A method and system for intelligent identification and diagnosis of defects of a VCSEL chip lens

Abstract

The present application relates to a kind of VCSEL chip lens defect intelligent identification and diagnosis method and system, it is related to defect identification technical field, including the following steps, by polarized light source irradiation is carried out optical imaging to the VCSEL chip lens, obtains polarization reflection image;Based on the polarization reflection image carries out gray threshold segmentation, obtains independent lens area;To the independent lens area carries out multi-scale pixel matrix convolution, obtains local edge profile, and based on the local edge profile determines whether the VCSEL chip lens exists defect geometric feature;If there is, then based on the defect geometric feature carries out wafer physical coordinate mapping, obtains defect distribution atlas;Based on the defect distribution atlas is carried out defect type diagnosis to the VCSEL chip lens, generates defect type report, solves the technical problem that significant missed detection and misjudgment exist in micron level defect identification in traditional technology.

Description

Technical Field

[0001] This invention relates to the field of defect identification technology, and in particular to a method and system for intelligent identification and diagnosis of defects in VCSEL chip lenses. Background Technology

[0002] In the manufacturing process of VCSEL (Vertical-Cavity Surface-Emitting Laser) chips, the integrity of the lens structure has a decisive impact on device performance. Currently, the industry commonly employs optical microscopy combined with manual visual inspection or traditional image processing algorithms to detect defects in VCSEL chip lenses. A typical approach involves illuminating the wafer surface with ordinary white light or an LED light source, acquiring the reflected image using a high-resolution camera, and then extracting the lens contour using fixed-threshold segmentation or edge detection operators (such as Canny and Sobel) to determine the presence of anomalies such as cracks, chipping, or deformation. This method has been widely applied in wafer-level appearance inspection before semiconductor packaging, and is particularly suitable for initial screening in high-volume production lines.

[0003] However, the inventors discovered that because VCSEL chip lenses are typically made of highly reflective materials and have complex surface curvatures, they are prone to strong glare, specular reflection, or local overexposure under normal non-polarized illumination conditions, causing key edge information in the image to be obscured or distorted. This makes it difficult for traditional edge detection methods based on grayscale gradients to reliably extract the true geometric contours of the lens, especially in the identification of micron-level defects (such as subpixel cracks or slight collapses), resulting in significant missed detections and false positives. Summary of the Invention

[0004] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides a method for intelligent identification and diagnosis of defects in VCSEL chip lenses, comprising: Optical imaging is performed on the lens of the VCSEL chip by illuminating it with a polarized light source to obtain a polarized reflection image; Based on the polarization reflection image, grayscale threshold segmentation is performed to obtain independent lens regions; Multi-scale pixel matrix convolution is performed on the independent lens region to obtain local edge contours, and the presence of defective geometric features in the VCSEL chip lens is determined based on the local edge contours. If present, a wafer physical coordinate mapping is performed based on the defect geometric features to obtain a defect distribution map; Based on the defect distribution map, the VCSEL chip lens is used to diagnose the defect type and generate a defect type report.

[0006] Furthermore, optical imaging of the VCSEL chip lens is performed using a polarized light source to obtain a polarized reflection image, including: The surface of the VCSEL chip lens is obliquely illuminated by a linearly polarized light source at an incident angle of 45° to obtain the difference in reflected light intensity between the s-polarization component and the p-polarization component. Based on the difference in reflected light intensity, polarization differential imaging is performed on the VCSEL chip lens to obtain an initial polarization reflection image. The initial polarized reflection image is subjected to background light suppression processing to obtain a background light distribution map, and the background light distribution map is subtracted pixel by pixel from the initial polarized reflection image to obtain the polarized reflection image.

[0007] Furthermore, based on the polarization reflection image, grayscale thresholding is performed to obtain independent lens regions, including: Histogram analysis of the polarization reflection image yields a bimodal distribution curve. Based on the bimodal distribution curve, the polarization reflection image is truncated at the valley bottom pixels to obtain the initial screening lens mask; The pixel distance transformation is performed on the initial screening lens mask to obtain a center distance matrix, and the watershed boundary is blocked based on the center distance matrix to obtain an independent lens region.

[0008] Furthermore, based on the center distance matrix, a watershed boundary is blocked on the initial screening lens mask to obtain an independent lens region, including: The center distance matrix is ​​subjected to a maximum search to obtain local extreme value pixels, and the local extreme value pixels are independently assigned numerical values ​​to obtain a seed label matrix; The initial screening lens mask is expanded layer by layer based on the seed tag matrix to obtain the pixel expansion region, and adjacent collision detection is performed based on the pixel expansion region to obtain the intersection boundary line. Based on the intersection boundary line, the background pixels of the initial screening lens mask are replaced to obtain a separation mask image, and the connected components of the separation mask image are extracted to obtain independent lens regions.

[0009] Furthermore, multi-scale pixel matrix convolution is performed on the independent lens region to obtain local edge contours, including: Spatial sliding convolution is performed on the independent lens regions to obtain a multi-scale smooth image; Based on the multi-scale smoothed image, cross-scale gray-level subtraction is performed on the independent lens region to obtain a multi-scale difference image; Multi-directional edge matrix convolution is performed on the multi-scale difference image to obtain a multi-directional edge feature map, and local extremum filtering is performed on the multi-scale difference image based on the multi-directional edge feature map to obtain discrete defect pixels. The discrete defect pixels are merged to obtain a continuous defect region, and the outer boundary of the continuous defect region is tracked to obtain the local edge contour.

[0010] Furthermore, the multi-scale difference image is convolved with a multi-directional edge matrix to obtain a multi-directional edge feature map, including: The first-order derivative convolution kernel sliding calculations were performed on the multi-scale difference image in four directions: 0°, 45°, 90°, and 135°, to obtain four sets of directional response maps. Pixel-level absolute value calculations are performed on the four sets of directional response maps to obtain four sets of directional gradient magnitude maps; Based on the four sets of directional gradient magnitude maps, pixel-wise maximum value fusion is performed to obtain an anisotropic response map, and neighborhood mean smoothing is performed based on the anisotropic response map to obtain a multi-directional edge feature map.

[0011] Furthermore, the outer boundary of the continuous defect region is traced to obtain the local edge contour, including: The continuous defect region is scanned pixel by pixel to obtain the outermost starting pixel, and the continuous defect region is searched clockwise based on the outermost starting pixel to obtain the initial closed boundary line. The initial closed boundary line is filtered by vertex distance threshold to obtain a set of inflection point coordinates. Based on the set of inflection point coordinates, spline curve interpolation is performed on the initial closed boundary line to obtain the local edge contour.

