Autonomous analysis method and system for microscopic image

Abstract

The invention discloses an autonomous analysis method and system of a microscopic image, and relates to the technical field of microscopic images, a plurality of particle marks are determined according to scale strip information of the microscopic image and detection of a particle area, a threshold segmentation mode or a hierarchical watershed mode is selectively triggered based on the plurality of particle marks, and the microscopic image is analyzed according to the threshold segmentation mode or the hierarchical watershed mode. According to the particle marking method and device, the corresponding particle result is output, region stripping of the microscopic image is achieved, the particle region and the information bar region are managed and controlled, the accuracy of multiple particle marks is guaranteed, and the accuracy of the particle result is further improved. Meanwhile, a plurality of particle statistical indexes in the microscopic image are determined based on detection of the particle result, an autonomous analysis report of particles related to the microscopic image is constructed along the plurality of particle statistical indexes, and multi-factor control is performed on the plurality of particle statistical indexes, so that autonomous analysis of the microscopic image is facilitated; therefore, the accuracy of an autonomous analysis report of particles involved in the microscopic image is improved.

Description

Autonomous Analysis Methods and Systems for Microscopic Images Technical Field

[0001] This invention relates to the technical field of microscopic images, and more particularly to an autonomous analysis method and system for microscopic images. Background Technology

[0002] Scanning electron microscopy (SEM) and transmission electron microscopy (TEM) are important tools for observing and analyzing the morphology of microscopic particles in materials science, nanotechnology, and biomedicine. Quantitative analysis of particles in microscopic images is necessary, including particle count, size distribution measurement, and particle density calculation. These analytical results are crucial for evaluating material properties, optimizing fabrication processes, and quality control.

[0003] In existing technologies, traditional manual measurement methods typically require researchers to manually identify scale bars in microscopic images, read their physical length values, and manually measure the pixel length of the scale bars using image processing software (such as ImageJ) to calculate the conversion ratio between pixels and actual physical dimensions. Researchers need to manually measure the diameter or area of ​​each particle in the image one by one, and export the measurement data to software such as Excel for statistical analysis. They may even manually draw particle size distribution histograms. There is no control over particle areas and information bar areas, which affects the accuracy of multiple particle markings and results in low accuracy of autonomous analysis reports involving particles in microscopic images. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides an autonomous analysis method and system for microscopic images.

[0005] This invention provides an autonomous analysis method for microscopic images, comprising: acquiring a microscopic image; determining particle regions and information bar regions based on region stripping of the microscopic image; determining scale bar information of the microscopic image based on the identification of the microscopic image; determining multiple particle markers based on the scale bar information and particle region detection of the microscopic image; selectively triggering a threshold segmentation method or a hierarchical watershed method based on the multiple particle markers to output corresponding particle results; determining multiple particle statistical indicators in the microscopic image based on the particle results detection; and constructing an autonomous analysis report of particles involved in the microscopic image along the multiple particle statistical indicators.

[0006] This invention provides a particle analysis system for silicon wafer microscopic images based on a dynamic window. This system is applied to the aforementioned autonomous analysis method for microscopic images. The system includes: a region stripping module for acquiring microscopic images and determining particle regions and information bar regions based on the region stripping; an image recognition module for determining scale bar information of the microscopic image based on the recognition; a particle result module for determining multiple particle markers based on the scale bar information and particle region detection, selectively triggering threshold segmentation or hierarchical watershed methods based on the multiple particle markers to output corresponding particle results; and an autonomous analysis module for determining multiple particle statistical indicators in the microscopic image based on the particle result detection, and constructing an autonomous analysis report of particles involved in the microscopic image along these indicators.

[0007] Compared with existing technologies, the beneficial effects of this invention are as follows: It acquires microscopic images, determines particle regions and information bar regions based on region stripping of the microscopic images; determines scale bar information of the microscopic images based on image recognition; determines multiple particle markers based on the scale bar information and particle region detection; and selectively triggers threshold segmentation or hierarchical watershed methods based on these multiple particle markers to output corresponding particle results. This achieves region stripping of the microscopic images, enabling control over particle and information bar regions, ensuring the accuracy of multiple particle markers, and further improving the accuracy of particle results. Simultaneously, it determines multiple particle statistical indicators in the microscopic images based on particle result detection, and constructs an autonomous analysis report of particles involved in the microscopic images along these indicators. Multi-factor control of these particle statistical indicators facilitates autonomous analysis of the microscopic images, thereby improving the accuracy of the autonomous analysis report of particles involved in the microscopic images. Attached Figure Description

[0008] Figure 1 is a flowchart illustrating the autonomous analysis method for microscopic images in an embodiment of the present invention; Figure 2 is a flowchart illustrating step S11 in the autonomous analysis method for microscopic images in an embodiment of the present invention; Figure 3 is a flowchart illustrating step S12 in the autonomous analysis method for microscopic images in an embodiment of the present invention; Figure 4 is a flowchart illustrating step S13 in the autonomous analysis method for microscopic images in an embodiment of the present invention; Figure 5 is a flowchart illustrating step S14 in the autonomous analysis method for microscopic images in an embodiment of the present invention; Figure 6 is a structural diagram illustrating the particle analysis system for silicon wafer microscopic images based on a dynamic window in an embodiment of the present invention; Figure 7 is an architecture diagram illustrating the particle analysis method for silicon wafer microscopic images based on a dynamic window in an embodiment of the present invention; Figure 8 is a flowchart illustrating the actual process of the particle analysis method for silicon wafer microscopic images based on a dynamic window in an embodiment of the present invention. Detailed Implementation

[0009] 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.

[0010] Please refer to Figures 1 to 5. A self-analysis method for microscopic images is applied to microscopic image scenarios. The self-analysis method for microscopic images includes: Step S11: Acquire a microscopic image and determine the particle region and information bar region based on the region stripping of the microscopic image; Step S12: Determine the scale bar information of the microscopic image based on the recognition of the microscopic image; Step S13: Determine multiple particle markers based on the scale bar information and particle region detection of the microscopic image, and selectively trigger threshold segmentation or hierarchical watershed method based on multiple particle markers to output the corresponding particle results; Step S14: Determine multiple particle statistical indicators in the microscopic image based on the detection of the particle results, and construct a self-analysis report of particles involved in the microscopic image along the multiple particle statistical indicators.

