Image splicing technology for digital microscopic scanning system

Abstract

The invention belongs to the technical field of microscopic image panoramic stitching, and provides an image stitching technology for a digital microscopic scanning system. Comprising the steps of panoramic focal length strategy setting, mobile shooting, label processing, local feature point extraction, single image matrix determination, overall splicing and non-blank overlapping region splicing, sliding feature fine alignment path deviation data calculation, panoramic focal length strategy adjustment and gap fusion processing. According to the method, the ResNet-18 model is used for performing high-density region judgment, so that the feature point detection efficiency is improved; through a linear gradual change technology and a recursive algorithm, the splicing problem of non-blank overlapping areas is solved, and the quality of the spliced image is improved; weighted average fusion is carried out through a Gaussian function, the problems of brightness difference, color non-uniformity, edge blurring and the like of a splicing area are solved, and the visual effect of the image is improved. By judging the position deviation function, the objective table is adjusted in real time, the operation process is simplified, and the cooperative efficiency of scanning and splicing is improved.

Description

Technical Field

[0001] The present invention relates to an image splicing technology for a digital microscopic scanning system, and in particular to an image splicing technology for a digital microscopic scanning system. Background Art

[0002] Microscopy, a key observation tool in fields such as biomedicine and materials science, has evolved from manual microscopic observation to automated microscopic scanning and finally to digital microscopic scanning. In the early days, researchers relied on manual microscopes to observe and analyze samples. This method often required a high level of operator skill and patience, as the acquisition of microscopic images was tedious and prone to human error. With technological advancements, automated microscopic scanning has gradually emerged, enabling high-precision scanning of large samples. This advancement has greatly improved work efficiency while reducing human error. However, whether manual or automated, traditional microscopy techniques typically only capture a single image within a limited field of view. While this single image provides high-resolution detail, it often fails to capture the full picture of the sample due to the limited field of view. With the increasing demands of research, relying solely on single images is no longer sufficient to meet the demand for large-scale, high-resolution images. In histological and biological tissue research, obtaining complete panoramic images to visualize the overall structure of the sample has become particularly important. Against this backdrop, image stitching technology has emerged. Image stitching aims to stitch multiple overlapping microscopic images into a complete, seamless panoramic image. Through this technology, researchers can obtain a wide range of sample observations while retaining the details of the original image, greatly expanding the application range and research depth of the microscope.

[0003] Existing image stitching technologies require processing large amounts of image data, particularly in large-scale microscopic scanning. As the resolution and number of images increase, computational requirements increase exponentially. Image stitching often involves complex image processing algorithms, such as feature matching and image transformation, which require significant computational resources. Consequently, when stitching large samples, this often leads to low system computational efficiency and prolonged processing time, severely impacting the real-time and efficiency of image stitching. Although existing stitching algorithms can combine multiple images into a single panoramic image, stitching artifacts are still prone to appear at image edges or in overlapping areas. These artifacts include color discrepancies, uneven brightness, and blurred transitions, impacting the overall visual quality and accuracy of the image. The presence of these stitching artifacts poses challenges to accurate sample analysis, particularly in fields requiring high precision, such as pathology and materials science. Current image stitching technologies often rely on traditional microscopic scanning methods and lack effective integration with modern integrated microscopic scanning systems. With the advancement of microscopic imaging technology, automated, integrated microscopic scanning systems are becoming increasingly mainstream. However, existing stitching technologies often operate independently of the scanning device, requiring stitching processing after image acquisition, failing to achieve seamless integration with the microscope system. This disconnect not only affects the efficiency of image acquisition and stitching, but also increases operational complexity. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide an image stitching technology for a digital microscopy scanning system. By using the ResNet-18 model to judge high-density areas, the efficiency of feature point detection is improved; by using linear gradient technology and recursive algorithm, the stitching problem of non-blank overlapping areas is solved, and the quality of the stitched image is improved; by using Gaussian function for weighted average fusion, problems such as brightness difference, color unevenness and edge blur in the stitching area are solved, and the visual effect of the image is improved; by judging the position deviation function, real-time adjustment of the stage is achieved, the operation process is simplified, and the collaborative efficiency of scanning and stitching is improved.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] An image stitching technology for a digital microscopy scanning system, comprising:

[0007] Set the panoramic focus strategy;

[0008] The panoramic focal length strategy is used to control a preset stage to move and shoot the target sample to obtain a local image;