[0012] Furthermore, based on the aforementioned defect geometric features, wafer physical coordinate mapping is performed to obtain a defect distribution map, including: The local centroid coordinates of the defect geometric features are extracted to obtain a single lens centroid coordinate set, and the array period is fitted to the single lens centroid coordinate set to obtain an array reference coordinate grid. Based on the array reference coordinate grid, the single lens centroid coordinate set is indexed by row and column to obtain an indexed defect coordinate table. Align the indexed defect coordinate table with a wafer-level coordinate system to obtain a defect distribution map.

[0013] Furthermore, based on the defect distribution map, defect type diagnosis is performed on the VCSEL chip lens, and a defect type report is generated, including: Cluster the defect coordinates in the defect distribution map, and group the coordinates of each defect coordinate that are less than the preset neighborhood radius into the same defect cluster area to obtain at least one defect cluster area. For each of the defect clusters, a circumscribed contour is fitted to obtain the geometry of the defect cluster and the dimensional parameters of its circumscribed contour. The geometric shape is matched with multiple standard process geometric templates in a preset process defect map library, and the contour overlap between each standard process geometric template and the geometric shape is calculated. Select the process defect category corresponding to the standard process geometry template with the highest contour overlap as the process defect category of the current defect cluster area; Based on the process defect category and the dimensional parameters of the outer contour of the defect cluster area and the defect coordinate points, diagnostic information is arranged to generate a defect type report.

[0014] This invention also provides a defect intelligent identification and diagnosis system for VCSEL chip lenses, comprising: The imaging module is used to perform optical imaging on the lens of the VCSEL chip by illuminating it with a polarized light source to obtain a polarized reflection image; The segmentation module is used to perform grayscale threshold segmentation based on the polarization reflection image to obtain independent lens regions; The judgment module is used to perform multi-scale pixel matrix convolution on the independent lens region to obtain local edge contours, and determine whether the VCSEL chip lens has defective geometric features based on the local edge contours. The mapping module is used to perform wafer physical coordinate mapping based on the defect geometric features if the defect exists, to obtain a defect distribution map; The diagnostic module is used to diagnose the defect type of the VCSEL chip lens based on the defect distribution map and generate a defect type report.

[0015] This invention provides a method for intelligent defect identification and diagnosis of VCSEL chip lenses, comprising the following steps: optical imaging of the VCSEL chip lens by irradiation with a polarized light source to obtain a polarized reflection image; grayscale threshold segmentation based on the polarized reflection image to obtain independent lens regions; multi-scale pixel matrix convolution of the independent lens regions to obtain local edge contours, and determining whether the VCSEL chip lens has defect geometric features based on the local edge contours; if so, wafer physical coordinate mapping based on the defect geometric features to obtain a defect distribution map; and defect type diagnosis of the VCSEL chip lens based on the defect distribution map to generate a defect type report. This method solves the technical problem of significant missed detections and misjudgments in the identification of micron-level defects in traditional technologies, and achieves automated intelligent detection of micron-level defects in VCSEL chip lenses with high sensitivity, high accuracy, localization, and classification, effectively overcoming the bottlenecks of high missed detection rates and high misjudgment rates in traditional detection technologies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of a method for intelligent identification and diagnosis of defects in a VCSEL chip lens according to an embodiment of the present invention; Figure 2 In this invention Figure 1 A schematic diagram of step S1 in the middle; Figure 3 In this invention Figure 1 A schematic diagram of step S2 in the middle; Figure 4 This is a structural block diagram of the intelligent defect identification and diagnosis system for VCSEL chip lenses in an embodiment of the present invention.

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

[0019] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0020] The following describes in detail, with reference to the accompanying drawings, a method for intelligent identification and diagnosis of defects in a VCSEL chip lens according to an embodiment of the present invention.

[0021] Figure 1 This invention provides a method for intelligent identification and diagnosis of defects in a VCSEL chip lens, comprising: Step S1: Optical imaging is performed on the VCSEL chip lens by illuminating it with a polarized light source to obtain a polarized reflection image.

[0022] Specifically, in implementation, a linearly polarized light source is used to illuminate the surface of the VCSEL chip lens at a specific incident angle. The polarization direction of this light source is orthogonal or parallel to the analyzer in the imaging optical path, thereby suppressing stray light reflected from the specular surface. A high-resolution CMOS camera simultaneously acquires the intensity distribution of the polarized light reflected from the lens surface, forming a polarized reflection image. For example, an image is acquired at a wavelength of 650 nm with a 45° illumination angle and a 0° analysis direction, effectively highlighting the details of the microstructure edges. In this embodiment, the above scheme can significantly reduce the overexposure effect in highly reflective areas, improving the signal-to-noise ratio and contour extraction accuracy of subsequent image segmentation.

[0023] Step S2: Perform grayscale threshold segmentation based on the polarized reflection image to obtain independent lens regions.

[0024] Specifically, for the aforementioned polarized reflection image, it is first converted into an 8-bit grayscale image. Then, the Otsu adaptive thresholding method is used to calculate the global segmentation threshold. If the grayscale of the lens region in the image is concentrated in the range of 180~240, it is binarized using this dynamic threshold. Subsequently, noise patches with an area of ​​less than 50 pixels are removed by connected component labeling, retaining independent regions that conform to the geometric dimensions of the VCSEL chip lens. If necessary, morphological opening operations are used to eliminate edge burrs. In this embodiment, the above scheme can effectively separate adjacent lens units, avoid regional adhesion caused by reflection crosstalk, and provide a clear input mask for subsequent multi-scale convolution.

[0025] Step S3: Perform multi-scale pixel matrix convolution on the independent lens region to obtain local edge contours, and determine whether the VCSEL chip lens has defective geometric features based on the local edge contours.

[0026] Specifically, the independent lens regions are input into a multi-scale convolution module, where pixel matrix convolutions are performed sequentially using 3×3, 5×5, and 7×7 Sobel kernels in the horizontal and vertical directions to extract gradient responses at different scales. Subsequently, the edge maps at each scale are weighted and fused, and sub-pixel-level contour points are located using non-maximum suppression. If a region's contour curvature abruptly exceeds a threshold (e.g., a local curvature radius less than 2 μm) or exhibits a non-closed fracture, a defective geometric feature is identified. In this embodiment, the above scheme can effectively capture micron-level edge chipping or cracks and other structural anomalies, avoiding the loss of detail caused by single-scale filtering.

[0027] Step S4: If the defect exists, perform wafer physical coordinate mapping based on the defect geometric features to obtain a defect distribution map.

[0028] Specifically, after determining the presence of defect geometric features, the system uses the position data fed back by the wafer stage encoder and the center of the field of view during imaging to transform the defect pixel coordinates to the wafer global coordinate system through an affine transformation matrix. The pixel spacing is converted to physical size according to the objective lens magnification (e.g., 50× corresponds to 0.2 μm / pixel), and associated with the wafer ID and die row and column number to generate a geographically tagged defect record. Multiple defect points are spatially interpolated to form a two-dimensional heatmap, i.e., a defect distribution map. In this embodiment, the above scheme achieves precise positioning and visual mapping of micron-level defects on an 8-inch VCSEL wafer, facilitating subsequent process traceability and yield analysis.

