Microscopic view adaptive positioning and high-power imaging guiding method

Abstract

The invention discloses a microscopic visual field adaptive positioning and high-power imaging guiding method, and belongs to the technical field of image processing. The method comprises the following steps: firstly, training a Yov11-FocalModulation network by using a training sample graph; loading a to-be-detected sample to a microscope loader, scanning by using a low-power objective lens, and splicing to obtain a panorama; inputting the panorama into the trained network to obtain a prediction box; screening out a smooth prediction frame according to the texture complexity and the gradient variation; and finally, converting the center coordinate of the smooth prediction frame to obtain a target physical coordinate, and after the objective table is moved, switching a high-power objective lens to scan an area in the frame to obtain a high-power microscopic image. According to the invention, by intelligently positioning the region of interest, the high-power imaging efficiency is improved, and redundant data is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a microscopic field of view adaptive positioning and high-magnification imaging guidance method. Background Art

[0002] Microscopic imaging technology is widely used in fields such as biomedical research (microbial staining, pathological section analysis), materials science (such as semiconductor wafer defect detection), and industrial quality inspection. With the growing demand for high-resolution imaging, the use of high-magnification objectives (such as 40×, 60×, and above) has become a key means of capturing microscopic details. However, the physical characteristics of high-magnification objectives (small field of view and shallow depth of field) lead to a dual bottleneck in efficiency and data volume when scanning the entire sample. Traditional high-magnification microscopes require a mechanically moving stage to locate and image each area, which is time-consuming and generates a large number of redundant images (such as blank areas or non-target tissue).

[0003] To solve this problem, the existing technology generally adopts two types of solutions: ‌1. Manually annotate local scans‌ The sample is scanned with a low-magnification objective lens to create a panoramic mosaic of the sample. The operator manually marks the scan area on the panoramic mosaic, and then uses a high-magnification objective lens to scan these manually marked areas. This method requires manual intervention for each scanned sample, and its efficiency is difficult to meet the needs of high-throughput business. 2. Pre-planned path continuous scanning: By pre-planning the scanning path, a high-magnification objective lens is used to uniformly capture high-magnification images in a uniformly designated area. However, this method cannot accurately scan the truly valuable sample areas when the sample content is irregularly distributed and there are large individual differences. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the present invention provides a method for adaptively positioning a microscopic field of view and guiding high-magnification imaging, which solves the problems of low scanning efficiency and excessive redundant image data caused by the small field of view during the scanning of heterogeneous samples by existing high-magnification objective lenses.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a microscopic field adaptive positioning and high-magnification imaging guidance method, comprising the following steps: S1. Use the training sample graph to train the Yolov11-FocalModulation network to obtain the trained Yolov11-FocalModulation network; S2. Load the sample to be tested into the microscope loader, scan the sample to be tested using a low-magnification objective lens to obtain multiple local images, and stitch the multiple local images together to obtain a panoramic image of the sample to be tested; S3. Input the panorama of the sample to be detected into the trained Yolov11-FocalModulation network to obtain the predicted frame; S4. Filter the prediction frames based on texture complexity and gradient change, and retain smooth prediction frames; S5. Perform coordinate transformation on the center coordinates of the smooth prediction frame to obtain the target physical coordinates, move the stage to the target physical coordinates, switch the high-magnification objective lens to scan the area within the smooth prediction frame, and obtain a high-magnification microscopic image.

[0006] Furthermore, S1 includes the following sub-steps: S11. Scan the entire field of view of the slide sample using a low-power microscope to obtain multiple partial images, and stitch the multiple partial images together to obtain a panoramic image of the slide sample; S12, manually marking abnormal areas on the panoramic image to obtain a training sample image; S13. Use the training sample graph to train the Yolov11-FocalModulation network to obtain a trained Yolov11-FocalModulation network.

[0007] Furthermore, the structure of the Yolov11-FocalModulation network in S1 is: based on the original Yolov11 network architecture, the SPPF unit in the original Yolov11 network architecture is replaced by the FocalModulation unit.