[0011] Referring to Figure 2, in step S11, the specific steps are as follows: S111: Acquire a microscopic image and scan the microscopic image. At this time, scan upwards line by line from the bottom of the microscopic image and analyze the pixel features of each line. Perform row brightness gradient detection on the pixel features of each line, determine the brightness gradient peak through the average brightness difference within the sliding window, and initially locate the position corresponding to the brightness gradient peak as the candidate boundary; S112: In each candidate boundary, mark the variance of the row pixels of each segmentation line, and perform multiple constraints based on the gradient dimension and variance dimension to determine the target row corresponding to the segmentation line, and determine the final clipping line by combining multiple neighborhoods of the target row. Determine the particle region and information bar region based on the final clipping line and the microscopic image. The particle region refers to the region in the microscopic image that contains the particles to be analyzed; the information bar region refers to the device information region usually contained at the bottom of the microscopic image. The information bar region does not contain any particles to be analyzed.

[0012] In the embodiments of this application, a microscopic image is acquired and the microscopic image is scanned. At this time, the image is scanned line by line from the bottom and the pixel features of each line are analyzed. The row brightness gradient is detected for the pixel features of each line. The brightness gradient peak is determined by the average brightness difference within the sliding window, and the position corresponding to the brightness gradient peak is initially located as a candidate boundary. The position corresponding to the brightness gradient peak is initially located as a candidate boundary.

[0013] At this point, input the microscopic image, initialize the scan pointer to the last row of the image (Height-1), and set the scan termination condition, such as stopping the scan at 80% of the image height to avoid misjudging the horizontal structure inside the image as a boundary. At the same time, construct a sliding window with a height of Hw (usually 3-5 rows). For each row i in the window, calculate the average grayscale value Li of all pixels in that row, and calculate the average brightness difference ΔL = |Li-Li+Hw| between the first and last rows in the window as the inter-row brightness gradient at that position. Therefore, the sliding window is used to calculate the average value instead of the brightness of a single pixel, which aims to smooth the local brightness fluctuations caused by the particles themselves or background noise in the SEM image and prevent false gradient peaks caused by the brightness changes of a single particle.

[0014] As the window moves from bottom to top, the gradient value ΔL is monitored in real time; a gradient threshold Tg is set; when ΔL>Tg, the current position is recorded as a candidate point for the gradient peak; the row position Pgrad corresponding to the maximum gradient value is marked as the initial boundary, which usually corresponds to the junction of the black background of the information bar and the bright background of the image content area.

[0015] For the initially located boundary row Pgrad and its neighborhood (e.g., Pgrad±2), the gray-level variance σ2 of each row of pixels is calculated; the information bar segmentation lines of the microscopic image are usually pure black or pure white, with extremely low gray-level variance; only when the gradient ΔL at this position is greater than the threshold Tg and the variance σ2 is less than the extremely low threshold Tv (approaching 0) is it confirmed as a valid segmentation line; the aim is to exclude brightness abrupt bands formed by dense particle arrangement inside the image, ensuring that the locked information bar boundary is the real one.

[0016] A refined search is performed near the initial boundary Pgrad of the dual-core (e.g., within 5-10 rows above and below); the average brightness of each row within this range is calculated, and the row Pfinal with the lowest average brightness (smallest gray value) is selected as the final cropping line; using the prior knowledge that there is usually a black dividing bar (darkest feature) between the information bar and the content area, the cropping line is ensured to fall precisely on the dividing bar, completely removing the pixel interference of the dividing line itself and preserving the complete and effective image content.

[0017] Specifically, the microscopic image is a SEM photograph of nanomaterial particles with a resolution of 1024×768. The bottom area of ​​the image, about 60 lines high, is an information bar with a black background, containing white text (such as "5kVSE") and a 1-pixel-wide black horizontal dividing line that separates the information bar from the particle area above. The particle area itself has uneven brightness, with some areas having denser and darker particles and others being brighter.

[0018] The system constructs a sliding window with a height of 3 rows upwards starting from row 768 (bottom); when the window moves to near row 708, it covers the boundary between the black information bar (average gray level of about 20) and the above particle area (average gray level of about 120); the calculated ΔL=|20-120|=100, which far exceeds the preset threshold Tg=30; therefore, row 708 is initially marked as a candidate boundary Pgrad.

[0019] The system calculates the variance for row 708 and its neighborhood. Since row 708 is a pure black dividing line, its pixel values ​​are almost all 0, and the calculated variance σ2≈0.5, which meets the requirement of extremely low variance. Assuming that although the dark area in row 400 has a large gradient, its row variance σ2 will be significantly higher (e.g., >50) because the area contains grain texture, it is judged as an invalid boundary and is excluded.

[0020] Within the tiny neighborhood of rows 705 to 712, the system calculates the average brightness row by row: row 710 (black background for information bar): average brightness 20; row 709 (black background for information bar): average brightness 18; row 708 (black dividing line): average brightness 5 (darkest); row 707 (start of particle content area): average brightness 100. The system determines that row 708 has the lowest average brightness and identifies it as the final clipping line Pfinal. The system automatically clips all pixel areas from row 708 down, retaining only the effective particle area from row 1 to row 707 for subsequent particle analysis. At this point, the text information and dividing line interference at the bottom are completely removed without any loss of particle content.

[0021] Furthermore, within each candidate boundary, the variance of the row pixels of each segmentation line is marked, and multiple constraints are applied based on the gradient dimension and variance dimension to determine the target row corresponding to the segmentation line. The final clipping line is determined by combining multiple neighborhoods of the target row. Based on the final clipping line and the microscopic image, the particle region and the information bar region are determined. The particle region refers to the region in the microscopic image that contains the particles to be analyzed; the information bar region refers to the device information region usually contained at the bottom of the microscopic image. The information bar region does not contain any particles to be analyzed, thus accommodating the overall considerations of the final clipping line and the microscopic image and ensuring the accuracy of the particle region and the information bar region.

[0022] At this time, for all candidate boundary positions detected in step S111 (i.e., the rows corresponding to the luminance gradient peaks), the pixel data of each row and its tiny neighborhood (such as 1 - 2 rows above and below) are extracted; the variance of the pixel gray values of the candidate rows is calculated; the variance reflects the consistency of pixel luminance, and the lower the variance, the more single - colored the row is (such as pure black or pure white); the luminance mutations (large gradients) caused by changes in particle density in the microscopic image are usually accompanied by relatively high texture variances; while the information bar dividing line is usually a solid - color line with extremely low variance; calculating the variance aims to distinguish these two situations.

[0023] Dual - constraint conditions: Set a judgment threshold. Only when both of the following two conditions are met simultaneously, the candidate boundary is confirmed as a valid dividing line: Gradient dimension: The inter - row luminance gradient ΔL > Tgrad (to ensure there is an obvious luminance jump); Variance dimension: The variance of row pixels σ2 < Tvar (to ensure that this position is a solid - color line rather than a particle area with complex texture). If only the gradient is large but the variance is high, it is determined as interference from texture features inside the image (such as a large - area dark region) and is excluded; if only the variance is low but the gradient is small (such as a uniform background), it is ignored and not processed. The only best row selected through the above dual - constraint screening is marked as the target row (dividing line).