[0029] Step S5: Based on the defect distribution map, perform defect type diagnosis on the VCSEL chip lens and generate a defect type report.

[0030] Specifically, based on the geometric characteristic parameters of each defect in the defect distribution map, such as area, perimeter, rate of curvature change, and spatial arrangement pattern, it is compared with a preset defect template library. This template library includes typical types such as edge chipping, cracks, pits, and foreign matter attachment. For example, when a defect region exhibits a high curvature abrupt change and its length is greater than 5 μm and its width is less than 1 μm, it is classified as a microcrack. The system matches point by point according to wafer coordinates and summarizes the statistics to generate a structured report containing defect category, quantity, and location index. In this embodiment, the above scheme can realize the automatic classification and source tracing of VCSEL chip lens defects, significantly improving the efficiency of failure analysis and the accuracy of process feedback.

[0031] In a specific embodiment, such as Figure 2 As shown, optical imaging of the VCSEL chip lens is performed using a polarized light source to obtain a polarized reflection image, including: S11, the surface of the VCSEL chip lens is obliquely illuminated by a linearly polarized light source at an incident angle of 45° to obtain the difference in reflected light intensity between the s-polarization component and the p-polarization component, and polarization differential imaging is performed on the VCSEL chip lens based on the difference in reflected light intensity to obtain an initial polarization reflection image. S12, perform background light suppression processing on the initial polarization reflection image to obtain a background light distribution map, and subtract the background light distribution map pixel by pixel from the initial polarization reflection image to obtain a polarization reflection image.

[0032] Specifically, to obtain a polarized reflection image by optically imaging the VCSEL chip lens using a polarized light source, an illumination and imaging system with controllable polarization must first be constructed. This system uses a linearly polarized light source as the illumination source, whose emitted light, after passing through a polarizer, forms a beam with a fixed polarization direction. This beam is obliquely incident on the surface of the VCSEL chip lens at a 45° incident angle. This angle is chosen based on a comprehensive consideration of the reflection characteristics of the s-polarization component and the p-polarization component at the semiconductor material interface, effectively stimulating the reflection difference between the two while avoiding signal attenuation or saturation due to excessively small or large incident angles.

[0033] During illumination, a rotating analyzer is positioned at the s-direction perpendicular to the incident plane and the p-direction within the incident plane, respectively, to sequentially acquire two sets of reflection images. Since VCSEL chip lenses are typically made of high-refractive-index compound semiconductor materials, their curved surface structure exhibits different Fresnel reflection behaviors under different polarization states. Particularly at the lens edges or in regions with abrupt microstructural changes, p-polarized light exhibits significantly reduced reflectivity due to its proximity to the Brewster angle, while s-polarized light maintains a relatively high reflection intensity. This physical mechanism results in a noticeable brightness difference between the two sets of images in areas with defects or geometric anomalies.

[0034] Subsequently, the two sets of reflection images corresponding to the s-polarization and p-polarization components, respectively, are subjected to pixel-by-pixel difference processing, i.e., the p-polarization image is subtracted from the s-polarization image to generate the initial polarized reflection image. This difference operation can effectively highlight the polarization-sensitive features caused by changes in surface morphology, while suppressing common-mode background information in uniform material regions, making potential defect areas more clearly identifiable in the image.

[0035] To further improve image quality, background light suppression processing needs to be performed on the initial polarization reflection image. Specifically, a set of blank field images can be acquired using the same optical path and parameters without placing the sample, or a slowly changing background light distribution map can be extracted by applying morphological top-hat transformation combined with low-pass filtering to the initial image itself. This background map mainly reflects non-sample-related interference factors such as uneven light source illumination, lens vignetting, and ambient stray light.

[0036] Finally, the obtained background light distribution map is subtracted pixel by pixel from the initial polarized reflectance image, and the result is cropped in grayscale to ensure that all pixel values ​​are within the effective display range, thus obtaining the polarized reflectance image. This image not only preserves the surface details revealed by the original polarization difference but also significantly reduces the impact of systematic background noise.

[0037] In this embodiment, the above scheme effectively enhances the visual contrast of micron-level geometric anomalies on the VCSEL chip lens surface by introducing a polarization differential imaging mechanism and combining it with a background light correction strategy, providing high-quality, high signal-to-noise ratio image input for subsequent defect contour extraction and type diagnosis.

[0038] In a specific embodiment, such as Figure 3 As shown, grayscale thresholding is performed on the polarized reflection image to obtain independent lens regions, including: S21, perform histogram statistics on the polarization reflection image to obtain a bimodal distribution curve; S22, based on the bimodal distribution curve, the polarization reflection image is truncated at the valley bottom pixel to obtain the initial screening lens mask; S23, perform pixel distance transformation on the initial screening lens mask to obtain a center distance matrix, and perform watershed boundary blocking on the initial screening lens mask based on the center distance matrix to obtain an independent lens region.

[0039] Specifically, grayscale thresholding is performed on the polarized reflection image to extract independent lens regions. The specific implementation process is as follows: First, grayscale histogram statistics are performed on the corrected polarized reflection image. Due to the high reflectivity of the VCSEL chip lens surface and the weak reflection of the background area (such as the wafer passivation layer or gaps), the image grayscale distribution usually exhibits a clear bimodal shape: one peak corresponds to the high-brightness area of ​​the lens, and the other corresponds to the low-reflection background. This bimodal distribution curve can be obtained through a standard histogram accumulation algorithm, without the need for a preset model, relying only on pixel grayscale frequency statistics.

[0040] After obtaining the bimodal distribution curve, the valley position between the two peaks is identified, representing the lowest probability point of gray-level transition between the two regions. The gray value corresponding to this valley is used as a truncation threshold to binarize the original polarization reflection image; all pixels with gray values ​​higher than this threshold are set to 1 (foreground), and the rest are set to 0 (background), thus generating a preliminary lens mask. Although this mask can roughly outline the area where the lens is located, due to the small spacing between adjacent lenses or uneven local reflection, multiple lenses often clump together, making it unsuitable for direct analysis of individual devices.

[0041] To address the adhesion problem, further processing of the initial lens mask is required. This involves introducing a pixel distance transformation operation: starting with each foreground pixel in the mask, the Euclidean distance to the nearest background pixel is calculated, generating a center distance matrix of the same size as the original image. This matrix forms a local maximum "peak" within each lens region, typically located near the geometric center. This center distance matrix is ​​then used as the input topographic map for the watershed algorithm. However, to avoid over-segmentation by traditional watershed algorithms, a boundary blocking strategy is applied between local minima—that is, boundary lines are only allowed to be generated in regions where distance values ​​decrease significantly, while suppressing false segmentation in smooth transition areas. Specifically, by detecting local maxima points in the center distance matrix and using them as seed points for region growing, while setting blocking barriers at steep gradient descents between adjacent seed points, the originally adhered lens regions are effectively separated.

