structural features. S7. Complex wavelet domain high-frequency detail fusion: Complex wavelet transform is used to complete multi-frame detail fusion for the detail layer corresponding to the undulating region, and the fusion jump is smoothly eliminated by consistency check to reconstruct the detail layer; S8, Seamless Blending and Color Correction: Using a low-NA full-focus reference image as a guide, smooth the blending weight boundaries, and then perform pixel remapping on the multi-channel image to output a complete large depth-of-field microscopic image.

[0010] Furthermore, in step S1, the number of axial sparse heights M≥3, the adjacent sampling interval of low NA is equal to the physical depth of field of the low numerical aperture system, and the sampling interval is 10 times or more of the high NA imaging scan step size.

[0011] Furthermore, in step S2, a three-point logarithmic Gaussian peak interpolation model is used to solve for the focus height. The interpolation model formula is: , in, The axial height value of the low-NA (non-polarization) sensor corresponds to the maximum absolute response of the focus measurement, while FM represents the local focus measurement. For low NA axial sampling interval, , and respectively using coordinates Within the local window centered on the element, at three consecutive large step heights , and The focus measurement value obtained from the calculation is as follows. coordinates The height of the focal plane.

[0012] Furthermore, in step S3, the spatial fluctuation characteristic index is the geometric fluctuation index. The calculation formula is: , In the formula, Let D be the local isotropic gradient vector of the focal plane topography map, and let D be the local statistical window. This represents the average height within a local statistical window. To adjust the balance between the weighting of the undulation gradient and the absolute elevation difference; Set flatness threshold ,when When this local field of view is determined to be a flat feature region, a small depth sampling parameter is assigned; when At that time, the local field of view is determined to be a region with significant undulations, and large depth sampling parameters are assigned.

[0013] Furthermore, in step S4, the axial focal plane adjustment in high NA mode is achieved by an electrically controlled liquid lens, and the optical path focal plane is adjusted by changing the input current of the electrically controlled liquid lens.

[0014] Furthermore, in step S6, Gaussian filtering is applied to the high NA image to complete dual-scale decomposition and obtain the base layer. With detail layer : , In the formula, axial height High NA local slice image, For pixel coordinates, Let m be the height of the focal plane along the axial direction. It is a Gaussian kernel with integral action; For detail layer Constructing the local two-dimensional structure tensor matrix: , In the formula, For detail layer abbreviation; The maximum eigenvalue of the structure tensor is used as the basis for sharpness evaluation to generate an initial detail weight map.

[0015] Furthermore, step S7 specifically includes: S71. A complex wavelet transform is performed on the detail layer using a complex Debach third-order orthogonal complex wavelet basis to obtain multi-directional complex wavelet subband coefficients. S72. Select the coefficient corresponding to the largest wavelet coefficient modulus as the fusion coefficient to generate the focus layer decision location map. ; S73, to When performing sub-band consistency verification, if the merging coefficients of two sub-bands in the three directions of the same pixel are mapped to the same acquisition height, the third sub-band is forcibly replaced with the complex wavelet coefficients corresponding to that height. S74, to A 3×3 spatial majority voting smoothing mechanism is used to smooth the height decision graph; S75. Perform inverse complex wavelet transform on the corrected complex wavelet coefficients to reconstruct the high-frequency fusion detail layer.

[0016] Furthermore, in step S8, guided by the low NA full-focus reference image, a guided filtering algorithm is used to smooth the stitching weight map between the single-frame image of the flat feature region and the reconstructed image of the detail layer of the undulating region, in order to eliminate the step artifacts in the stitching of the flat and undulating regions.

[0017] Furthermore, the multi-channel pixel remapping in step S8 satisfies: , In the formula, For RGB color channels, To output color channel pixel values, Axial position The pixel value of the k-th color channel in the slice image. To theoretically fuse the brightness of grayscale images, This is a local grayscale projection of the multifocal image sequence actually acquired under high NA conditions, with all output pixels taken from the true optical response of the original acquired image.

[0018] An adaptive non-uniform depth-of-field extended microscopic imaging system includes: A multi-numerical aperture time-controlled illumination unit is used to switch between outputting low numerical aperture and high numerical aperture illumination beams. An electronically controlled liquid lens focusing unit is located in the imaging optical path and rapidly changes the axial imaging focal plane of the system via electrical signals. The sample carrier displacement stage is used to carry three-dimensional undulating samples and achieve precise axial and lateral displacement. Image acquisition sensor, used to acquire low-NA defocus images and high-NA axial slice images; The main control computing unit is electrically connected to the multi-numerical aperture time-controlled illumination unit, the electronically controlled liquid lens focusing unit, the image acquisition sensor, and the sample bearing displacement stage. The main control computing unit stores and runs the equipment control program and image processing algorithm of the above imaging method, and outputs high-resolution large depth-of-field fused microscopic images.