[0024] Centered on the determined target row, set a tiny search neighborhood range (for example, 5 rows above to 5 rows below the target row); within this neighborhood, calculate the average luminance value of pixels row by row; select the row with the lowest average luminance (the smallest gray value) as the final cropping line; using the characteristic that there is generally a black separator bar in the microscopic image, perform "fine - tuning" near the gradient peak to ensure that the cropping line precisely falls on the black separator bar, thereby avoiding leaving pixels of the dividing line in the particle area and causing edge noise.

[0025] According to the row number Ycut of the determined final cropping line, segment the original microscopic image; retain the image area with row numbers from 11 to Ycut - 1, which is defined as the particle area (ROI), and this area contains all valid particles to be analyzed; at the same time, discard the area with row numbers from Ycut to the bottom of the image, which is defined as the information bar area (including non - particle information such as voltage and scale text).

[0026] Specifically, there is an information bar at the bottom of the microscopic image; there is a 1 - pixel - wide black horizontal separator line near the 708th row. Below the separator line is a black background and white text, and above is a microscopic image of non - uniform gray - scale nanoparticles; in addition, there is a darker area formed by particle aggregation at the 400th row, which is likely to cause misjudgment of the boundary.

[0027] The system checks the candidate boundaries found by S111; the pixel grayscale values ​​in row 708 (the position of the black dividing line) are basically all 0, and the variance σ7082 is calculated to be 0.2 (extremely low); the system also checks the candidate boundaries in row 400 (grainy dark area). Since this row contains grainy texture with varying brightness, the variance σ4002 is calculated to be 85.5 (relatively high); optionally, the variance data can be used as the preset data for this scene.

[0028] Meanwhile, line 708 is judged as follows: brightness gradient ΔL=100 (satisfying >30) and variance 0.2 (satisfying <5); the system judges it as a valid dividing line and marks it as the target line; line 400 is judged as follows: although the brightness gradient is large, the variance σ2=85.5 far exceeds the threshold, the system judges it as grain texture interference and removes it.

[0029] The system performs a fine search within the five neighboring rows above and below row 708 (i.e., rows 703 to 713); calculates the average brightness of each row: row 710 (black background for information bar): average brightness 20; row 709 (black background for information bar): average brightness 18; row 708 (black dividing line): average brightness 2 (lowest); row 707 (start of particle content area): average brightness 115; the system determines that row 708 has the lowest average brightness and locks it as the final clipping line.

[0030] The system performs a cropping operation, retaining the image data from row 1 to row 707 as the particle region for subsequent particle counting and size analysis; it automatically removes the data from row 708 to row 768, successfully stripping away the information bar area with text information at the bottom, and ensuring that the dividing line itself is not included in the analysis area, thus guaranteeing the purity of the statistical data.

[0031] Referring to Figure 3, the specific steps in step S12 are as follows: S121: The microscopic image is input into a dual-channel detection network, and two detection channels are activated to capture candidate lines with different features: Channel A uses Canny edge detection combined with Hough transform to specifically extract clear and continuous straight lines in the microscopic image; Channel B uses multi-level threshold binarization and contour extraction to process blurred or discontinuous scale lines and selects contours with aspect ratios within a preset range as candidate lines; S122: All candidate lines are comprehensively scored, and excessively long lines are penalized by a penalty mechanism, and the candidate region with the highest score is marked. The candidate region with the highest score is sent to the OCR engine for recognition to output the OCR recognition result, and the scale bar information of the microscopic image is determined based on the analysis of the OCR recognition result; If the OCR recognition result is empty or the recognized content does not conform to the physical unit characteristics, the system will automatically backtrack and select the candidate region with the second highest score to be sent back to the OCR engine for recognition.

[0032] In the embodiments of this application, the microscopic image is input into a dual-channel detection network, and two detection channels are activated to capture candidate lines with different features: Channel A uses Canny edge detection combined with Hough transform to specifically extract clear and continuous straight lines in the microscopic image, while Channel B uses multi-level threshold binarization and contour extraction to process blurred or discontinuous scale lines and selects contours with aspect ratios within a preset range as candidate lines.

[0033] At this point, the microscopic image is input into a dual-channel detection network, and two detection channels are activated to capture candidate lines with different features: Channel A: Line extraction based on Canny edge detection and Hough transform; Gaussian filtering is applied to the microscopic image to reduce noise and suppress the interference of particle noise in the image on edge detection; Canny operator is applied to extract edge features in the image; the Canny algorithm has a low error rate and high positioning accuracy, and can effectively detect strong edges between the scale bar and the background; the edge image is input into the Hough transform; the Hough transform fits collinear line segments from discrete edge points by voting in the parameter space; among the lines output by the Hough transform, preliminary screening is performed according to the characteristics of microscopic imaging: at the same time, angle constraint: lines that deviate from the horizontal direction (such as the absolute value of the rotation angle > 5∘) are filtered out, and lines that are close to the horizontal are retained; length constraint: excessively short line segments are removed, and only lines with a length exceeding the minimum pixel threshold are retained.

[0034] Channel B: Morphological processing based on multi-level threshold binarization and contour extraction; considering that some scale bars may have low contrast or be blurry, the system does not rely on a single fixed threshold, but uses a set of incremental threshold sequences (e.g., from grayscale value 50 to 200, with a step size of 10) to perform multiple binarization processes on the image; contour extraction is performed on each binarization result to obtain the outer contour of all connected white regions; based on the morphological characteristics of scale bars, which are usually long and thin rectangles, the aspect ratio of the minimum bounding rectangle of each contour is calculated; only contours with an aspect ratio greater than a preset threshold (e.g., AspectRatio>5) are retained, and interference blocks that are close to square (e.g., text blocks or particle clusters) are removed. The advantage of this channel is that even if the scale bar is visually broken, as long as the binarized contour of its remaining part conforms to the long and thin feature, it can still be extracted as a valid candidate.

[0035] The line segments extracted by channel A are merged with the contours extracted by channel B; the spatial overlap between the lines is calculated. If the spatial positions of the objects detected by the two channels are highly overlapping (e.g., IoU>0.7), they are determined to be the same object, and deduplication and merging are performed to avoid duplicate calculations; the final candidate line set is generated, which includes both clear and complete scales as well as blurry or discontinuous scales, providing a data basis for subsequent scoring decisions.