[0009] Inputting the local image into the pre-trained ResNet-18 model for label processing to obtain a high-density area judgment result;

[0010] Extracting local feature points of the local image using ORB technology and pyramid multi-scale image strategy according to the high-density area judgment result;

[0011] Determining the isographic matrix of the adjacent partial images using the local feature points;

[0012] Using the isographic matrix to perform overall splicing of the partial images, and using a linear gradient technique and a recursive algorithm to perform splicing of non-blank overlapping areas of adjacent partial images to obtain a temporary spliced ​​image;

[0013] Performing sliding feature fine alignment of feature edge topology on the temporary spliced ​​image to obtain a preprocessed spliced ​​image;

[0014] When the pre-processed stitched image is aligned, the path of the stage is monitored using a position deviation judgment function to obtain path deviation data, and the panoramic focus strategy is adjusted using the path deviation data;

[0015] The grayscale values ​​of the pre-processed stitched images are weighted averaged and fused using a Gaussian function to obtain a panoramic optimized image.

[0016] Preferably, the calculation formula of the isometric matrix is:

[0017]

[0018] Wherein, a and b are the horizontal and vertical pixel coordinates of the local image respectively; x' and y'

[0019] are the pixel coordinates of the temporary spliced ​​image in the horizontal and vertical directions respectively; H is the isographic matrix; hnn.

[0020] Preferably, the position deviation determination function is:

[0021]

[0022] in, x and y are the pixel coordinates of the pre-processed stitched image in the horizontal and vertical directions, respectively; WNCC(u, v) is the position deviation determination function; I1(x, y) and I2(x+u, y+v) are the pixel values ​​of the adjacent temporary stitched images at (x, y) and (x+u, y+v), respectively; and are the average values ​​of I1(x,y) and I2(x+u,y+v); w(x,y) is the weight at (x,y); bright (x,y) is the brightness of the pixel at (x,y); is the average pixel brightness.

[0023] Preferably, the scanning mode of the panoramic focus strategy is serpentine scanning.

[0024] Preferably, the training process of the ResNet-18 model includes:

[0025] Collect cell images to be trained;

[0026] Using SIFT technology to detect feature points of the cell image to be trained, and obtain densely distributed feature point data;

[0027] Gridding the cell image to be trained based on the densely distributed feature point data to obtain a grid image;

[0028] Calculating the number of feature points of each grid in the grid image to obtain feature point distribution data;

[0029] Calculating the coordinate range of the grid with the densest feature points in the grid image using the feature point distribution data, and performing label processing on the coordinate range to obtain a labeled image;

[0030] The cell image to be trained and the label image are input into a preset ResNet-18 network for training to obtain the trained ResNet-18 model.

[0031] Preferably, the parameter for controlling the movement of the stage is the number of pulses; the calculation formula for the number of pulses is:

[0032] p = distance_row / u;

[0033] Wherein, p is the number of pulses; distance_row is the lateral movement distance of a single shot; and u is the pulse equivalent of the motor in the stage.

[0034] Preferably, performing seam fusion processing on the original panoramic image using a Gaussian function to obtain a panoramic optimized image includes:

[0035] Extracting pixel grayscale values ​​of overlapping areas in the original panoramic image;

[0036] Using a Gaussian function to perform weighted averaging on the pixel grayscale values ​​to obtain an optimized grayscale value;

[0037] The optimized grayscale value is updated to the original panoramic image to obtain the panoramic optimized image.

[0038] Preferably, the grid size of the grid image is 300 pixels.

[0039] The present invention discloses the following technical effects:

[0040] The present invention provides an image stitching technology for a digital microscopic scanning system. By utilizing the ResNet-18 model to perform high-density area judgment, the defect of low feature point detection efficiency in the prior art is solved, and rapid extraction of local image feature points is achieved. By using linear gradient technology and a recursive algorithm, the stitching problem of non-blank overlapping areas is solved, and the quality of the stitched image is optimized. By performing weighted average fusion using a Gaussian function, problems such as brightness difference, color unevenness, and edge blur in the stitching area are solved, and the visual effect of the stitched image is improved. By judging the position deviation function, the shortcomings of the conventional scanning method, such as low efficiency, long time, and high complexity, are solved, and real-time adaptive adjustment of the stage is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 A schematic diagram of the microscopic cell image stitching process provided by an embodiment of the present invention;