[0008] Furthermore, S4 includes the following sub-steps: S41, converting the prediction frame from the RGB color space to the LAB color space, and extracting the brightness channel in the LAB color space; S42, performing detail enhancement on the brightness channel of the predicted frame to obtain a detail enhanced frame; S43, adjusting the brightness of the detail enhancement frame to obtain a brightness adjustment frame; S44, extracting texture complexity of the brightness adjustment frame; S45, obtaining the X-direction gradient and the Y-direction gradient of the brightness adjustment frame, and calculating the gradient change; S46, calculating texture smoothness according to the gradient variation and texture complexity; S47. Retain the prediction box whose texture smoothness is greater than the smoothness threshold.

[0009] Furthermore, S42 includes the following sub-steps: S421, performing an opening operation on the brightness channel of the prediction frame using a 51x51 elliptical kernel to obtain an opening operation result; S422: Subtract the opening operation result from the brightness channel of the predicted frame to obtain a detail enhancement frame.

[0010] Furthermore, the formula for brightness adjustment in S43 is: ;

[0011] Among them, L AD is the brightness adjustment frame, L is the brightness channel of the prediction frame, and E is the detail enhancement frame.

[0012] Furthermore, the calculation formula of texture complexity in S44 is: ;

[0013] Among them, C is the texture complexity, l i is the brightness value of the i-th pixel in the brightness adjustment frame, N is the number of pixels in the brightness adjustment frame, and i is the pixel number.

[0014] Furthermore, S45 includes the following sub-steps: S451, slide a 5×5 X-direction Sobel kernel on the brightness adjustment frame and calculate the X-direction gradient of each sliding position; S452, slide a 5×5 Y-direction Sobel kernel on the brightness adjustment frame and calculate the Y-direction gradient of each sliding position; S453, calculating the standard deviation of the X-direction gradient at each sliding position to obtain the X-direction gradient standard deviation; S454, calculating the standard deviation of the Y-direction gradient at each sliding position to obtain the Y-direction gradient standard deviation; S455: Add the standard deviation of the gradient in the X direction and the standard deviation of the gradient in the Y direction to obtain the gradient variation.

[0015] Furthermore, the formula for calculating texture smoothness in S46 is: ;

[0016] Among them, S is the texture smoothness, C is the texture complexity, and G is the gradient change.

[0017] Furthermore, S5 includes the following sub-steps: S51. Get the center coordinates of the smoothed prediction box: ;

[0018] Among them, X p is the horizontal coordinate of the center coordinate, Y pis the ordinate of the center coordinate, x1 is the abscissa of the first corner point of the smooth prediction box, x2 is the abscissa of the second corner point of the smooth prediction box, x3 is the abscissa of the third corner point of the smooth prediction box, x4 is the abscissa of the fourth corner point of the smooth prediction box, y1 is the ordinate of the first corner point of the smooth prediction box, y2 is the ordinate of the second corner point of the smooth prediction box, y3 is the ordinate of the third corner point of the smooth prediction box, and y4 is the ordinate of the fourth corner point of the smooth prediction box; S52. Perform coordinate transformation on the center coordinates of the smoothed prediction frame to obtain the target physical coordinates: ;

[0019] Among them, X z is the horizontal coordinate of the target physical coordinate, Y z is the vertical coordinate of the target physical coordinate, X s Y is the physical starting point for calibration of the upper left corner of the slide where the sample to be tested is located on the stage. s P is the physical starting calibration vertical coordinate of the upper left corner of the slide sample where the sample to be tested is located on the stage, d The coefficient for converting pixel value to physical distance; S53, moving the stage to the target physical coordinates, switching the high-magnification objective lens to scan the area within the smooth prediction frame, and obtaining a high-magnification microscopic image.