[0019] The advantages of this invention are: The full-field focusing topography map is reconstructed by sparse axial sampling in low numerical aperture mode. The field of view is partitioned according to the surface undulation characteristics of the sample. Then, a differentiated high numerical aperture image acquisition strategy is adopted. Compared with the traditional global uniform axial scanning scheme, it can avoid a large number of redundant image acquisition operations in flat areas, effectively reduce the duration of light irradiation on the sample, alleviate the adverse effects of continuous illumination, and reduce the resource overhead of image data storage and subsequent image processing. In the high numerical aperture stage, layered imaging is achieved by adjusting the axial focal plane, which eliminates the need to frequently drive the mechanical structure carrying the sample or objective lens to move back and forth, thus avoiding imaging distortion problems caused by factors such as hysteresis and backlash error during mechanical displacement. Abandoning the traditional spatial domain pixel hard stitching rules, a cross-scale consistency seamless fusion algorithm was designed, combining dual-scale structural tensor feature extraction with phase-preserving complex two-dimensional wavelet transform. This algorithm utilizes structural tensors to perform robustness measurement of microscopic details at the boundaries, performs spatial and sub-band consistency verification in the complex wavelet domain using phase stability, and uses low-NA full-focus images as guidance to perform edge-preserving smoothing on the fusion weight distribution, eliminating step edge artifacts and synthesizing an ultra-high resolution and physically continuous virtual large depth-of-field image. By introducing a multi-channel pixel remapping mechanism, the final output color pixels are all selected from the real pixels with the highest matching degree in the original high NA image, rather than relying on pixel weighting to generate virtual pixels. This can avoid color distortion problems such as spectral shift and false colors during multi-channel fusion, and ensure the authenticity of color information in the output image. It can adaptively adjust the acquisition strategy according to the morphological fluctuations of the sample itself, and can adapt to three-dimensional observation samples with different surface morphological features, thus having good scene adaptability. Attached Figure Description

[0020] Figure 1 This is a flowchart of the adaptive non-uniform depth-of-field extended microscopic imaging method of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] This embodiment discloses an adaptive non-uniform depth-of-field extended microscopic imaging method. Please refer to the flowchart below. Figure 1 .

[0023] S1, Aperture-coded sparse topography pre-collection The system first uses aperture time-varying control to quickly switch the lighting source to a low aperture mode, at which point the physical depth of field of the system is significantly expanded.

[0024] Set the system's low NA axial step span to Satisfying the relation: In the formula, The axial step size required for high-NA multilayer fusion is typically , For physical depth of field, The scaling factor is used to adjust the sampling step size under low NA conditions. This achieves sampling step distances 10 to 50 times higher than high NA (Non-Analog) sampling step distances. The system controls the actuators at a fixed and large-span axial height. (Where k = 1, 2, ..., M, and M = 3 or 4 at extremely low frame rates) Full-view shooting is performed, without needing to focus the samples, directly capturing and obtaining an extremely sparse low-NA out-of-focus base map sequence. .

[0025] S2, Global Focus Topography Reconstruction Taking advantage of the highly regular, single-peak Gaussian distribution characteristic of the focus measurement curve under low NA (no multi-peak interference), this algorithm can resolve a high-precision focus surface topography map covering the entire field of view within milliseconds using an efficient three-point logarithmic Gaussian interpolation mathematical model. .

[0026] Based on physical and geometric optics models, under low-NA, large-depth-of-field illumination, the attenuation process of high-frequency edges and micro-textures can be approximated as a convolution process of the Gaussian point spread function (PSF), whose focus measure function... The axial variation curve exhibits excellent single-peak symmetry and a wide envelope, without generating multi-peak disturbances. (Define arbitrary spatial coordinates) The local focus measure function at point is: In the formula, For A local normalized statistical window centered on the center, This is either a Laplacian operator or a first-order high-pass wavelet operator. In this case, the focus metric curve is at the actual physical focal plane height. The vicinity can be accurately modeled as a Gaussian envelope model: In the formula, To focus on the peak value of the measurement center, To achieve effective physical depth of field at low numerical apertures Strongly correlated constants.