[0043] Figure 2 A flowchart of microscopic cell image stitching provided by an embodiment of the present invention;

[0044] Figure 3 Flowchart of the key algorithms provided by the embodiments of the present invention;

[0045] Figure 4 A schematic diagram of a scanning method provided by an embodiment of the present invention;

[0046] Figure 5 An example diagram of dividing a feature point dense area provided by an embodiment of the present invention;

[0047] Figure 6 A schematic diagram of images to be spliced ​​provided in an embodiment of the present invention;

[0048] Figure 7 A minimum feature point area map provided by an embodiment of the present invention;

[0049] Figure 8 An example diagram of SIFT combined with the flann matcher provided in an embodiment of the present invention;

[0050] Figure 9 A schematic diagram of edge extraction results provided by an embodiment of the present invention;

[0051] Figure 10A schematic diagram of the splicing result provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] The purpose of the present invention is to provide an image stitching technology for a digital microscopy scanning system. The technology improves the efficiency of feature point detection by using the ResNet-18 model to judge high-density areas. It solves the stitching problem of non-blank overlapping areas and improves the quality of the stitched image through linear gradient technology and recursive algorithm. It solves the problems of brightness difference, color unevenness and edge blur in the stitching area and improves the visual effect of the image through weighted average fusion using Gaussian function. It realizes real-time adjustment of the stage by judging the position deviation function, simplifies the operation process, and improves the collaborative efficiency of scanning and stitching.

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Figure 1 A schematic diagram of the microscopic cell image stitching process provided by an embodiment of the present invention, Figure 2 The flowchart of microscopic cell image stitching provided by the embodiment of the present invention is as follows: Figure 3 The key algorithm flow chart provided by the embodiment of the present invention is as follows: Figure 1 、 Figure 2 as well as Figure 3 As shown, the present invention provides an image stitching technology for a digital microscopic scanning system, comprising:

[0056] Step 100: Setting the panoramic focus strategy;

[0057] Step 200: Using a panoramic focal length strategy to control a preset stage to move and shoot a target sample to obtain a partial image;

[0058] Step 300: Input the local image into the pre-trained ResNet-18 model for label processing to obtain a high-density area judgment result;

[0059] Step 400: extracting local feature points of the local image using ORB technology and pyramid multi-scale image strategy based on the high-density area judgment result;

[0060] Step 500: Determine the isographic matrix of adjacent local images using local feature points;

[0061] Step 600: Using a isographic matrix to perform overall stitching of the partial images, and using a linear gradient technique and a recursive algorithm to stitch non-blank overlapping areas of adjacent partial images to obtain a temporary stitched image;

[0062] Step 700: performing sliding feature fine alignment of feature edge topology on the temporary spliced ​​image to obtain a pre-processed spliced ​​image;

[0063] Step 800: When pre-processing the stitched image for alignment, use the position deviation function to monitor the path of the stage to obtain path deviation data, and use the path deviation data to adjust the panoramic focus strategy;

[0064] Step 900: Using a Gaussian function to perform weighted average fusion on the grayscale values ​​of the pre-processed stitched images to obtain a panoramic optimized image.

[0065] Preferably, the calculation formula of the isometric matrix is:

[0066]

[0067] Wherein, a and b are the horizontal and vertical pixel coordinates of the local image respectively; x' and y'

[0068] are the pixel coordinates of the temporary spliced ​​image in the horizontal and vertical directions respectively; H is the isographic matrix; hnn.

[0069] Specifically, the position deviation determination function is:

[0070]

[0071] in,

[0072] ; x and y are the pixel coordinates in the horizontal and vertical directions of the preprocessed spliced ​​image respectively; WNCC(u, v is the position deviation function; I1(x, y) and I2(x+u, y+v

[0073] are the pixel values ​​of the adjacent temporary spliced ​​images at (x, y) and (x+u, y+v) respectively; and are the average values ​​of I1(x,y) and I2(x+u,y+v); w(x,y) is the weight at (x,y); bright (x,y) is the brightness of the pixel at (x,y); is the average pixel brightness.

[0074] refer to Figure 4 The scanning mode of the panoramic focus strategy is serpentine scanning.