[0036] Specifically, the microscopic image contains a ruler bar; however, due to imaging reasons, the ruler bar is semi-blurred, and the lines on the right side are slightly broken; there are also a large number of noise points formed by nanoparticles and text characters in the image background. Channel A performs (clear line extraction): the system performs Canny edge detection on image A. Since the contrast of the left side of the ruler bar is still acceptable, the Canny operator successfully extracts the edge of the left line; the Hough transform fits a line segment L1L1 (pixel coordinates: [100,720]→[400,721]) based on these edge points; since the right side of the ruler bar is broken and blurred, Canny failed to detect the complete edge, so channel A only extracts the left half of the ruler bar line segment.

[0037] Channel B execution (blurred / discontinuous line extraction): The system initiates multi-level threshold scanning; when the threshold is set to 120, the blurred ruler bars in the image are binarized into thin white stripes. Although there are breaks in the middle, the overall shape is still connected or approximately connected; the contour extraction algorithm captures the contour C1 of this region; the system calculates the minimum bounding rectangle of C1, with a width of 350 pixels and a height of 4 pixels. The aspect ratio is calculated to be 350 / 4=87.5, which far exceeds the preset threshold of 5.0; at the same time, Channel B also extracts the contour of the text character "5μm", but its aspect ratio is close to 1.0, and it is discarded by the system as an invalid shape.

[0038] The system compares line segment L1 in channel A with contour C1 in channel B; it finds that L1 is completely covered by C1; the system determines that they are the same target and adds C1 (which contains a more complete ruler shape) to the final candidate line set; results: although single edge detection (channel A) is limited by the break and does not capture the whole picture, the dual-channel mechanism successfully captures the complete ruler strip candidate object through morphological analysis (channel B), avoiding missed detections due to poor image quality.

[0039] Furthermore, all candidate lines are comprehensively scored, and excessively long lines are penalized by a penalty mechanism to reduce their scores. The candidate region with the highest score is marked and sent to the OCR engine for recognition to output the OCR recognition result. The scale bar information of the microscopic image is determined based on the analysis of the OCR recognition result. If the OCR recognition result is empty or the recognized content does not conform to the physical unit characteristics, the system will automatically backtrack and select the candidate region with the second highest score to be sent to the OCR engine for recognition again. This comprehensive consideration of the analysis of OCR recognition results ensures the accuracy of the scale bar information of the microscopic image.

[0040] At this point, for all candidate lines output in step S121, extract the following key geometric features: positional features: the vertical coordinate (Y value) of the candidate line. Usually, the ruler bar is located in the bottom area of ​​the image, and the closer to the bottom and within the effective particle area, the higher the score; horizontality: the angle (θ) between the line and the horizontal axis; lines close to 0∘ score the highest; length features: the pixel length (L) of the line; lines with moderate length (within the range of conventional ruler sizes) score higher; set the weight coefficients for each feature and calculate the comprehensive score S for each candidate line; scoring formula example: S=w1·Pscore+w2·Hscore+w3·Lscore, where w is the weight; sort all candidate lines in descending order according to the comprehensive score SS to generate a candidate list.

[0041] Meanwhile, microscopic images often contain edge dividing lines that run through the left and right sides, and their length is often close to or equal to the image width, which can easily be confused with the ruler bars; a length threshold Lmax is introduced (usually set to more than 90% of the image width); for candidate lines with length L>Lmax, a large penalty is directly applied (e.g., multiplying the total score by 0.1 or directly subtracting a fixed score); even if other features of the line (such as levelness) are good, it will be ranked at the end of the candidate list due to the excessive length penalty, effectively avoiding misselection.

[0042] Select the candidate line with the highest score as the current target object; based on the coordinate position of the line, expand outward by a certain pixel range (e.g., 20 pixels above and 10 pixels below) to generate a rectangular cropping area containing the ruler line and the numerical text above it; input the cropping area into the OCR engine for text recognition; focus on extracting numerical characters and unit characters; output the recognized text string and its confidence score.

[0043] Logical verification of OCR recognition results: Non-empty verification: whether the recognition result is empty or garbled; Physical unit verification: whether the recognition result contains elements from the predefined physical unit list; If the recognition result satisfies both "non-empty" and "contains valid units", the recognition is considered successful, and the coordinates of the candidate area and the recognized content are output as the final ruler bar information; If the recognition fails (empty or without units), the system automatically backtracks to the second highest-scoring candidate line in the candidate list and repeats the process of "ROI cropping > OCR recognition > verification"; Termination condition: loop execution until a valid ruler is successfully recognized, or the candidate list is traversed.

[0044] Specifically, the microscopic image contains two main horizontal lines: the true scale bar: located at the bottom Y=700, 300 pixels long, horizontal, with "100nm" above it; and the edge separator line: located at Y=760 (above the information bar), 1024 pixels long (running through the entire image), a pure black line with no text.

[0045] Line 1 (ruler bar): Position is appropriate (+80 points), levelness is perfect (+100 points), length of 300 pixels meets expectations (+90 points); Total score: 270 points; Line 2 (separator line): Position is too low (+60 points), levelness is perfect (+100 points), length is 1024 pixels; Triggered excessively long line penalty (deduct 200 points); Total score: 60+100-200=-40 points; Line 1 is ranked 1st in the row, and Line 2 is ranked 2nd in the row.

[0046] The system locks line 1, crops the area above it, and the OCR recognizes the string "100nm". The string is not empty and contains the unit "nm" (in the unit list). Recognition is successful, and the physical length of the ruler is determined to be 100nm, the pixel length is 300px, and the pixel resolution is calculated to be 0.333nm / px.

[0047] Assuming the text above line 1 is blurry, resulting in an empty OCR output, the system detects a recognition failure and immediately initiates a backtracking mechanism. It automatically selects the second-best object in the list—line 2. The area of ​​line 2 is cropped for OCR, and the recognition result is empty (or has no physical units). The system continues to backtrack to select the third option (if any) until a valid object is found or a report of "no ruler detected" is submitted. Through this scoring and closed-loop verification mechanism, the system successfully uses geometric features to prioritize and lock onto the real ruler, effectively suppressing the interference of long dividing lines and ensuring the accurate extraction of ruler information.

[0048] Referring to Figure 4, the specific steps in step S13 are as follows: S131: Binarize the particle region and output the corresponding binarized image. Based on the combination of the binarized image and the scale bar information of the microscopic image, determine the corresponding connected components to identify and mark multiple independent particle markers; S132: Collect the processing area formed by multiple particle markers and perform hierarchical processing based on the processing area. At this time, the first level is: before performing distance transformation, mark particles with an area smaller than a specific threshold as protected objects; the second level is: perform distance transformation on the remaining adhered areas; the third level is: dynamically generate a foreground threshold based on the image signal-to-noise ratio to determine the particle center; the fourth level is: define the unknown region and perform watershed segmentation of the adhered parts to merge the segmentation results with the small particles corresponding to the protected objects and output the corresponding particle results.