[0027] To calculate the true continuous focus plane with sub-micron precision under extremely sparse Z-axis sampling, a three-point logarithmic Gaussian peak interpolation model is established. For arbitrary pixel coordinates First, the height point with the highest measured response in the entire sequence is retrieved. and its adjacent points and The corresponding measure function values ​​are respectively , and .

[0028] Since the logarithmic form of the Gaussian function forms a quadratic parabola: make Then there are three measurement equations with the same height difference. By solving symmetric differences and the vertex of the parabola, a continuous focusing surface height map with sub-micron precision can be directly derived analytically. : S3, Field of View Undulation Region Division Define the spatial first-order partial derivative gradient vector of the three-dimensional surface topography as: To effectively suppress minute high-frequency disturbances and computational noise introduced by interpolation, the system employs an isotropic guided filter to smooth the topography map while preserving its edges. The system calculates the local spatial gradient and variance of the topography map at the focal plane, quantitatively assessing the geometric fluctuation index of the current field of view. Its expression is a linear weighted function of the spatial mean of the local topographic gradient magnitude and the physical standard deviation of the height in that region: In the formula, This represents the average height within the local patch D. This is a normalization coefficient used to balance the fluctuations in slope and absolute elevation.

[0029] The system sets a flatness judgment threshold. This threshold is related to the theoretical physical depth of field under high NA. Related, usually set .

[0030] Based on the magnitude of the fluctuation index, the microscopic field of view is divided into flat feature regions and significantly undulating feature regions: Flat feature region The surface undulations in this area are minimal, falling entirely within a single high-NA depth-of-field section; Steep feature areas, The surface of this area exhibits a distinct sloping gradient, folded microstructures, or multiple layers of suspended cells.

[0031] S4, High NA Adaptive Differential Sampling The system performs parameterized physical acquisition with highly non-uniform parameters: For flat areas, the system assumes that the samples in these areas have no height undulations and configures small-depth adaptive fusion parameters. In this region, the system controls the adjustable lens (ETL) to directly jump one step at a time, only... Single frames are captured at high altitudes (or only within an extremely narrow envelope to cope with environmental disturbances). Two high-NA images were captured internally, minimizing exposure load.

[0032] For regions with significant undulations, the system is configured with large-depth adaptive fusion parameters. In this region, the system's height at the topographic center is determined. A large-scale focused stack envelope is allocated around the perimeter. To determine the absolute elevation difference of local fluctuations, and to control the electro-hydraulic lens to perform micro-stepping (step distance) within this depth. High-density Z-stack acquisition ensures that all depth information is accurately captured.

[0033] To further quantify this level of adaptive sampling process, the table below shows the non-uniform resampling parameter configurations performed by the system based on the local fluctuation index.

[0034] Table 1 System uniform resampling parameter configuration S5, Low NA Global Reference Graph Construction High-frequency components are extracted from the low-NA base map sequence acquired in step S1 to synthesize a large depth-of-field macroscopic low-NA full-focus reference image. .

[0035] S6. Image Multiscale Decomposition and Focus Weight Extraction The acquired high NA local slice sequence is denoted as First, the algorithm employs an isotropic Gaussian filter to perform two-scale decomposition, dividing the image into smooth, low-frequency contour-containing base layers. With detail layers containing rich texture details : For details To make it insensitive to high-frequency image noise, the focus intensity evaluation no longer uses Laplacian high-pass filtering, but instead establishes a local two-dimensional structure tensor matrix J: In the formula, It is a Gaussian kernel with integral action. The structure tensor is solved... Extract the largest eigenvalue from the eigenvalues. As a robust measure of pixel focus intensity, it generates an initial detail weight map corresponding to high-NA images. ; S7, Complex Wavelet Domain High-Frequency Detail Fusion Complex wavelet transform is used to fuse multi-frame details in the detail layers corresponding to the undulating regions, and consistency checks are used to smooth out fusion jumps and reconstruct the detail layers. Perform CWT decomposition on the detail layer image to obtain complex wavelet coefficients. ,in Indicates the decomposition scale. These represent sub-bands in different directions (horizontal, vertical, diagonal). Complex coefficients can be expressed as amplitude and phase: Because the maximum amplitude corresponds to the sharpest and highest contrast local physical micro-features in the image, the maximum absolute value fusion rule is used to synthesize the high-frequency coefficients of complex wavelets: To suppress misjudgments caused by isolated noise, subband consistency checks and spatial consistency checks are performed in the complex wavelet domain: at any spatial location, if the coefficients of two subbands in the three directions originate from the same physical height... Then, the corresponding coefficients of the third sub-band will be forcibly corrected to those from that layer. The complex wavelet coefficients; simultaneously, the decision location coordinate diagram. conduct The neighborhood majority voting mechanism is smooth, avoiding spatial discontinuities. This ensures that the fused complex wavelet coefficients exhibit phase-fidelity properties. Perform inverse complex wavelet transform (ICWT) to reconstruct the high-frequency detail reconstruction layer. S8, seamless edge blending and color correction Since the non-undulating regions only have one or two high-NA images, while the undulating regions have fused high-frequency details reconstructed through CWT. Seamless integration of the two is essential. In Phase 1, a completely continuous, artifact-free low-NA macroscopic full-focus image is synthesized using a sparse low-NA image sequence and fast frequency domain maximum filtering. Although the image has a low physical resolution, it contains fully realistic macroscopic brightness topology and layer boundary gradient features.