[0042] In practical applications, if two lenses in a VCSEL array slightly overlap due to manufacturing deviations, they will appear as a single connected region in the initial screening mask. After distance transformation, two distinct peaks will appear inside. The watershed boundary blocking mechanism will then define a precise boundary between the two peaks, ultimately outputting separate independent lens regions.

[0043] In this embodiment, the above scheme achieves accurate individual separation of high-density VCSEL chip lens arrays by combining histogram double-peak valley truncation, pixel distance transformation and controlled watershed segmentation, avoiding the erroneous merging caused by adhesion in the traditional threshold method, and providing a reliable regional positioning basis for subsequent lens-by-lens defect analysis.

[0044] In a specific embodiment, histogram statistics are performed on the polarization reflection image to obtain a bimodal distribution curve, including: The polarization reflection image is globally accumulated to obtain an initial gray-level histogram, and the initial gray-level histogram is smoothed by a sliding window mean to obtain a smoothed frequency curve. Local maxima search is performed on the smoothed frequency curve to obtain the primary and secondary bimodal extreme points; Based on the primary and secondary bimodal extreme points, the smooth frequency curve is truncated to obtain a bimodal distribution curve.

[0045] Specifically, the polarization reflection image is subjected to histogram statistics to obtain a bimodal distribution curve. This process involves several progressive operations. First, a global grayscale frequency accumulation is performed on the entire polarization reflection image: all pixels in the image are traversed, and the frequency of each grayscale level (typically 0 to 255) is counted to construct an initial grayscale histogram. This histogram directly reflects the distribution density of different brightness components in the image. However, due to imaging noise, local reflection fluctuations, or quantization errors, the original frequency data often exhibits jagged fluctuations, which is not conducive to subsequent extreme point identification.

[0046] To suppress such high-frequency disturbances, a sliding window mean smoothing process needs to be applied to the initial grayscale histogram. Specifically, a window of fixed width (e.g., 5 grayscale levels) is selected and slid along the grayscale axis point by point. At each position, the arithmetic mean of the frequencies within the window is calculated, and this mean is used to replace the original frequency of the center grayscale level, ultimately generating a continuous and clearly trending smooth frequency curve. The choice of window size must balance noise suppression and peak shape preservation—too large a window will cause the double peaks to merge, while too small a window will result in insufficient smoothing. In VCSEL chip lens imaging scenarios, due to the high contrast between the lens area and the background, a window of 3 to 7 levels is usually sufficient to effectively preserve the double-peak structure.

[0047] Subsequently, a local maximum search is performed on the resulting smoothed frequency curve. This operation is accomplished by comparing the frequency value corresponding to each gray level with its neighboring values ​​(e.g., the two gray levels before and after it): if the frequency of a point is higher than its left and right neighbors, it is determined to be a local maximum. In a typical VCSEL polarization reflection image, due to the sharp contrast between the high reflectivity of the lens surface and the surrounding low reflectivity area, the smoothed curve usually presents two significant peaks—one corresponding to the lens body (primary peak), and the other corresponding to the background or passivation layer area (secondary peak). These two extreme points are the primary and secondary bimodal extreme points, and their gray level positions are denoted as G1 and G2 (G1>G2), respectively.

[0048] After identifying the primary and secondary bimodal extreme points, the smoothed frequency curve is truncated using the grayscale intervals corresponding to these two points as boundaries. Specifically, the curve segment from slightly below the left valley of the secondary peak to slightly above the right valley of the primary peak is retained, while the low-frequency tails at both ends are removed, thus focusing on the effective distribution area truly involved in threshold segmentation. This truncation operation not only compresses irrelevant grayscale ranges but also enhances the discriminability of the bimodal features. The final output is the bimodal distribution curve used for subsequent valley detection.

[0049] For example, in the inspection of a batch of GaAs-based VCSEL wafers, after the polarization reflection image was processed by the above process, the main peak was located near gray level 180, and the secondary peak was located near gray level 60. There was a clear concave valley between the two, which provided a reliable basis for subsequent precise threshold setting.

[0050] In this embodiment, the above scheme effectively extracts the bimodal distribution curve characterizing the difference between the lens and the background by performing sliding mean smoothing, local maximum location and interval truncation on the initial histogram, avoiding misjudgment caused by noise interference, and providing a stable and reproducible statistical basis for subsequent valley pixel truncation.

[0051] In a specific embodiment, the initial screening lens mask is subjected to watershed boundary blocking based on the center distance matrix to obtain an independent lens region, including: The center distance matrix is ​​subjected to a maximum search to obtain local extreme value pixels, and the local extreme value pixels are independently assigned numerical values ​​to obtain a seed label matrix; The initial screening lens mask is expanded layer by layer based on the seed tag matrix to obtain the pixel expansion region, and adjacent collision detection is performed based on the pixel expansion region to obtain the intersection boundary line. Based on the intersection boundary line, the background pixels of the initial screening lens mask are replaced to obtain a separation mask image, and the connected components of the separation mask image are extracted to obtain independent lens regions.

[0052] Specifically, the watershed boundary blocking of the initial screening lens mask based on the center distance matrix is ​​used to obtain independent lens regions. The specific implementation process is as follows: First, a local maximum search is performed on the center distance matrix: traverse each foreground pixel in the matrix (i.e., the position with a value of 1 in the initial screening lens mask), and determine whether its distance value is strictly greater than the distance values ​​of all its eight neighboring pixels; if this condition is met, the pixel is marked as a local extremum pixel. Under typical imaging conditions of VCSEL chip lens arrays, each adhered lens cluster usually corresponds to one or more such extremum points, whose positions are roughly located near the geometric center of each lens.

[0053] Subsequently, all detected local extremum pixels are assigned distinct integer labels (e.g., starting from 1 and incrementing sequentially), generating a seed label matrix. This matrix retains a unique identifier only at the extremum pixel locations, setting all other locations to zero, thus providing a clear starting point for subsequent region growth. Next, using this seed label matrix as the initial state, layer-by-layer pixel expansion is performed within the foreground region defined by the initial screening lens mask: in each iteration, all currently non-zero labeled boundary pixels are expanded one step towards their four- or eight-neighborhood unlabeled foreground pixels, inheriting their original labels; this process continues until all foreground pixels are labeled or adjacent regions come into contact.

[0054] During the expansion process, the system simultaneously performs adjacent collision detection: once the expansion fronts of two different labels meet at the same pixel location, the pixel is determined to be part of the intersection boundary line. The intersection boundary line is essentially the boundary formed by the growth competition of different lens regions, and its direction is determined by the gradient distribution of the center distance matrix, naturally tending to extend along the perpendicular bisector of the line connecting the centers of the two lenses. After all expansion is completed, all pixels on the intersection boundary line are forcibly set to background values ​​(i.e., 0) in the initial screening lens mask, realizing the physical cutting of the adhesion region, thereby generating a separation mask map.