[0049] In the embodiments of this application, the particle region is binarized and the corresponding binarized image is output. The corresponding connected components are determined based on the combination of the binarized image and the scale bar information of the microscopic image, so as to identify and mark multiple independent particle markers. This approach takes into account the overall combination of the binarized image and the scale bar information of the microscopic image, ensuring the accuracy of the corresponding connected components.

[0050] At this point, the particle region coordinates output in step S112 are called to crop out image data containing only effective particles from the original microscopic image, completely eliminating interference from the bottom information bar; Gaussian blur filtering (default kernel size such as 5×55×5) is applied to the cropped image to smooth high-frequency noise in the image and reduce isolated noise points in the subsequent binarization process.

[0051] Read the binarization threshold parameter in the system configuration (default is 100); if in fully automatic mode, use Otsu's maximum inter-class variance method to automatically calculate the optimal segmentation threshold, and divide the image grayscale histogram into two classes: particles (foreground) and background; traverse the image pixels, set pixels with grayscale values ​​greater than the threshold to 255 (white / foreground), and pixels with grayscale values ​​less than or equal to the threshold to 0 (black / background); generate a standard binarized image, where white areas represent potential particles.

[0052] Perform connected component analysis on the binarized image (e.g., based on 8-neighborhood or 4-neighborhood connectivity) to detect all independent white pixel blocks in the image; for each detected connected component (i.e., candidate particle), calculate its geometric features, including: pixel area: the total number of pixels in the connected component; circumscribed contour: the sequence of pixel coordinates of the edge of the connected component.

[0053] Combining the minimum particle area (e.g., 5px) and maximum particle area (e.g., 200px) parameters in the system configuration; noise removal: deleting connected components with an area smaller than the minimum threshold (usually image background noise or small impurities); background block removal: deleting connected components with an area larger than the maximum threshold (usually background afterimages that have not been completely removed or large stains); at the same time, calculating the roundness R=4π×Area / Perimeter2 for the retained connected components; deleting connected components with a roundness lower than the minimum roundness threshold (e.g., 0.5), and removing long strips or extremely irregular artifacts; assigning a unique label ID (LabelID) to all valid connected components that pass the screening, generating a label map (LabelMap) containing multiple particle labels.

[0054] Specifically, after processing by S112, the effective particle area of ​​the microscopic image is from row 00 to row 700700; the image contains a large number of spherical nanoparticles of about 50nm, as well as a small amount of background noise; the system crops out the image area of ​​row 0-7000-700; a 5×5 Gaussian blur is applied, which makes the edges of the originally fine particles in the image smooth, while the random high-frequency noise in the background is reduced.

[0055] The system uses the Otsu algorithm to calculate the grayscale histogram and determines the optimal segmentation threshold to be 118. Binarization is then performed: pixels with a grayscale value > 118 are turned white (grain), and pixels with a grayscale value ≤ 118 are turned black (background). A clear black-and-white binary image is generated, in which the grain appears as clear white spots.

[0056] The system scanned the binary image and detected 150 independent white connected regions; the area of ​​each connected region was calculated; for example: connected region #001: area 120px, complete outline; connected region #002: area 3px, point-like; connected region #003: area 5000px, large area patch-like.

[0057] The system sets a minimum area of ​​5px and a maximum area of ​​200px. Connected component #002 (area 3px) is identified as noise and removed. Connected component #003 (area 5000px) is identified as background residue and removed. Connected component #001 (area 120px) is retained. At the same time, the system sets a minimum roundness of 0.5. Connected component #001 is calculated to have a roundness of 0.85 (close to a perfect circle) and is retained. The system ultimately identified 120 valid particles and generated a unique ID (1 to 120) for each, completing the transformation from image to "particle tag set" and preparing for the hierarchical watershed processing of S132.

[0058] Furthermore, the processing area formed by multiple particle markers is collected, and hierarchical processing is performed based on this processing area. At this time, the first level is: before performing distance transformation, particles with an area smaller than a certain threshold are marked as protected objects; the second level is: distance transformation is performed on the remaining adhered areas; the third level is: foreground threshold is dynamically generated according to the image signal-to-noise ratio to determine the particle center; the fourth level is: unknown areas are defined and watershed segmentation is performed to merge the segmentation results with the small particles corresponding to the protected objects, and the corresponding particle results are output. This realizes the region stripping of the microscopic image, so as to control the particle area and the information bar area, ensuring the accuracy of multiple particle markers and further improving the accuracy of particle results.

[0059] At this point, a hierarchical processing pipeline is established; the first stage pre-separates independent small particles to avoid them being mistakenly cut or swallowed up in subsequent complex distance transformation and watershed operations; the remaining region generates seed points through distance transformation and dynamic threshold segmentation, uses the watershed algorithm to handle adhesion, and finally merges the results to achieve complete particle segmentation.

[0060] Level 1: Small Particle Protection and Extraction (Preprocessing): Perform connected component analysis on the binarized image generated in step S131, and calculate the pixel area of ​​each independent connected component (particle marker); set an area threshold AminAmin (which can be adaptively calculated based on the total number of pixels in the image or preset by the user); identify connected components with an area smaller than AminAmin as "small particle protection objects"; extract these protection objects directly from the image to be processed, move them into the final result set, and mask (remove) these small particle regions from the original image to form a "remaining image to be processed" containing only large particles and adhered regions; this prevents small particles from being covered or merged by the "potential field" of large particles in subsequent distance transformations.

[0061] Level 2: Distance Transformation of Remaining Regions (Morphological Processing): For the remaining image to be processed after removing small particles (mainly including large particles and adhered areas); perform a distance transformation algorithm on the binary image; calculate the Euclidean distance from each foreground pixel (particle pixel) in the image to the nearest background pixel; the larger the distance value of the pixel, the closer it is to the geometric center of the particle; output a grayscale distance map, where the grayscale value represents the distance from the point to the background, providing a basis for subsequent seed point extraction.

[0062] Level 3: Adjustable Foreground Threshold and Particle Center (Seed) Generation: Analyze the signal-to-noise ratio (SNR) or histogram distribution of the distance map to evaluate image quality and particle contrast; dynamically generate a foreground threshold Tfore based on the SNR, which is typically defined as a percentage of the maximum value in the distance map (e.g., 30%–50%); mark pixel regions in the distance map with gray values ​​greater than Tfore as "defined foreground" (i.e., seed points for particle centers), these seed points represent the core region of each individual particle and do not overlap.