[0036] Therefore, this invention utilizes this macroscopic low-NA full-focus image. As a guide image, a guide filter is applied to smooth the edge distribution of the stitching weight boundary between the single-frame image in the flat region and the fusion detail layer in the undulating region under high NA. Since the guide image has the characteristic of continuous large depth of field in the transition region, the guide filter forces the boundary line of the weight map to align with the physical structure of the low NA macro full-focus image, eliminating the hard seams formed in the transition zone due to non-uniform physical scanning, and synthesizing an ultra-high resolution and physically continuous virtual large depth of field image.

[0037] To completely avoid non-convexity errors and false color and pseudo-color biases in multi-channel color imaging after fusion, this algorithm introduces a multi-channel pixel reassignment algorithm when reconstructing the color full-focus image. For the final synthesized color vector image in channels... pixel value at Instead of simply performing a weighted summation, it replaces the actual acquired pixel values ​​from the original high-NA multifocal image stack that are closest to the theoretically predicted brightness by fusion with the new values. This mechanism forces all three-dimensional spatial point data of the final full-focus image to be strictly derived from the physical response actually recorded by the microscopic system, avoiding computational overflow and spectral distortion, and ensuring the high spectral and color fidelity required for pathological analysis or semiconductor measurements.

[0038] The table below compares in detail the differences in system control characteristics between the method of the present invention and the traditional equidistant focusing scanning microscope.

[0039] Table 2. Differences in system control characteristics between the method of the present invention and the traditional equidistant focusing scanning microscope. The table below compares the comprehensive performance indicators of the method of the present invention with those of the traditional full-axis equal-step scanning fusion method and the improved parent patent technology targeted by the present invention.

[0040] Table 3 Comparison of Comprehensive Performance Indicators The method of this invention achieves decisive performance advantages in all core dimensions, including the number of axial movements, high-frequency PSNR performance at image edges, color fidelity, and phototoxicity suppression.

[0041] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive non-uniform depth-of-field extended microscopic imaging method, characterized in that, include: S1. Aperture-encoded sparse topography pre-acquisition: Switch the microscope system to low numerical aperture illumination mode, acquire multiple sets of sparse height defocused images along the axis, and obtain a low NA image sequence. S2. Global Focus Topography Reconstruction: Based on low NA image sequences, the focus measurement of each region is calculated, and the focus height of the entire field of view is solved through an interpolation model to generate a focus surface topography map. S3. Field of view undulation region division: Based on the focal plane topography map, calculate the spatial undulation feature index, and combine it with a preset threshold to divide the imaging field of view into flat feature region and undulating significant region. S4, High NA Adaptive Differential Sampling: Switch the microscope system to high numerical aperture imaging mode and perform axial non-uniform image acquisition in different regions; in flat areas, directly locate to the height corresponding to the topography map to acquire clear images, and in undulating areas, perform Z-stack sequence acquisition with the topography height as the center. S5. Construction of low-NA global reference map: Synthesize a low-NA full-focus reference image with continuous depth-of-field features using a low-NA off-focus image sequence; S6. Image multi-scale decomposition and focus weight extraction: Decompose all high NA acquired images to obtain a base layer that retains low-frequency topology and a detail layer that retains high-frequency texture, and generate an initial detail weight map based on local structural features. S7. Complex wavelet domain high-frequency detail fusion: Complex wavelet transform is used to complete multi-frame detail fusion for the detail layer corresponding to the undulating region, and the fusion jump is smoothly eliminated by consistency check to reconstruct the detail layer; S8, Seamless Blending and Color Correction: Using a low-NA full-focus reference image as a guide, smooth the blending weight boundaries, and then perform pixel remapping on the multi-channel image to output a complete large depth-of-field microscopic image.

2. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S1, the number of axial sparse heights M≥3, the adjacent sampling interval of low NA is equal to the physical depth of field of the low numerical aperture system, and the sampling interval is 10 times or more of the high NA imaging scan step size.

3. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S2, a three-point logarithmic Gaussian peak interpolation model is used to solve for the focus height. The interpolation model formula is: ， in, The axial height value of the low-NA (non-polarization) sensor corresponds to the maximum absolute response of the focus measurement, while FM represents the local focus measurement. For low NA axial sampling interval, , and respectively using coordinates Within the local window centered on the element, at three consecutive large step heights , and The focus measurement value obtained from the calculation is as follows. coordinates The height of the focal plane.

4. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S3, the spatial fluctuation characteristic index is the geometric fluctuation index. The calculation formula is: ， In the formula, Let D be the local isotropic gradient vector of the focal plane topography map, and let D be the local statistical window. This represents the average height within a local statistical window. To adjust the balance between the weighting of the undulation gradient and the absolute elevation difference; Set flatness threshold ,when When this local field of view is determined to be a flat feature region, a small depth sampling parameter is assigned; when At that time, the local field of view is determined to be a region with significant undulations, and large depth sampling parameters are assigned.

5. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S4, the axial focal plane adjustment in high NA mode is achieved by an electrically controlled liquid lens, and the optical path focal plane is adjusted by changing the input current of the electrically controlled liquid lens.

6. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S6, Gaussian filtering is applied to the high NA image to complete dual-scale decomposition and obtain the base layer. With detail layer : ， In the formula, axial height High NA local slice image, For pixel coordinates, Let m be the height of the focal plane along the axial direction. It is a Gaussian kernel with integral action; For detail layer Constructing the local two-dimensional structure tensor matrix: ， In the formula, For detail layer abbreviation; The maximum eigenvalue of the structure tensor is used as the basis for sharpness evaluation to generate an initial detail weight map.

7. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, Step S7 specifically includes: S71. A complex wavelet transform is performed on the detail layer using a complex Debach third-order orthogonal complex wavelet basis to obtain multi-directional complex wavelet subband coefficients. S72. Select the coefficient corresponding to the largest wavelet coefficient modulus as the fusion coefficient to generate the focus layer decision location map. ; S73, to When performing sub-band consistency verification, if the merging coefficients of two sub-bands in the three directions of the same pixel are mapped to the same acquisition height, the third sub-band is forcibly replaced with the complex wavelet coefficients corresponding to that height. S74, to A 3×3 spatial majority voting smoothing mechanism is used to smooth the height decision graph; S75. Perform inverse complex wavelet transform on the corrected complex wavelet coefficients to reconstruct the high-frequency fusion detail layer.

8. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S8, guided by the low NA full-focus reference image, a guided filtering algorithm is used to smooth the stitching weight map between the single-frame image of the flat feature region and the detailed layer reconstruction image of the undulating region, in order to eliminate the step artifacts in the stitching of the flat and undulating regions.

9. The adaptive non-uniform depth-of-field extended microscopic imaging method according to claim 1, characterized in that, In step S8, the multi-channel pixel remapping satisfies: ， In the formula, For RGB color channels, To output color channel pixel values, Axial position The pixel value of the k-th color channel in the slice image. To theoretically fuse the brightness of grayscale images, This is a local grayscale projection of the multifocal image sequence actually acquired under high NA conditions, with all output pixels taken from the true optical response of the original acquired image.

10. An adaptive non-uniform depth-of-field extended microscopic imaging system, characterized in that, include: A multi-numerical aperture time-controlled illumination unit is used to switch between outputting low numerical aperture and high numerical aperture illumination beams. An electronically controlled liquid lens focusing unit is located in the imaging optical path and rapidly changes the axial imaging focal plane of the system via electrical signals. The sample carrier displacement stage is used to carry three-dimensional undulating samples and achieve precise axial and lateral displacement. Image acquisition sensor, used to acquire low-NA defocus images and high-NA axial slice images; The main control computing unit is electrically connected to the multi-numerical aperture time-controlled illumination unit, the electrically controlled liquid lens focusing unit, the image acquisition sensor, and the sample bearing displacement stage. The main control computing unit stores and runs the device control program and image processing algorithm of any of the imaging methods described in claims 1 to 9, and outputs a high-resolution, large-depth-of-field fused microscopic image.

Images