[0055] Finally, connected component extraction is performed on the separated mask image: using a standard two-pass scanning method or disjoint-set data structure algorithm, all sets of foreground pixels that are connected to each other are identified, and each connected region is assigned a unique number. Each number corresponds to an independent lens region with clear boundaries and no overlapping or adhering elements, which can be directly used for subsequent single-unit analysis. For example, in a high-density VCSEL array image, the initial screening mask misclassified three adjacent lenses as a single connected block. After the above process, the center distance matrix showed three obvious peaks, and two intersecting boundary lines were formed during the expansion process, ultimately successfully segmenting three independent lens regions.

[0056] In this embodiment, the above scheme effectively solves the adhesion problem caused by small lens spacing or uneven surface reflection by constructing seed markers under the guidance of the center distance matrix, implementing controlled pixel expansion, and using the intersection boundary line for background replacement. It achieves accurate individual separation of each unit in the VCSEL chip lens array, laying a reliable foundation for subsequent defect location and parameter measurement.

[0057] In a specific embodiment, multi-scale pixel matrix convolution is performed on the independent lens region to obtain local edge contours, including: Spatial sliding convolution is performed on the independent lens regions to obtain a multi-scale smooth image; Based on the multi-scale smoothed image, cross-scale gray-level subtraction is performed on the independent lens region to obtain a multi-scale difference image; Multi-directional edge matrix convolution is performed on the multi-scale difference image to obtain a multi-directional edge feature map, and local extremum filtering is performed on the multi-scale difference image based on the multi-directional edge feature map to obtain discrete defect pixels. The discrete defect pixels are merged to obtain a continuous defect region, and the outer boundary of the continuous defect region is tracked to obtain the local edge contour.

[0058] Specifically, multi-scale pixel matrix convolution is performed on the independent lens regions to extract local edge contours. The specific implementation process is as follows: First, a spatial sliding convolution operation is performed within each independent lens region: a set of Gaussian kernel functions are selected, with standard deviations σ of 1.0, 2.0, and 3.0, corresponding to different levels of smoothness. Each Gaussian kernel is slid pixel by pixel across the lens region image, and a local weighted average is calculated to generate three multi-scale smoothed images. These images retain the structural information of the original lens region at different spatial frequencies. The larger the scale, the more blurred the details, but the more stable the overall shape.

[0059] Next, cross-scale gray-level subtraction is performed based on the aforementioned multi-scale smoothed images: the smoothed image at a smaller scale (e.g., σ=1.0) is subtracted from the smoothed image at a larger scale (e.g., σ=2.0 or σ=3.0), resulting in the corresponding multi-scale difference image. This difference operation is essentially a bandpass filter, capable of enhancing gray-level changes within a specific scale range, such as abrupt reflections caused by lens edges or minor surface depressions. In VCSEL lens imaging, if a scratch or chipped edge exists, it is still clearly visible in the σ=1.0 image, but has been smoothed out in the σ=3.0 image. After subtracting the two, the defect location will exhibit a clear positive or negative response.

[0060] Subsequently, each multi-scale difference image is convolved with edge detection kernels in multiple directions. Commonly used directions include 0°, 45°, 90°, and 135°, and the corresponding convolution kernels can be Sobel or Prewitt operators. Through this multi-directional edge matrix convolution, several multi-directional edge feature maps are obtained, each highlighting the gradient response in a certain direction. These feature maps are then fused by taking the maximum absolute value at each pixel position to form a comprehensive edge response map. Local extremum filtering is then performed on this map on the multi-scale difference image: for each pixel, if its grayscale value is a maximum or minimum value within its 3×3 neighborhood, and the absolute value exceeds a preset threshold (e.g., 10% of the dynamic range of the difference image), it is marked as a discrete defect pixel.

[0061] These discrete defect pixels are typically distributed as points or short lines and require further integration. Pixel connectivity merging is performed using the four-neighbor connectivity criterion, aggregating spatially adjacent defect points into continuous defect regions. Finally, chain code tracing or Freeman coding is used to trace the outer boundary of each continuous defect region, sequentially recording the coordinate sequence of boundary pixels to obtain closed or semi-closed local edge contours.

[0062] For example, when there is a micron-level defect at the edge of a VCSEL lens, the region appears as alternating bright and dark stripes in the difference image with σ=1.0 and σ=3.0. After multi-directional convolution, the response is strongest in the 45° direction. Local extremum screening accurately captures its position, and after connection and merging, a complete defect region is formed. Boundary tracking finally outputs its outer contour line.

[0063] In this embodiment, the above-mentioned scheme effectively extracts the precise edge contour of weak defects on the surface of VCSEL lens through multi-scale smoothing, cross-scale difference, multi-directional edge response fusion and local extremum screening. It overcomes the problem of easy missed detection or false detection in single-scale or single-directional detection, and significantly improves the completeness and positioning accuracy of defect geometric characterization.

[0064] In a specific embodiment, the multi-scale difference image is convolved with a multi-directional edge matrix to obtain a multi-directional edge feature map, including: The first-order derivative convolution kernel sliding calculations were performed on the multi-scale difference image in four directions: 0°, 45°, 90°, and 135°, to obtain four sets of directional response maps. Pixel-level absolute value calculations are performed on the four sets of directional response maps to obtain four sets of directional gradient magnitude maps; Based on the four sets of directional gradient magnitude maps, pixel-wise maximum value fusion is performed to obtain an anisotropic response map, and neighborhood mean smoothing is performed based on the anisotropic response map to obtain a multi-directional edge feature map.

[0065] Specifically, the multi-scale difference image is convolved with a multi-directional edge matrix to generate a multi-directional edge feature map. The specific implementation process is as follows: First, the multi-scale difference image is slidably calculated with four fixed-direction first-derivative convolution kernels, which are 0°, 45°, 90°, and 135° respectively. The 0° direction kernel typically adopts a form like […]. The horizontal Sobel operator [1, 0, +1] uses its transpose in the 90° direction for detecting vertical edges, while diagonally arranged kernel templates are used in the 45° and 135° directions, for example [[ 2, 1, 0], [ [1, 0, +1], [0, +1, +2]] and their mirror variants. During convolution, the kernel slides pixel by pixel on the image, and the output at each position is a weighted sum of the kernel coefficients and the corresponding local pixel value, ultimately forming four sets of directional response maps. Each set of maps reflects the rate of grayscale change in that direction, and can be positive or negative.

[0066] Subsequently, pixel-level absolute value operations were performed on these four sets of directional response maps: each pixel position was traversed, and the absolute value of its response value was taken to eliminate the influence of sign and retain only the gradient intensity information, resulting in four sets of directional gradient magnitude maps. Since surface defects of VCSEL lenses (such as microcracks or edge chipping) may exhibit different orientations in different directions, the magnitude map in a single direction often can only capture part of the edge, so further fusion is required.