[0063] Level 4: Unknown Region Definition, Watershed Segmentation, and Result Merging: Between the seed point (foreground) and the background, an "unknown region" is defined. This region typically corresponds to the adhered parts between particles or areas with blurred boundaries, and is the key region that the watershed algorithm needs to address. The watershed algorithm is executed. The algorithm simulates a "flooding" process, spreading outwards from different seed points and establishing a watershed at the point where the "dams" meet, thus cutting the adhered particles apart in the unknown region. The large / adhered particle results obtained from the watershed algorithm segmentation are merged with the "small particle protection objects" pre-extracted and protected in Level 1. A final complete marker map (particle result) containing all particles (independent, tiny, and separated adhered particles) is generated.

[0064] Specifically, the microscopic image contains nanoparticles of various sizes; there are a large number of independent small particles with a diameter of about 20 nm in the image, as well as clusters of large particles with a diameter of about 100 nm that are stuck together; Level 1: Small particle protection: The system calculates the area of ​​all connected regions; the protection threshold is set to 50px (corresponding to a circle with a diameter of about 8 nm); the system finds 50 independent small particles in the image, these 50 small particles are immediately marked as "protected objects", directly stored in the final result, and temporarily removed from the image to prevent them from being lost in subsequent calculations.

[0065] Level 2: Distance transformation of the remaining region: The remaining part of the image is mainly composed of several large areas of adherent clusters; the system performs distance transformation on these cluster regions; in a triangular adherent cluster composed of 3 particles, the system generates a distance map; the central region of each particle inside the cluster is highlighted (large distance value), while the boundary of the particles (valley) is dark (small distance value).

[0066] Level 3: Seed generation: The system calculates the maximum value of the distance map to be 25.0 (pixel distance); based on the image signal-to-noise ratio, the dynamic foreground threshold is set to 40% of the maximum value, i.e., 10.0; in the above triangular clusters, only the three core extreme points with a distance value > 10.0 are selected as "seed points"; the system accurately locks the independent centers of these three adhering particles.

[0067] Level 4: Watershed Segmentation and Merging: The system uses these three seed points as injection points to execute the watershed algorithm on the unknown adhesion region (the contact surface between particles); the algorithm constructs a watershed at the contact surface, successfully cutting the three particles that are stuck together; the system merges these three newly segmented large particles with the 50 small particles protected in the first step; finally, it outputs a complete particle labeling image containing 53 particles (50 small particles + 3 large particles), with no oversegmentation or missed detection.

[0068] Referring to Figure 5, in step S14, the specific steps are as follows: S141: Dynamically detect the particle result and determine the combination of various indicators during the detection. Based on the identification of each indicator combination, determine the corresponding particle statistical indicators to output multiple particle statistical indicators, including particle number, average diameter, size distribution, particle density, and roundness; S142: Mark the multiple particle statistical indicators and arrange their positions to construct an autonomous analysis report of particles involved in the microscopic image. This autonomous analysis report integrates the original image, the result image with marked particle outlines, the statistical histogram, and various parameter setting information; store the autonomous analysis report in the database and support users to view, export, delete, and overwrite detection of historical data.

[0069] In the embodiments of this application, the particle results are dynamically detected, and various index combinations are determined during the detection. Based on the identification of each index combination, the corresponding particle statistical index is determined to output multiple particle statistical indices, including particle number, average diameter, size distribution, particle density, and roundness. This takes into account the overall consideration of the identification of various index combinations and ensures the accuracy of the corresponding particle statistical indices.

[0070] At this point, the system performs dynamic detection and geometric feature extraction of particle results, traverses the final particle result labeled image output in step S132, visits each independent particle connected domain, and calculates basic geometric parameters such as pixel area and pixel perimeter for each particle.

[0071] The system performs index combination identification and physical quantity conversion, calls the scale bar information obtained in step S122 to calculate the pixel resolution (i.e., the actual physical length represented by each pixel), and calculates the physical index combination of each particle based on the resolution and geometric parameters; specifically including: obtaining the physical area by dividing the pixel area by the square of the resolution; assuming the particle is round and inferring the equivalent diameter from the area to eliminate the influence of the irregular shape of the particle on the size characterization; and calculating the roundness through a specific formula to describe the degree to which the particle is close to round, with a value range of 0 to 1, the closer to 1, the rounder the particle.

[0072] The system calculates and outputs particle statistical indicators. It directly counts the total number of effective particles NN identified in the image, calculates the arithmetic mean of the equivalent diameter of all particles as the average diameter, divides the particle diameter into preset intervals to construct a particle size distribution histogram, and calculates the number of particles per unit area, i.e., particle density, by combining the total area of ​​the particle region.

[0073] Specifically, the microscopic image has been segmented and identified 53 particles. The physical length of the scale bar is known to be 200nm, corresponding to a pixel length of 400px. Therefore, the pixel resolution is 0.5nm / px, and the effective particle area is 100μm2. The system detected particle number #01, whose pixel area is 400px2 and pixel perimeter is 80px. Particle number #02 was detected, whose pixel area is 900px2 and pixel perimeter is 120px.

[0074] For particle #01, its physical area was calculated to be 1600 nm², its equivalent diameter to be approximately 45.1 nm, and its roundness to be approximately 0.785. For particle #02, its physical area was calculated to be 3600 nm², its equivalent diameter to be approximately 67.7 nm, and its roundness to be approximately 0.785. In the particle statistical index calculation and output stage, the system counted a total of N=53 particles, calculated the average diameter Davg=58.4 nm, generated size distribution data (e.g., 15 particles in the 40-50 nm range, 20 particles in the 50-60 nm range, etc.), and calculated the particle density to be 0.53 particles / μm². The system output a statistical report containing all the above indicators (quantity, average diameter, distribution table, density, and roundness of each particle), realizing accurate quantitative analysis of the microscopic morphology of particles in the microscopic image.

[0075] Furthermore, multiple particle statistical indicators are labeled and their positions are arranged to construct an autonomous analysis report of particles in the microscopic image. This autonomous analysis report integrates the original image, the result image with labeled particle outlines, statistical histograms, and various parameter settings. The autonomous analysis report is stored in a database, and users can view, export, delete, and perform overlay detection on historical data. Simultaneously, multi-factor control is implemented for multiple particle statistical indicators to facilitate autonomous analysis of the microscopic image, thereby improving the accuracy of the autonomous analysis report of particles in the microscopic image.

[0076] At this point, the system performs statistical indicator labeling and location arrangement, standardizes and labels the multi-dimensional particle statistical indicators (such as particle number, average diameter, minimum / maximum diameter, average roundness, and particle density) output in step S141, and records the parameter configurations used in the analysis (such as binarization threshold, watershed threshold, and scale information) as metadata.