[0075] Specifically, the training process of the ResNet-18 model includes:

[0076] Collect cell images to be trained;

[0077] Using SIFT technology to detect feature points of the cell image to be trained, and obtain densely distributed feature point data;

[0078] Gridding the cell image to be trained based on the densely distributed feature point data to obtain a grid image;

[0079] Calculating the number of feature points of each grid in the grid image to obtain feature point distribution data;

[0080] Calculating the coordinate range of the grid with the densest feature points in the grid image using the feature point distribution data, and performing label processing on the coordinate range to obtain a labeled image;

[0081] The cell image to be trained and the label image are input into a preset ResNet-18 network for training to obtain the trained ResNet-18 model.

[0082] Preferably, the parameter for controlling the movement of the stage is the number of pulses; the calculation formula for the number of pulses is:

[0083] p = distance_row / u;

[0084] Wherein, p is the number of pulses; distance_row is the lateral movement distance of a single shot; and u is the pulse equivalent of the motor in the stage.

[0085] Furthermore, the original panoramic image is subjected to seam fusion processing using a Gaussian function to obtain a panoramic optimized image, including:

[0086] Extracting pixel grayscale values ​​of overlapping areas in the original panoramic image;

[0087] Using a Gaussian function to perform weighted averaging on the pixel grayscale values ​​to obtain an optimized grayscale value;

[0088] The optimized grayscale value is updated to the original panoramic image to obtain the panoramic optimized image.

[0089] Preferably, the grid size of the grid image is 300 pixels.

[0090] Specifically, the focus strategy and path planning: Before capturing a panoramic image, the focus strategy for the captured area is calculated to determine the optimal focus position for each shot. Combined with the motion path of the stage, this ensures consistent focus across each row of images, thereby improving stitching accuracy.

[0091] Furthermore, the stage is motion-controlled. The stage's motion control system is the core of this embodiment. By controlling the stage's motor, the position information for each shot can be accurately determined. For each scanned line, the stage calculates the corresponding pixel offset based on the set movement distance, providing basic data for image stitching.

[0092] Preferably, the training of the residual network: generally, high-density areas are distributed in dense cell areas, and the morphological characteristics of cells are obviously different from normal cells, such as uneven cell size, irregular shape, nuclear atypia, and in the microscopic image after staining, specific areas are stained darker or lighter. The image of urine-based cells provided by a hospital is used as a reference data set. Here, a minimum area with dense feature points is divided into an array featuremin of length 4, where the elements represent the pixel x coordinate, pixel y coordinate, pixel width width, and pixel height height. The standard for demarcating the area is to use SIFT (Scale-Invariant Feature Transform) detection. The advantage of this method is that it has strong anti-noise ability and rich detection feature points, so it is more suitable for use as a data set for dense feature point areas. Example diagram for dividing dense feature point areas Figure 5 .

[0093] Furthermore, after SIFT feature point detection, the image is divided into grids, and the number of feature points in each grid is counted. Here, the grid size grid_size is defined as 300 pixels. The density matrix density_matrix is ​​initialized, the number of feature points in each grid is calculated, and the grid with the densest feature points max_desity_index is found. The coordinate range is calculated:

[0094] top_left_x=max_col*grid_size

[0095] top_left_y=max_row*grid_size

[0096] bottom_right_x=(max_col+1)*grid_size

[0097] bottom_right_y=(max_row+1)*grid_size

[0098] The regional data of the above features are labeled and trained using the lightweight network model ResNet-18. This model can achieve a 96% accuracy rate in high-density areas, which improves the speed of feature point screening and thus improves the stitching efficiency.

[0099] refer to Figure 6 、 Figure 7 and Figure 8 Coarse feature point detection: Based on the judgment of the previous model (the lightweight network model ResNet-18), a more accurate high-density feature point range is obtained. Within this pixel area, feature points are detected using ORB (Oriented FAST and Rotated BRIEF) combined with a pyramid multi-scale image strategy. These feature points are used in the subsequent calculation of the homography matrix to achieve image stitching.

[0100] Specifically, the isotropic matrix is ​​determined. The isotropic matrix between the two rows of images is determined by calculating the characteristic points. The isotropic matrix is ​​used for image transformation to enable seamless splicing of the two rows of images. The isotropic matrix mapping process is as follows:

[0101]

[0102] Here, H is a 3x3 isotropic matrix containing 8 independent parameters (because the matrix can be multiplied by a non-zero scalar without changing, h33 is usually normalized to 1).