[0020] The beneficial effects of the present invention are as follows: the present invention processes the panoramic image of the sample to be detected through the trained Yolov11-FocalModulation network to obtain the suspected area where the target is located, that is, the prediction frame, and then screens out a smooth prediction frame through texture complexity and gradient change, effectively filtering out worthless areas and avoiding redundant scanning. The present invention utilizes the synergy of panoramic scanning splicing with a low-power objective lens and precise imaging with a high-power objective lens. In the detection of heterogeneous samples, the present invention not only uses the low-power objective lens to quickly cover the entire sample to obtain macro information, but also relies on the high-power objective lens to focus on key areas to capture micro details, breaking through the limitations of the field of view of the high-power objective lens, greatly improving scanning efficiency, and reducing the generation of redundant image data; at the same time, coordinate transformation and precise movement control of the stage ensure that high-power imaging accurately focuses on the target area, achieves accurate shooting of the target, improves scanning efficiency, and reduces redundant image data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of a method for adaptive positioning of microscopic field of view and high-magnification imaging guidance; Figure 2 This is a panoramic view of the slide sample in S11; Figure 3 is a schematic diagram of the training sample graph; Figure 4Schematic diagram of the target box loss of the training set; Figure 5 Schematic diagram of target box loss for the validation set; Figure 6 This is the change graph of accuracy when using the training set; Figure 7 This is the recall rate change diagram when using the training set; Figure 8 This is the average accuracy change graph when using the validation set; Figure 9 This is a schematic diagram of the structure of the Yolov11-FocalModulation network; Figure 10 It is a panoramic view of the samples to be tested in S2; Figure 11 Schematic diagram of the filtered prediction box; Figure 12 It is a high-magnification microscopic image obtained based on the filtered prediction box. DETAILED DESCRIPTION

[0022] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0023] like Figure 1 As shown, a microscopic field adaptive positioning and high-magnification imaging guidance method includes the following steps: S1. Use the training sample graph to train the Yolov11-FocalModulation network to obtain the trained Yolov11-FocalModulation network; S2. Load the sample to be tested into the microscope loader, scan the sample to be tested using a low-magnification objective lens to obtain multiple local images, and stitch the multiple local images together to obtain a panoramic image of the sample to be tested; S3. Input the panorama of the sample to be detected into the trained Yolov11-FocalModulation network to obtain the predicted frame; S4. Filter the prediction frames based on texture complexity and gradient change, and retain smooth prediction frames; S5. Perform coordinate transformation on the center coordinates of the smooth prediction frame to obtain the target physical coordinates, move the stage to the target physical coordinates, switch the high-magnification objective lens to scan the area within the smooth prediction frame, and obtain a high-magnification microscopic image.

[0024] In this embodiment, S1 includes the following sub-steps: S11. Use a low-power microscope to scan the entire field of view of the slide sample to obtain multiple local images, and stitch the multiple local images together to obtain a panoramic view of the slide sample, such as Figure 2 As shown; S12, manually mark the abnormal areas on the panoramic image to obtain a training sample image, such as Figure 3 As shown; S13. Use the training sample graph to train the Yolov11-FocalModulation network to obtain a trained Yolov11-FocalModulation network.

[0025] In step S1, the dataset annotated by professionals is divided into a training set and a validation set in a ratio of 7:3. This is then fed into the Yolov11-FocalModulation network for training and validation. After multiple rounds of training, the optimal model parameters are achieved when the recall and precision metrics of the Yolov11-FocalModulation network no longer increase.

[0026] like Figure 4 and 5 As shown in , when the number of samples in the training set and validation set reaches 200-400, the error no longer decreases significantly (indicating that the fit to the training data is good). Figure 6 As shown in , when the number of samples in the training set reaches 200~400, the accuracy of the Yolov11-FocalModulation network no longer increases. Figure 7 As shown in , when the number of samples in the training set reaches 200~400, the recall rate of the Yolov11-FocalModulation network no longer increases. Figure 8 As shown in the figure, when the number of samples in the validation set reaches 200~400, the average precision no longer increases.

[0027] In this embodiment, the structure of the Yolov11-FocalModulation network in S1 is as follows: based on the original Yolov11 network architecture, the SPPF unit in the original Yolov11 network architecture is replaced with the FocalModulation unit, such as Figure 9 shown.

[0028] exist Figure 9In the Yolov11 network architecture, the abbreviation FM stands for FocalModulation unit. SPPF, a spatial pyramid pooling method, extracts multi-scale feature information by pooling features at different scales. However, SPPF treats features in all regions identically and does not distinguish between important and unimportant areas. Therefore, SPPF is not ideal for finding suitable viewing areas in microscopic images. Therefore, the SPPF unit in Yolov11 is replaced by the FocalModulation unit, which incorporates an attention mechanism.

[0029] Focal Modulation is an attention mechanism that weights each region based on its importance. It first calculates the importance factor for each region and then weights features accordingly. This allows the model to focus more on the more important, visually relevant regions of the image, improving its ability to extract visually relevant features.

[0030] In this embodiment, the panoramic image of the sample to be detected in S2 is as follows: Figure 10 shown.