[0067] The fusion operation employs a pixel-wise maximum value strategy: at the same coordinate position, the values ​​of the four sets of directional gradient magnitude maps are compared, and the largest value is selected as the output value for that point, thereby constructing an anisotropic response map. This map integrates the most significant edge responses across all directions, effectively enhancing structural abrupt changes in any orientation while suppressing weak response noise in non-dominant directions. However, due to potential high-frequency perturbations in the difference image itself, the anisotropic response map often exhibits isolated bright spots or spikes near the edges.

[0068] Therefore, neighborhood mean smoothing is applied to the anisotropic response map: a 3×3 square window is selected, slid across the image, and the arithmetic mean of all pixels within the window is calculated. This mean value is then used to replace the center pixel value. Although this operation slightly blurs the edges, it effectively suppresses salt-and-pepper artifacts and makes the edge energy distribution more continuous. After this processing, the final result is a multi-directional edge feature map, which retains the strong response characteristics of multi-directional edges and has good spatial smoothness, facilitating subsequent extreme value selection.

[0069] For example, in a VCSEL lens image, a diagonal micro-scratch shows a high amplitude in the 45° response map, but almost no response in the 0° and 90° maps; after maximum value fusion, the scratch is completely preserved in the anisotropic response map, and after neighborhood mean smoothing, its edge contour becomes continuous and free of stray points.

[0070] In this embodiment, the above scheme constructs a multi-directional edge feature map that is sensitive to edges of any orientation and is robust to noise by means of directional first derivative convolution, absolute value transformation, maximum value fusion and neighborhood smoothing. This provides high-quality input for the accurate extraction of discrete defect pixels and significantly improves the orientation adaptability and stability of VCSEL lens surface micro-defect detection.

[0071] In a specific embodiment, the outer boundary of the continuous defect region is traced to obtain a local edge contour, including: The continuous defect region is scanned pixel by pixel to obtain the outermost starting pixel, and the continuous defect region is searched clockwise based on the outermost starting pixel to obtain the initial closed boundary line. The initial closed boundary line is filtered by vertex distance threshold to obtain a set of inflection point coordinates. Based on the set of inflection point coordinates, spline curve interpolation is performed on the initial closed boundary line to obtain the local edge contour.

[0072] Specifically, the multi-scale difference image is convolved with a multi-directional edge matrix to generate a multi-directional edge feature map. The specific implementation process is as follows: First, the multi-scale difference image is slidably calculated with four fixed-direction first-derivative convolution kernels, which are 0°, 45°, 90°, and 135° respectively. The 0° direction kernel typically adopts a form like […]. The horizontal Sobel operator [1, 0, +1] uses its transpose in the 90° direction for detecting vertical edges, while diagonally arranged kernel templates are used in the 45° and 135° directions, for example [[ 2, 1, 0], [ [1, 0, +1], [0, +1, +2]] and their mirror variants. During convolution, the kernel slides pixel by pixel on the image, and the output at each position is a weighted sum of the kernel coefficients and the corresponding local pixel value, ultimately forming four sets of directional response maps. Each set of maps reflects the rate of grayscale change in that direction, and can be positive or negative.

[0073] Subsequently, pixel-level absolute value operations were performed on these four sets of directional response maps: each pixel position was traversed, and the absolute value of its response value was taken to eliminate the influence of sign and retain only the gradient intensity information, resulting in four sets of directional gradient magnitude maps. Since surface defects of VCSEL lenses (such as microcracks or edge chipping) may exhibit different orientations in different directions, the magnitude map in a single direction often can only capture part of the edge, so further fusion is required.

[0074] The fusion operation employs a pixel-wise maximum value strategy: at the same coordinate position, the values ​​of the four sets of directional gradient magnitude maps are compared, and the largest value is selected as the output value for that point, thereby constructing an anisotropic response map. This map integrates the most significant edge responses across all directions, effectively enhancing structural abrupt changes in any orientation while suppressing weak response noise in non-dominant directions. However, due to potential high-frequency perturbations in the difference image itself, the anisotropic response map often exhibits isolated bright spots or spikes near the edges.

[0075] Therefore, neighborhood mean smoothing is applied to the anisotropic response map: a 3×3 square window is selected, slid across the image, and the arithmetic mean of all pixels within the window is calculated. This mean value is then used to replace the center pixel value. Although this operation slightly blurs the edges, it effectively suppresses salt-and-pepper artifacts and makes the edge energy distribution more continuous. After this processing, the final result is a multi-directional edge feature map, which retains the strong response characteristics of multi-directional edges and has good spatial smoothness, facilitating subsequent extreme value selection.

[0076] For example, in a VCSEL lens image, a diagonal micro-scratch shows a high amplitude in the 45° response map, but almost no response in the 0° and 90° maps; after maximum value fusion, the scratch is completely preserved in the anisotropic response map, and after neighborhood mean smoothing, its edge contour becomes continuous and free of stray points.

[0077] In this embodiment, the above scheme constructs a multi-directional edge feature map that is sensitive to edges of any orientation and is robust to noise by means of directional first derivative convolution, absolute value transformation, maximum value fusion and neighborhood smoothing. This provides high-quality input for the accurate extraction of discrete defect pixels and significantly improves the orientation adaptability and stability of VCSEL lens surface micro-defect detection.

[0078] In a specific embodiment, wafer physical coordinate mapping is performed based on the defect geometric features to obtain a defect distribution map, including: The local centroid coordinates of the defect geometric features are extracted to obtain a single lens centroid coordinate set, and the array period is fitted to the single lens centroid coordinate set to obtain an array reference coordinate grid. Based on the array reference coordinate grid, the single lens centroid coordinate set is indexed by row and column to obtain an indexed defect coordinate table. Align the indexed defect coordinate table with a wafer-level coordinate system to obtain a defect distribution map.

[0079] Specifically, a wafer physical coordinate mapping is performed based on the defect geometric features to generate a defect distribution map. The specific implementation process is as follows: First, local centroid coordinate extraction is performed on the defect geometric features corresponding to each independent lens region: for each identified continuous defect region, the arithmetic mean of all its pixel coordinates is calculated, i.e. ,in This represents the total number of pixels within the defective area. This represents the position of each pixel in the image coordinate system. This centroid point approximately represents the center position of the defect within a single lens, thus converging the centroids of all lenses to form a single-lens centroid coordinate set.