[0103] Preferably, non-empty overlapping areas are processed. This embodiment avoids the problem of non-empty overlapping areas by accurately calculating the distance and pixel offset of each row movement. This not only improves the stitching efficiency, but also reduces redundant calculations and saves a lot of time. This embodiment sets the number of pulses p = distance_row / u, where distance_row is the distance of horizontal movement of each image shot, and u is the pulse equivalent of the selected model motor. In this way, the distance distance_row and the pixel offset overlap of each row movement can be accurately calculated, and a linear gradient (linear interpolation) is used to fuse the overlapping part of the two pictures, that is, the area with a width of overlap and a height of the original image. In this way, the problem of failure to stitch microscopic local images containing non-blank overlapping areas is avoided in the stitching process of each row, which not only improves the stitching efficiency, but also reduces redundant calculations and saves a lot of time.

[0104] Furthermore, a recursive splicing strategy is used. To obtain a maximum and symmetrical overlapping area, this embodiment adopts a recursive splicing strategy. Through a recursive algorithm, the splicing between each row is performed step by step (row by row), ensuring the maximization and symmetry of the splicing area, thereby improving the splicing quality.

[0105] Specifically, the sliding features of the feature edge topology are precisely aligned. First, edge detection is performed. Since there are small gaps between the overlapping image boundaries of cells during segmentation, a closing operation (dilation followed by erosion) is required to eliminate them. The convolution kernel of the closing operation is set to a 3*3 matrix of all ones, ultimately bridging the narrow gaps in the cell boundaries. The steps are as follows:

[0106] S1: Binary thresholding

[0107] GrayImage=cv2.cvtColor(Image,cv2.COLOR_BGR2GRAY)

[0108] #Split color channels and convert color space RGB->HSV

[0109] ret,Region1=cv2.threshold(GrayImage,180,255,cv2.THRESH_BINARY_INV)

[0110] ImageBlur=cv2.GaussianBlur(Region1,(1,1),5)#Gaussian denoising

[0111] ImageMediaBlur = cv2.medianBlur (ImageBlur, 15) # median filtering

[0112] S2: Erosion, remove small spots in the image

[0113] ImageErode=cv2.erode(ImageMediaBlur,kernel1)

[0114] S3: expansion, restoration and amplification

[0115] ImageDilate=cv2.dilate(ImageErode,kernel,iterations=3)

[0116] S4: Closing operation, removing small blocks inside the object

[0117] ImageClose=cv2.morphologyEx(ImageDilate,cv2.MORPH_CLOSE,kernel)

[0118] The results obtained are referenced Figure 9 .

[0119] Furthermore, a search area is defined around the matched feature points, with a fixed-size window (30x30 window) centered around the initially obtained feature point pairs. Within this window, all points with edge pixel values ​​in the tissue cell edge image are searched. In this way, the path can be regarded as a collection of these edge points. After image segmentation preprocessing, the tissue cell boundary appearing in the image is used as the search path to narrow the search area and further improve the matching speed and matching accuracy. During the image stitching process, the position deviation function is used to monitor the stage's route in real time, and the improved normalized cross-correlation function NCC (Normalized Cross-Correlation) is used, and its calculation formula is as follows:

[0120]

[0121] Specifically, gap fusion processing. For gaps that may appear in the image stitching process, a new grayscale value is generated by traversing the overlapping area and performing weighted averaging on the grayscale value of each pixel to achieve a smooth transition of the image. The stitching gaps are automatically adjusted and fused to ensure the integrity and continuity of the final panoramic image, and the stitching traces caused by the grayscale value differences of different images in the overlapping area are reduced, making the result more natural and coherent. For gaps that may appear in the image stitching process, a new grayscale value is generated by traversing the overlapping area and performing weighted averaging on the grayscale value of each pixel to achieve a smooth transition of the image. Through weighted average fusion, the weight of the distance to the middle part of the overlapping area is large, and the weight of the edge part is small. According to the distance function of the overlapping area, the Gaussian function is used here to set the weight to achieve automatic adjustment and fusion of the stitching gaps, ensuring the integrity and continuity of the final panoramic image, and reducing the stitching traces caused by the grayscale value differences of different images in the overlapping area, making the result more natural and coherent. Panoramic stitching effect reference Figure 10 .