[0031] In the present invention, the predicted box obtained in S3 includes: four corner point coordinates.

[0032] In this embodiment, S4 includes the following sub-steps: S41, converting the prediction frame from the RGB color space to the LAB color space, and extracting the brightness channel in the LAB color space; S42, performing detail enhancement on the brightness channel of the predicted frame to obtain a detail enhanced frame; S43, adjusting the brightness of the detail enhancement frame to obtain a brightness adjustment frame; S44, extracting texture complexity of the brightness adjustment frame; S45, obtaining the X-direction gradient and the Y-direction gradient of the brightness adjustment frame, and calculating the gradient change; S46, calculating texture smoothness according to the gradient variation and texture complexity; S47. Retain the prediction box whose texture smoothness is greater than the smoothness threshold.

[0033] The present invention extracts the brightness channel that is sensitive to illumination changes in the LAB color space, eliminates color interference, enhances the details of the brightness channel of the prediction frame, and then adjusts the brightness to obtain a brightness adjustment frame that strengthens the cracking characteristics. On the one hand, the present invention reflects the texture clarity of the brightness adjustment frame through the texture complexity of the brightness adjustment frame. On the other hand, the present invention calculates the gradient change by the X-direction gradient and the Y-direction gradient to measure the texture in different directions. The present invention measures the smoothness of the prediction frame in two ways and selects the prediction frame with a smoothness greater than the smoothness threshold, such as Figure 11 shown.

[0034] In this embodiment, S42 includes the following sub-steps: S421, performing an opening operation on the brightness channel of the prediction frame using a 51x51 elliptical kernel to obtain an opening operation result; S422: Subtract the opening operation result from the brightness channel of the predicted frame to obtain a detail enhancement frame.

[0035] The formula for obtaining the detail enhancement frame is: ;

[0036] Among them, E is the detail enhancement box, L is the brightness channel of the prediction box, b is the 51x51 elliptical kernel, It is an opening operation.

[0037] In this embodiment, the formula for brightness adjustment in S43 is: ;

[0038] Among them, L AD is the brightness adjustment frame, L is the brightness channel of the prediction frame, and E is the detail enhancement frame.

[0039] The present invention uses a 51x51 elliptical kernel to perform an opening operation (erode first and then dilate) on the brightness channel, which can effectively remove the non-uniform lighting background, enhance low-contrast details, and enhance the crack characteristics of low-contrast areas. The 2x detail enhancement frame is then added to the brightness channel of the prediction frame to obtain a brightness adjustment frame, which not only retains the basic brightness information but also amplifies the enhanced details.

[0040] In this embodiment, the calculation formula of texture complexity in S44 is: ;

[0041] Among them, C is the texture complexity, l i is the brightness value of the i-th pixel in the brightness adjustment frame, N is the number of pixels in the brightness adjustment frame, and i is the pixel number.

[0042] In this embodiment, S45 includes the following sub-steps: S451, slide a 5×5 X-direction Sobel kernel on the brightness adjustment frame and calculate the X-direction gradient of each sliding position; S452, slide a 5×5 Y-direction Sobel kernel on the brightness adjustment frame and calculate the Y-direction gradient of each sliding position; S453, calculating the standard deviation of the X-direction gradient at each sliding position to obtain the X-direction gradient standard deviation; S454, calculating the standard deviation of the Y-direction gradient at each sliding position to obtain the Y-direction gradient standard deviation; S455: Add the standard deviation of the gradient in the X direction and the standard deviation of the gradient in the Y direction to obtain the gradient variation.

[0043] Different textures and edges (such as cracks) in an image have different rates of brightness change in the horizontal (X-direction) and vertical (Y-direction) directions. For example, cracks can exhibit various orientations. The X-direction gradient can highlight vertical edge variations (e.g., a horizontally extending crack will have a significant vertical brightness change), while the Y-direction gradient can highlight horizontal edge variations (e.g., a vertically extending crack will have a significant horizontal brightness change). By calculating the gradient in two orthogonal directions, we can fully capture the directional characteristics of edges and textures within a region, avoiding missing key features due to a single orientation.