[0077] The data structure layout of a system design report typically adopts a hierarchical or block layout, including header information containing image paths, analysis timestamps, scale information (resolution), and key parameter summaries; core statistical data area displaying the above statistical indicators in tabular or key-value pair form and providing intuitive numerical references; and detailed data area containing detailed data for each particle (number, diameter, area, roundness) to support subsequent in-depth analysis.

[0078] The system performs autonomous analysis report construction, embeds the input original microscopic image into the report as a reference benchmark for the analysis object, and draws particle outlines (such as using different colors to outline) or numbers on the original image or particle region image using the particle markers output in step S132, generating a result image marked with particle outlines to intuitively display the segmentation effect.

[0079] Meanwhile, the system automatically draws a particle size distribution histogram based on the particle size distribution data. The horizontal axis represents the diameter range, and the vertical axis represents the number or frequency percentage of particles, which intuitively reflects the dispersion of particles. The original image, the result image, the binarized image, and the histogram are stitched together into a comprehensive display image (Figure.png) in a four-grid or custom layout for easy and quick viewing.

[0080] The system stores reports and manages the database. After the analysis is completed, it automatically creates a subfolder named with a timestamp and pattern in the database folder (e.g., Automatic_sampleA_20250915_143052), and writes the original image, the color result image with outlines marked, the comprehensive analysis chart containing multiple views, and the structured data file containing the summary and single-particle data into this folder. In addition, the system performs an overwrite detection before writing to check whether there are historical records of images with the same name in the database. If they exist, the system automatically backs up or overwrites the old data to ensure the consistency between the analysis results and the current image status.

[0081] The system provides user interaction interfaces, including a graphical user interface (GUI) that displays a list of all historical analysis records for users to view historical data; a function that allows users to instantly load and display the corresponding comprehensive analysis chart and statistical data when clicking on a record; an export function that allows users to export the selected analysis results folder (including images and CSV data) to a user-specified local path with one click; a deletion management function that allows users to select one or more records to delete in order to free up database storage space; and a search function that supports quick retrieval and filtering of historical reports based on image name, time range, or specific parameters.

[0082] Specifically, the image (filename Sample_A_SEM.GIF) has undergone full-process analysis, identifying 53 particles with an average diameter of 58.4 nm. The system extracted data with a quantity of 53, an average diameter of 58.4 nm, and a roundness of 0.78. These indicators, along with the current time (2025-09-15 14:30:52) and parameters (threshold = 120, watershed = on), were packaged into the header summary information of the report.

[0083] The system reads Sample_A_SEM.GIF as the original image, reads the particle marker data, and draws green outlines on the image to generate Sample_A_result.png to clearly display the identified particles. Based on the diameter distribution of the 53 particles, it draws a statistical histogram showing that the peak values ​​are concentrated in the 50-60nm range. The system then stitches the original image, the binarized image, the outline marker image, and the size distribution histogram into a large four-grid image, Sample_A_figure.png.

[0084] The system creates a folder in the database directory, writes the aforementioned files and a statistics.csv file containing detailed data into this folder, and performs a coverage check. When an old record of the image from yesterday is detected in the database, an automatic pop-up prompt is displayed, and an overwrite update is performed to ensure that the current data is the latest version. During the user interaction phase, the user opens the system's "Results Management" interface and sees a new record in the list. After clicking on the record, the right panel immediately displays the generated Sample_A_figure.png, allowing the user to intuitively see the analysis results. Clicking the "Export" button, the system packages and copies the entire folder to the desktop, making it convenient for the user to directly insert statistical charts into the experimental report document.

[0085] Referring to Figure 7, in another embodiment of this application, SEM / TEM microscopic images are used as input, and a series of modular processing flows are used to finally output particle analysis results. The core processing flow, from left to right, consists of: a system initialization module, an image preprocessing module, a scale bar automatic recognition module, a particle recognition and analysis module, and an interactive editing module. Each module contains specific sub-functions. For example, the image preprocessing module is responsible for path compatibility processing, format conversion, noise filtering, and region detection. The scale bar automatic recognition module achieves automatic scale bar recognition through Hough transform detection, OCR text recognition, multi-strategy matching, and scale calibration. The particle recognition and analysis module includes steps such as adaptive threshold segmentation, contour extraction and filtering, watershed algorithm separation, and feature parameter calculation. In addition, the data storage and export module runs through the entire processing flow and is responsible for managing intermediate results and final data in the form of database storage, CSV export, image saving, and chart generation. Finally, all processing results are summarized into the analysis result output, generating particle size distribution, statistical data, and labeled images. The entire architecture design reflects the full automation of the process from image input to result output, while retaining interactive editing functions to ensure the accuracy and flexibility of the analysis.

[0086] Referring to Figure 8, in another embodiment of this application, the process involves "loading a microscopic image." The workflow then enters the "image preprocessing" stage, which includes three key steps: reading the grayscale image, performing Gaussian blur noise reduction, and then automatically identifying the scale bar using Hough transform and OCR technology. Simultaneously, the image preprocessing stage also includes "analysis region detection" to exclude the information bar area at the bottom of the image and performing binarization thresholding on the image. After preprocessing, the workflow enters the "particle recognition" stage. The system selects a "segmentation method" based on particle adhesion. If "watershed" is selected, the watershed algorithm is used for segmentation, followed by contour detection and extraction. If "traditional" is selected, traditional thresholding is performed directly. The identified particles then enter the "particle filtering" stage, where invalid particles are filtered by setting an area range and a roundness threshold. The filtered particles then enter the "statistical calculation" stage, where the area of ​​each particle (in nm) is calculated sequentially. 2 The parameters were calculated as follows: equivalent diameter (in nm), sphericity, and finally, particle density (in particles / μm). 2 Finally, the process enters the "results output" stage, presenting the analysis results in the form of visualization (contour annotation map), granular distribution histogram, database storage, and CSV data export, until the process ends. The entire flowchart clearly shows the fully automated analysis steps from image loading to result output, highlighting core innovations such as adaptive region stripping, automatic scale recognition, hierarchical watershed segmentation, and automated statistics.

[0087] Please refer to Figure 6, which is a schematic diagram of the structure of the particle analysis system for silicon wafer microscopic images based on dynamic windows in an embodiment of the present invention. The particle analysis system for silicon wafer microscopic images based on dynamic windows is applied to the above-mentioned autonomous analysis method for microscopic images. The particle analysis system for silicon wafer microscopic images based on dynamic windows includes: a region stripping module 21, used to acquire microscopic images and determine particle regions and information bar regions based on the region stripping of the microscopic images; an image recognition module 22, used to determine the scale bar information of the microscopic images based on the recognition of the microscopic images; a particle result module 23, used to determine multiple particle markers based on the scale bar information and particle region detection of the microscopic images, and selectively trigger threshold segmentation or hierarchical watershed method based on the multiple particle markers to output the corresponding particle results; and an autonomous analysis module 24, used to determine multiple particle statistical indicators in the microscopic images based on the detection of the particle results, and construct an autonomous analysis report of particles involved in the microscopic images along the multiple particle statistical indicators.