[0080] Subsequently, the coordinate set is used for array period fitting. Since VCSEL lenses are typically arranged in regular rectangles or hexagons on the wafer, the center-to-center distance between adjacent lenses is essentially constant. Therefore, autocorrelation analysis or Fourier spectrum peak detection can be performed on the centroid coordinate set in the X and Y directions respectively to determine the principal period length. and Then, the least squares method is used to fit all centroids to an ideal mesh model, which is based on the base point. With periodic vector , The final output array reference coordinate grid is constructed. This grid does not directly use the original image coordinates, but is a reconstructed idealized lattice used to correct positional shifts caused by imaging distortion or process deviations.

[0081] Based on this reference grid, row and column indexing is performed on the single-lens centroid coordinate set: each centroid point is projected to the nearest grid intersection point, using its row number. With column number As integer indices, an indexed defect coordinate table is formed. This table not only contains the original image coordinates but also associates them with the logical position of the lens on the wafer (e.g., row 32, column 45), allowing defects to be precisely located by device unit. For example, on a 6-inch GaAs wafer, if the measured... , Then, a defect centroid located at image coordinates (10256, 8942) may be assigned to (i=71, j=69) after matching, corresponding to a specific chip location on the wafer.

[0082] Finally, the indexed defect coordinate table is aligned with the wafer-level coordinate system. The wafer coordinate system is typically based on the wafer notch or flat edge, with the origin at the wafer's geometric center, and units in micrometers. Using the known calibration relationship between the imaging field of view and the wafer stage coordinates (such as an affine transformation matrix), the image coordinates are converted to physical coordinates, and combined with the index information to complete a global mapping, ultimately generating a defect distribution map. This map uses the wafer plane as a base map, marking the physical location of each defect and its corresponding lens number.

[0083] In this embodiment, the above scheme accurately maps local image defects to the wafer physical space through centroid extraction, period fitting, index allocation and coordinate system alignment, realizing a closed loop from pixel-level detection to wafer-level positioning, and providing a traceable and quantifiable spatial distribution basis for yield analysis and process feedback in VCSEL manufacturing.

[0084] In a specific embodiment, defect type diagnosis is performed on the VCSEL chip lens based on the defect distribution map, and a defect type report is generated, including: Cluster the defect coordinates in the defect distribution map, and group the coordinates of each defect coordinate that are less than the preset neighborhood radius into the same defect cluster area to obtain at least one defect cluster area. For each of the defect clusters, a circumscribed contour is fitted to obtain the geometry of the defect cluster and the dimensional parameters of its circumscribed contour. The geometric shape is matched with multiple standard process geometric templates in a preset process defect map library, and the contour overlap between each standard process geometric template and the geometric shape is calculated. Select the process defect category corresponding to the standard process geometry template with the highest contour overlap as the process defect category of the current defect cluster area; Based on the process defect category and the dimensional parameters of the outer contour of the defect cluster area and the defect coordinate points, diagnostic information is arranged to generate a defect type report.

[0085] Specifically, the defect type diagnosis and defect type report generation for the VCSEL chip lens based on the defect distribution map are implemented as follows: First, spatial clustering is performed on all defect coordinate points in the defect distribution map: a preset neighborhood radius (e.g., 50 micrometers) is set, and each defect point is traversed. If the Euclidean distance between two points is less than this radius, they are considered to belong to the same local region. Using DBSCAN or connected component labeling algorithms, points that meet the proximity condition are grouped into the same defect cluster area, ultimately resulting in several defect cluster areas. In actual VCSEL wafer inspection, if there is a lithographic alignment misalignment in a certain area, edge chipping may occur consecutively at multiple adjacent lens positions. These defect points are relatively close to each other, naturally forming a cluster area.

[0086] Next, the circumscribed contour of each defect cluster is extracted: first, the smallest closed boundary enclosing all points is fitted using the convex hull algorithm or the α-shape method; then, the geometric features of the circumscribed contour are calculated, including dimensional parameters such as the major axis, minor axis, area, eccentricity, and orientation angle. For example, the circumscribed contour of a linear scratch defect cluster caused by mask contamination is usually an elongated ellipse, with a major axis of up to 200 micrometers, a minor axis of only 15 micrometers, and an eccentricity close to 0.95.

[0087] Subsequently, the geometry was compared one by one with multiple standard process geometry templates in a pre-defined process defect map library. The map library contains templates corresponding to typical process anomalies, such as "sputtering particle contamination" appearing as a circle, "dry etching over-etching" appearing as a rectangular array, and "photoresist residue" appearing as irregular clumps. Shape matching uses the contour overlap index, which is calculated as the intersection-union ratio (IoU) or the reciprocal of the Hausdorff distance between the current circumscribed contour and the template contour at a normalized scale. A higher value indicates a better match. Assuming that the IoU between the contour of a certain cluster and the "mask fragment detachment" template reaches 0.82, which is significantly higher than other templates (such as "ion implantation offset" which is only 0.31), the process defect category corresponding to this template is selected as the diagnostic result for the current cluster.

[0088] Finally, based on the determined process defect category, the dimensional parameters of the outer contour (such as length, width, and area), and the wafer location information of the original defect coordinate point, the diagnostic information is formatted according to a preset format: each record contains fields such as defect ID, chip row and column index, physical coordinates, cluster size, matching template name, and confidence level, and the output is in the form of structured text or table, i.e., defect type report.

[0089] In this embodiment, the above-mentioned scheme achieves automatic classification from wafer-level defect distribution to specific process root causes through spatial clustering, circumscribed contour fitting, template matching and structured layout. This enables the differentiation of defects from different sources (such as lithography, etching or deposition anomalies) in the VCSEL manufacturing process, significantly improving the efficiency and accuracy of defect analysis and providing a direct basis for process window optimization and equipment maintenance.

[0090] The above describes the intelligent defect identification and diagnosis method for VCSEL chip lenses in embodiments of the present invention. The following describes the intelligent defect identification and diagnosis system for VCSEL chip lenses in embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the intelligent defect identification and diagnosis system for VCSEL chip lenses in this invention includes: Imaging module 1 is used to perform optical imaging on the lens of the VCSEL chip by illuminating it with a polarized light source to obtain a polarized reflection image; Segmentation module 2 is used to perform grayscale threshold segmentation based on the polarization reflection image to obtain independent lens regions; The judgment module 3 is used to perform multi-scale pixel matrix convolution on the independent lens region to obtain local edge contours, and determine whether the VCSEL chip lens has defective geometric features based on the local edge contours. Mapping module 4 is used to perform wafer physical coordinate mapping based on the defect geometric features if the defect exists, to obtain a defect distribution map; Diagnostic module 5 is used to diagnose the defect type of the VCSEL chip lens based on the defect distribution map and generate a defect type report.

[0091] In this embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, and will not be repeated here.