[0044] If the area is smooth, the X and Y gradient values ​​are mostly small and concentrated, and the standard deviation approaches 0. However, if the area contains complex textures such as cracks, the gradient values ​​fluctuate greatly (some areas have large gradients, others have small gradients), and the standard deviation increases significantly. The standard deviation can be used to convert the gradient "fluctuation" into a numerical indicator to distinguish smooth areas from areas with complex textures.

[0045] In this embodiment, the formula for calculating texture smoothness in S46 is: ;

[0046] Among them, S is the texture smoothness, C is the texture complexity, and G is the gradient change.

[0047] In this embodiment, S5 includes the following sub-steps: S51. Get the center coordinates of the smoothed prediction box: ;

[0048] Among them, X p is the horizontal coordinate of the center coordinate, Y pis the ordinate of the center coordinate, x1 is the abscissa of the first corner point of the smooth prediction box, x2 is the abscissa of the second corner point of the smooth prediction box, x3 is the abscissa of the third corner point of the smooth prediction box, x4 is the abscissa of the fourth corner point of the smooth prediction box, y1 is the ordinate of the first corner point of the smooth prediction box, y2 is the ordinate of the second corner point of the smooth prediction box, y3 is the ordinate of the third corner point of the smooth prediction box, and y4 is the ordinate of the fourth corner point of the smooth prediction box; S52. Perform coordinate transformation on the center coordinates of the smoothed prediction frame to obtain the target physical coordinates: ;

[0049] Among them, X z is the horizontal coordinate of the target physical coordinate, Y z is the vertical coordinate of the target physical coordinate, X s Y is the physical starting point for calibration of the upper left corner of the slide where the sample to be tested is located on the stage. s P is the physical starting calibration vertical coordinate of the upper left corner of the slide sample where the sample to be tested is located on the stage, d The coefficient for converting pixel value to physical distance; S53, move the stage to the target physical coordinates, switch the high-power objective lens to scan the area within the smooth prediction frame, and obtain a high-power microscopic image, such as Figure 12 shown.

[0050] Keep the sample on the stage stationary and move the stage to Coordinates, switch the high-power objective lens to scan the target area, and you can get a high-power microscopic image of the effective area of ​​the sample.

[0051] The present invention determines the target area for high-power objective lens scanning through professional annotation and deep learning models, and ultimately realizes accurate scanning of the designated area by the high-power objective lens, thereby improving the efficiency and accuracy of microscope sample analysis.

[0052] The present invention processes the panoramic image of the sample to be detected through the trained Yolov11-FocalModulation network to obtain the suspected area where the target is located, that is, the prediction frame. Then, through the texture complexity and gradient change, a smooth prediction frame is screened out, effectively filtering out worthless areas and avoiding redundant scanning. The present invention utilizes the synergy of low-power objective lens panoramic scanning splicing and high-power objective lens precise imaging. In the detection of heterogeneous samples, the present invention not only uses the low-power lens to quickly cover the entire sample to obtain macro information, but also relies on the high-power lens to focus on key areas to capture micro details, breaking through the limitations of the high-power objective lens field of view, greatly improving scanning efficiency, and reducing the generation of redundant image data; at the same time, coordinate transformation and precise movement control of the stage ensure that high-power imaging accurately focuses on the target area, achieves accurate shooting of the target, improves scanning efficiency, and reduces redundant image data.

[0053] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A microscopic field adaptive positioning and high-magnification imaging guidance method, characterized in that: The following steps are involved: S1. Use the training sample graph to train the Yolov11-FocalModulation network to obtain the trained Yolov11-FocalModulation network; S2. Load the sample to be tested into the microscope loader, scan the sample to be tested using a low-magnification objective lens to obtain multiple local images, and stitch the multiple local images together to obtain a panoramic image of the sample to be tested; S3. Input the panorama of the sample to be detected into the trained Yolov11-FocalModulation network to obtain the predicted frame; S4. Filter the prediction frames based on texture complexity and gradient change, and retain smooth prediction frames; S5. Perform coordinate transformation on the center coordinates of the smooth prediction frame to obtain the target physical coordinates, move the stage to the target physical coordinates, switch the high-magnification objective lens to scan the area within the smooth prediction frame, and obtain a high-magnification microscopic image.

2. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 1, characterized in that: The S1 comprises the following sub-steps: S11. Scan the entire field of view of the slide sample using a low-power microscope to obtain multiple partial images, and stitch the multiple partial images together to obtain a panoramic image of the slide sample; S12, manually marking abnormal areas on the panoramic image to obtain a training sample image; S13. Use the training sample graph to train the Yolov11-FocalModulation network to obtain a trained Yolov11-FocalModulation network.

3. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 1, characterized in that: The structure of the Yolov11-FocalModulation network in S1 is as follows: based on the original Yolov11 network architecture, the SPPF unit in the original Yolov11 network architecture is replaced by the FocalModulation unit.

4. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 1, characterized in that: The S4 comprises the following sub-steps: S41, converting the prediction frame from the RGB color space to the LAB color space, and extracting the brightness channel in the LAB color space; S42, performing detail enhancement on the brightness channel of the predicted frame to obtain a detail enhanced frame; S43, adjusting the brightness of the detail enhancement frame to obtain a brightness adjustment frame; S44, extracting texture complexity of the brightness adjustment frame; S45, obtaining the X-direction gradient and the Y-direction gradient of the brightness adjustment frame, and calculating the gradient change; S46, calculating texture smoothness according to the gradient variation and texture complexity; S47. Retain the prediction box whose texture smoothness is greater than the smoothness threshold.

5. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 4, characterized in that: The S42 includes the following sub-steps: S421, performing an opening operation on the brightness channel of the prediction frame using a 51x51 elliptical kernel to obtain an opening operation result; S422: Subtract the opening operation result from the brightness channel of the predicted frame to obtain a detail enhancement frame.

6. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 4, characterized in that: The formula for brightness adjustment in S43 is: ； Among them, L AD is the brightness adjustment frame, L is the brightness channel of the prediction frame, and E is the detail enhancement frame.

7. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 4, characterized in that: The calculation formula of the texture complexity in S44 is: ； Among them, C is the texture complexity, l i is the brightness value of the i-th pixel in the brightness adjustment frame, N is the number of pixels in the brightness adjustment frame, and i is the pixel number.

8. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 4, characterized in that: The S45 includes the following sub-steps: S451, slide a 5×5 X-direction Sobel kernel on the brightness adjustment frame and calculate the X-direction gradient of each sliding position; S452, slide a 5×5 Y-direction Sobel kernel on the brightness adjustment frame and calculate the Y-direction gradient of each sliding position; S453, calculating the standard deviation of the X-direction gradient at each sliding position to obtain the X-direction gradient standard deviation; S454, calculating the standard deviation of the Y-direction gradient at each sliding position to obtain the Y-direction gradient standard deviation; S455: Add the standard deviation of the gradient in the X direction and the standard deviation of the gradient in the Y direction to obtain the gradient variation.

9. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 4, characterized in that: The formula for calculating texture smoothness in S46 is: ； Among them, S is the texture smoothness, C is the texture complexity, and G is the gradient change.

10. The microscopic field adaptive positioning and high-magnification imaging guidance method according to claim 1, characterized in that: The S5 comprises the following sub-steps: S51. Get the center coordinates of the smoothed prediction box: ； Among them, X p is the horizontal coordinate of the center coordinate, Y p is the ordinate of the center coordinate, x1 is the abscissa of the first corner point of the smooth prediction box, x2 is the abscissa of the second corner point of the smooth prediction box, x3 is the abscissa of the third corner point of the smooth prediction box, x4 is the abscissa of the fourth corner point of the smooth prediction box, y1 is the ordinate of the first corner point of the smooth prediction box, y2 is the ordinate of the second corner point of the smooth prediction box, y3 is the ordinate of the third corner point of the smooth prediction box, and y4 is the ordinate of the fourth corner point of the smooth prediction box; S52. Perform coordinate transformation on the center coordinates of the smoothed prediction frame to obtain the target physical coordinates: ； Among them, X z is the horizontal coordinate of the target physical coordinate, Y z is the vertical coordinate of the target physical coordinate, X s Y is the physical starting point for calibration of the upper left corner of the slide where the sample to be tested is located on the stage. s P is the physical starting calibration vertical coordinate of the upper left corner of the slide sample where the sample to be tested is located on the stage, d The coefficient for converting pixel value to physical distance; S53, moving the stage to the target physical coordinates, switching the high-magnification objective lens to scan the area within the smooth prediction frame, and obtaining a high-magnification microscopic image.

Images