[0122] The beneficial effects of the present invention are as follows:

[0123] The present invention improves the efficiency of feature point detection by using the ResNet-18 model to judge high-density areas; solves the splicing problem of non-blank overlapping areas through linear gradient technology and recursive algorithm, and improves the quality of spliced ​​images; solves the problems of brightness difference, color unevenness and edge blur in the spliced ​​areas through weighted average fusion using Gaussian function, and improves the visual effect of the image; realizes real-time adjustment of the stage by judging the position deviation function, simplifies the operation process, and improves the collaborative efficiency of scanning and splicing.

[0124] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0125] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. An image stitching technology for a digital microscopic scanning system, characterized in that: include: Set the panoramic focus strategy; The panoramic focal length strategy is used to control a preset stage to move and shoot the target sample to obtain a local image; Inputting the local image into a pre-trained ResNet-18 model for label processing to obtain a high-density area judgment result; Extracting local feature points of the local image using ORB technology and pyramid multi-scale image strategy according to the high-density area judgment result; Determining the isographic matrix of the adjacent local images using the local feature points; The partial images are integrally spliced ​​using the isographic matrix, and the adjacent partial images are spliced ​​in non-blank overlapping areas using a linear gradient technique and a recursive algorithm to obtain a temporary spliced ​​image; Performing sliding feature fine alignment of feature edge topology on the temporary spliced ​​image to obtain a preprocessed spliced ​​image; When the pre-processed stitched image is aligned, the position deviation function is used to monitor the route of the stage to obtain path deviation data, and the panoramic focal length strategy is adjusted using the path deviation data; The grayscale values ​​of the preprocessed stitched images are weighted averaged and fused using a Gaussian function to obtain a panoramic optimized image.

2. The image stitching technology for a digital microscopic scanning system according to claim 1, characterized in that: The calculation formula of the isotropic matrix is: Among them, a and b are the pixel coordinates of the local image in the horizontal direction and the vertical direction respectively; x' and y' are the pixel coordinates of the temporary spliced ​​image in the horizontal direction and the vertical direction respectively; H is the isometric matrix; hnn is the element corresponding to the nth row and the nth column of the isometric matrix.

3. The image stitching technology for a digital microscopic scanning system according to claim 1, characterized in that: The position deviation determination function is: in, x and y are the pixel coordinates of the preprocessed stitched image in the horizontal direction and the vertical direction respectively; WNCC(u,v) is the position determination function; I1(x,y) and I2(x+u,y+v) are the pixel values ​​of the adjacent temporary stitched images at (x,y) and (x+u,y+v) respectively; and are the average values ​​of I1(x,y) and I2(x+u,y+v); w(x,y) is the weight at (x,y); bright (x,y) is the brightness of the pixel at (x,y); is the average pixel brightness.

4. The image stitching technology for a digital microscopic scanning system according to claim 1, characterized in that: The scanning mode of the panoramic focus strategy is serpentine scanning.

5. The image stitching technology for a digital microscopic scanning system according to claim 1, characterized in that: The training process of the ResNet-18 model includes: Collect cell images to be trained; Using SIFT technology to detect feature points of the cell image to be trained, and obtain densely distributed feature point data; Based on the densely distributed feature point data, the cell image to be trained is gridded to obtain a grid image; Calculating the number of feature points of each grid in the grid image to obtain feature point distribution data; Calculating the coordinate range of the grid with the densest feature points in the grid image using the feature point distribution data, and performing label processing on the coordinate range to obtain a label image; The cell image to be trained and the label image are input into a preset ResNet-18 network for training to obtain the trained ResNet-18 model.

6. The image stitching technology for a digital microscopic scanning system according to claim 1, characterized in that: The parameter for controlling the movement of the stage is the pulse number; the calculation formula for the pulse number is: p = distance_row / u; Wherein, p is the pulse number; distance_row is the lateral movement distance of a single shot; and u is the pulse equivalent of the motor in the stage.

7. The image stitching technology for a digital microscopic scanning system according to claim 1, characterized in that: The original panoramic image is subjected to seam fusion processing by using a Gaussian function to obtain a panoramic optimized image, including: Extracting the pixel grayscale values ​​of the overlapping areas in the original panoramic image; Using a Gaussian function to perform weighted averaging on the pixel grayscale values ​​to obtain an optimized grayscale value; The optimized grayscale value is updated to the original panoramic image to obtain the panoramic optimized image.

8. The image stitching technology for a digital microscopic scanning system according to claim 5, characterized in that: The grid size of the grid image is 300 pixels.

Images