Claims

1. A method for intelligent identification and diagnosis of defects in VCSEL chip lenses, characterized in that, include: Optical imaging is performed on the lens of the VCSEL chip by illuminating it with a polarized light source to obtain a polarized reflection image; Based on the polarization reflection image, grayscale threshold segmentation is performed to obtain independent lens regions; Multi-scale pixel matrix convolution is performed on the independent lens region to obtain local edge contours, and the presence of defective geometric features in the VCSEL chip lens is determined based on the local edge contours. If present, a wafer physical coordinate mapping is performed based on the defect geometric features to obtain a defect distribution map; Based on the defect distribution map, the VCSEL chip lens is used to diagnose the defect type and generate a defect type report.

2. The method for intelligent identification and diagnosis of defects in VCSEL chip lenses according to claim 1, characterized in that, Optical imaging of the VCSEL chip lens is performed using a polarized light source to obtain a polarized reflection image, including: The surface of the VCSEL chip lens is obliquely illuminated by a linearly polarized light source at an incident angle of 45° to obtain the difference in reflected light intensity between the s-polarization component and the p-polarization component. Based on the difference in reflected light intensity, polarization differential imaging is performed on the VCSEL chip lens to obtain an initial polarization reflection image. The initial polarized reflection image is subjected to background light suppression processing to obtain a background light distribution map, and the background light distribution map is subtracted pixel by pixel from the initial polarized reflection image to obtain the polarized reflection image.

3. The intelligent defect identification and diagnosis method for VCSEL chip lenses according to claim 1, characterized in that, Based on the polarization reflection image, grayscale thresholding is performed to obtain independent lens regions, including: Histogram analysis of the polarization reflection image yields a bimodal distribution curve. Based on the bimodal distribution curve, the polarization reflection image is truncated at the valley bottom pixels to obtain the initial screening lens mask; The pixel distance transformation is performed on the initial screening lens mask to obtain a center distance matrix, and the watershed boundary is blocked based on the center distance matrix to obtain an independent lens region.

4. The intelligent defect identification and diagnosis method for VCSEL chip lenses according to claim 3, characterized in that, Based on the center distance matrix, the watershed boundary of the initial screening lens mask is blocked to obtain an independent lens region, including: The center distance matrix is ​​subjected to a maximum search to obtain local extreme value pixels, and the local extreme value pixels are independently assigned numerical values ​​to obtain a seed label matrix; The initial screening lens mask is expanded layer by layer based on the seed tag matrix to obtain the pixel expansion region, and adjacent collision detection is performed based on the pixel expansion region to obtain the intersection boundary line. Based on the intersection boundary line, the background pixels of the initial screening lens mask are replaced to obtain a separation mask image, and the connected components of the separation mask image are extracted to obtain independent lens regions.

5. The intelligent defect identification and diagnosis method for VCSEL chip lenses according to claim 1, characterized in that, Multi-scale pixel matrix convolution is performed on the independent lens region to obtain the local edge contour, including: Spatial sliding convolution is performed on the independent lens regions to obtain a multi-scale smooth image; Based on the multi-scale smoothed image, cross-scale gray-level subtraction is performed on the independent lens region to obtain a multi-scale difference image; Multi-directional edge matrix convolution is performed on the multi-scale difference image to obtain a multi-directional edge feature map, and local extremum filtering is performed on the multi-scale difference image based on the multi-directional edge feature map to obtain discrete defect pixels. The discrete defect pixels are merged to obtain a continuous defect region, and the outer boundary of the continuous defect region is tracked to obtain the local edge contour.

6. The intelligent defect identification and diagnosis method for VCSEL chip lenses according to claim 5, characterized in that, The multi-scale difference image is convolved with a multi-directional edge matrix to obtain a multi-directional edge feature map, including: The first-order derivative convolution kernel sliding calculations were performed on the multi-scale difference image in four directions: 0°, 45°, 90°, and 135°, to obtain four sets of directional response maps. Pixel-level absolute value calculations are performed on the four sets of directional response maps to obtain four sets of directional gradient magnitude maps; Based on the four sets of directional gradient magnitude maps, pixel-wise maximum value fusion is performed to obtain an anisotropic response map, and neighborhood mean smoothing is performed based on the anisotropic response map to obtain a multi-directional edge feature map.

7. The method for intelligent identification and diagnosis of defects in VCSEL chip lenses according to claim 5, characterized in that, The outer boundary of the continuous defect region is traced to obtain the local edge contour, including: The continuous defect region is scanned pixel by pixel to obtain the outermost starting pixel, and the continuous defect region is searched clockwise based on the outermost starting pixel to obtain the initial closed boundary line. The initial closed boundary line is filtered by vertex distance threshold to obtain a set of inflection point coordinates. Based on the set of inflection point coordinates, spline curve interpolation is performed on the initial closed boundary line to obtain the local edge contour.

8. The intelligent defect identification and diagnosis method for VCSEL chip lenses according to claim 1, characterized in that, Based on the aforementioned defect geometric features, wafer physical coordinate mapping is performed to obtain a defect distribution map, including: The local centroid coordinates of the defect geometric features are extracted to obtain a single lens centroid coordinate set, and the array period is fitted to the single lens centroid coordinate set to obtain an array reference coordinate grid. Based on the array reference coordinate grid, the single lens centroid coordinate set is indexed by row and column to obtain an indexed defect coordinate table. Align the indexed defect coordinate table with a wafer-level coordinate system to obtain a defect distribution map.

9. The intelligent defect identification and diagnosis method for VCSEL chip lenses according to claim 1, characterized in that, Based on the defect distribution map, the VCSEL chip lens is subjected to defect type diagnosis, and a defect type report is generated, including: Cluster the defect coordinates in the defect distribution map, and group the coordinates of each defect coordinate that are less than the preset neighborhood radius into the same defect cluster area to obtain at least one defect cluster area. For each of the defect clusters, a circumscribed contour is fitted to obtain the geometry of the defect cluster and the dimensional parameters of its circumscribed contour. The geometric shape is matched with multiple standard process geometric templates in a preset process defect map library, and the contour overlap between each standard process geometric template and the geometric shape is calculated. Select the process defect category corresponding to the standard process geometry template with the highest contour overlap as the process defect category of the current defect cluster area; Based on the process defect category and the dimensional parameters of the outer contour of the defect cluster area and the defect coordinate points, diagnostic information is arranged to generate a defect type report.

10. A defect intelligent identification and diagnosis system for VCSEL chip lenses, characterized in that, include: The imaging module is used to perform optical imaging on the lens of the VCSEL chip by illuminating it with a polarized light source to obtain a polarized reflection image; The segmentation module is used to perform grayscale threshold segmentation based on the polarization reflection image to obtain independent lens regions; The judgment module is used to perform multi-scale pixel matrix convolution on the independent lens region to obtain local edge contours, and determine whether the VCSEL chip lens has defective geometric features based on the local edge contours. The mapping module is used to perform wafer physical coordinate mapping based on the defect geometric features if the defect exists, to obtain a defect distribution map; The diagnostic module is used to diagnose the defect type of the VCSEL chip lens based on the defect distribution map and generate a defect type report.

Images