[0088] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for autonomous analysis of microscopic images, characterized in that, include: Acquire microscopic images, and determine the particle region and information bar region based on the region stripping of the microscopic images; The scale bar information of the microscopic image is determined based on the identification of the microscopic image; Multiple particle markers are determined based on the scale bar information and particle region detection of the microscopic image. Threshold segmentation or hierarchical watershed method is selectively triggered based on multiple particle markers to output the corresponding particle results. Based on the detection of these particle results, multiple particle statistical indicators in the microscopic image are determined, and an autonomous analysis report of the particles involved in the microscopic image is constructed along these multiple particle statistical indicators.

2. The autonomous analysis method for microscopic images according to claim 1, characterized in that, The acquisition of the microscopic image, and the determination of the particle region and information bar region based on the region stripping of the microscopic image, includes: acquiring the microscopic image and scanning the microscopic image. At this time, the microscopic image is scanned line by line from the bottom upwards, and the pixel features of each line are analyzed. The row brightness gradient is detected for the pixel features of each line. The brightness gradient peak is determined by the average brightness difference within the sliding window, and the position corresponding to the brightness gradient peak is initially located as the candidate boundary.

3. The autonomous analysis method for microscopic images according to claim 2, characterized in that, The acquisition of the microscopic image, and the determination of the particle region and information bar region based on the region stripping of the microscopic image, further includes: marking the variance of the row pixels of each segmentation line in each candidate boundary, and applying multiple constraints based on the gradient dimension and variance dimension to determine the target row corresponding to the segmentation line, and determining the final clipping line by combining multiple neighborhoods of the target row, and determining the particle region and information bar region based on the final clipping line and the microscopic image. The particle region refers to the region in the microscopic image that contains the particles to be analyzed; the information bar region refers to the device information region usually contained at the bottom of the microscopic image, and the information bar region does not contain any particles to be analyzed.

4. The autonomous analysis method for microscopic images according to claim 1, characterized in that, The step of determining the scale bar information of the microscopic image based on the recognition of the microscopic image includes: inputting the microscopic image into a dual-channel detection network and activating two detection channels to capture candidate lines with different features: Channel A uses Canny edge detection combined with Hough transform to specifically extract clear and continuous straight lines in the microscopic image, while Channel B uses multi-level threshold binarization and contour extraction to process blurred or discontinuous scale lines and selects contours with aspect ratios that meet the preset range as candidate lines.

5. The autonomous analysis method for microscopic images according to claim 4, characterized in that, The process of determining the scale bar information of the microscopic image based on the recognition of the microscopic image further includes: comprehensively scoring all candidate lines, and using an excessively long line penalty mechanism to reduce the score of excessively long lines, marking the candidate region with the highest score, sending the candidate region with the highest score into the OCR engine for recognition, outputting the OCR recognition result, and determining the scale bar information of the microscopic image based on the analysis of the OCR recognition result; if the OCR recognition result is empty or the recognized content does not conform to the physical unit characteristics, the process will automatically backtrack and select the candidate region with the second highest score to be re-sent into the OCR engine for recognition.

6. The autonomous analysis method for microscopic images according to claim 1, characterized in that, The process of determining multiple particle markers based on the scale bar information of the microscopic image and the detection of particle regions, and selectively triggering threshold segmentation or hierarchical watershed methods based on multiple particle markers to output corresponding particle results includes: binarizing the particle region image and outputting the corresponding binarized image; determining the corresponding connected components based on the combination of the binarized image and the scale bar information of the microscopic image, so as to identify and mark multiple independent particle markers.

7. The autonomous analysis method for microscopic images according to claim 6, characterized in that, The process of determining multiple particle markers based on the scale bar information and particle region detection of the microscopic image, and selectively triggering threshold segmentation or hierarchical watershed methods based on multiple particle markers to output corresponding particle results, further includes: acquiring the processing area formed by multiple particle markers, and performing hierarchical processing based on the processing area. The first level involves marking particles with areas smaller than a specific threshold as protected objects before performing distance transformation; the second level involves performing distance transformation on the remaining adhered areas; the third level involves dynamically generating a foreground threshold based on the image signal-to-noise ratio to determine the particle center; and the fourth level involves defining unknown regions and performing watershed segmentation on the adhered parts to merge the segmentation results with the small particles corresponding to the protected objects, and outputting the corresponding particle results.

8. The autonomous analysis method for microscopic images according to claim 1, characterized in that, The process of determining multiple particle statistical indicators in the microscopic image based on the particle results and constructing an autonomous analysis report of particles involved in the microscopic image along the multiple particle statistical indicators includes: dynamically detecting the particle results and determining various indicator combinations during the detection; determining the corresponding particle statistical indicators based on the identification of each indicator combination; and outputting multiple particle statistical indicators, including particle number, average diameter, size distribution, particle density, and roundness.

9. The autonomous analysis method for microscopic images according to claim 8, characterized in that, The process of determining multiple particle statistical indicators in a microscopic image based on the particle results, and constructing an autonomous analysis report of particles involved in the microscopic image along these multiple particle statistical indicators, further includes: marking multiple particle statistical indicators and arranging their positions to construct an autonomous analysis report of particles involved in the microscopic image. This autonomous analysis report integrates the original image, the result image with marked particle outlines, statistical histograms, and various parameter setting information; storing the autonomous analysis report in a database, and supporting users to view, export, delete, and overwrite detections of historical data.

10. A particle analysis system based on microscopic images of silicon wafers using a dynamic viewing window, characterized in that, The particle analysis system for silicon wafer microscopic images based on dynamic windows is applied to the autonomous analysis method for microscopic images as described in any one of claims 1-9. The particle analysis system for silicon wafer microscopic images based on dynamic windows includes: a region stripping module for acquiring microscopic images and determining particle regions and information bar regions based on the region stripping of the microscopic images; an image recognition module for determining the scale bar information of the microscopic images based on the recognition of the microscopic images; a particle result module for determining multiple particle markers based on the scale bar information and particle region detection of the microscopic images, and selectively triggering threshold segmentation or hierarchical watershed methods based on the multiple particle markers to output the corresponding particle results; and an autonomous analysis module for determining multiple particle statistical indicators in the microscopic images based on the particle result detection, and constructing an autonomous analysis report of the particles involved in the microscopic images along the multiple particle statistical indicators.

Images