A defect detection method and device based on objective lens surface composite phase microstructure

Abstract

The application relates to a defect detection method and device based on an objective lens surface composite phase microstructure. The defect detection method comprises the following steps: acquiring a first polarization channel image and a second polarization channel image; performing angular average processing to obtain a one-dimensional radial intensity distribution; acquiring a radial interference carrier frequency F according to the one-dimensional radial intensity distribution; acquiring a first depth coordinate z1 of each defect according to the radial interference carrier frequency F; performing speckle texture feature extraction to acquire a depth index of each defect; performing differential operation to obtain a spin differential image and acquire a spin differential depth estimation value z2; and acquiring a three-dimensional space coordinate of the defect. The defect detection device comprises an illumination light source, a coaxial beam splitter, an objective lens group, an aperture diaphragm, a tube lens group, a polarization beam splitter and a double-channel image sensor. The application has high detection precision, considers resolution and large depth of field.

Description

Technical Field

[0001] This invention relates to the field of semiconductor inspection technology, and in particular to a defect detection method and device based on the composite phase microstructure of the objective lens surface. Background Technology

[0002] In semiconductor wafer defect detection, dual telecentric optical systems have outstanding advantages such as constant magnification, no parallax, and low distortion because both the entrance pupil and exit pupil are located at infinity. They are widely used for two-dimensional high-precision dimensional measurement of semiconductor wafers and precision optical components.

[0003] In the existing technology, the traditional dual telecentric system has the following technical defects: First, the depth of field of the high numerical aperture dual telecentric system is extremely shallow (usually only a few micrometers to tens of micrometers), which cannot directly cover the axial depth range where defects on the wafer surface and subsurface may be distributed (such as the upper and lower surfaces of the thin film layer, subsurface cracks), resulting in a severe limitation on its three-dimensional detection capability; Second, existing three-dimensional detection methods mainly rely on confocal axial scanning, structured light projection or multi-view stereo vision. Confocal scanning requires mechanical drive, is slow and sensitive to vibration, structured light is prone to failure on mirror or low contrast surfaces, and stereo vision requires dual optical paths and is highly dependent on texture. None of the above methods can achieve high resolution, large depth of field and high speed at the same time.

[0004] Therefore, it is necessary to provide a new defect detection method and device based on the composite phase microstructure of the objective lens surface to solve the above-mentioned problems in the prior art. Summary of the Invention

[0005] The technical problem to be solved by this application is how to provide a defect detection method and device based on the composite phase microstructure of the objective lens surface with high detection accuracy, while also taking into account resolution and large depth of field.

[0006] To address the aforementioned technical problems, according to embodiments of this application, a defect detection method based on a composite phase microstructure on the objective lens surface is provided, comprising the following steps: illuminating the sample to be tested with illumination light to generate reflected light; the reflected light is modulated by an objective lens assembly integrating a composite phase microstructure and then separated and collected by a polarization beam splitter to obtain a first polarization channel image and a second polarization channel image; performing angular averaging on the first polarization channel image and the second polarization channel image respectively to obtain a one-dimensional radial intensity distribution; obtaining a radial interference carrier frequency F based on the one-dimensional radial intensity distribution; obtaining the first depth coordinate z1 of each defect based on the radial interference carrier frequency F; extracting speckle texture features from the first polarization channel image and the second polarization channel image respectively to obtain the depth index of each defect; performing a difference operation on the first polarization channel image and the second polarization channel image to obtain a spin difference image; obtaining a spin difference depth estimate z2 based on the spin difference image; and reconstructing the first polarization channel image and the second polarization channel image by limiting the axial search range with the first depth coordinate z1, outputting images of each depth layer, and obtaining the three-dimensional spatial coordinates (x, y, z) of the defect.

[0007] According to an embodiment of this application, the step of performing angular averaging on the first polarization channel image and the second polarization channel image to obtain a one-dimensional radial intensity distribution includes: pre-locating defects in the first polarization channel image and the second polarization channel image to obtain diffuse spots corresponding to each defect; establishing a polar coordinate system with the geometric center of each diffuse spot as the origin; and performing integral averaging on the two-dimensional local image patch along the circumferential direction to obtain the one-dimensional radial intensity distribution.

[0008] According to an embodiment of this application, obtaining the radial interference carrier frequency F based on the one-dimensional radial intensity distribution includes extracting the peak frequency after performing a one-dimensional Fourier transform on the one-dimensional radial intensity distribution, and using it as the radial interference carrier frequency F.

[0009] According to an embodiment of this application, obtaining the first depth coordinate z1 of each defect based on the radial interference carrier frequency F includes obtaining the first depth coordinate z1 of each defect based on the mapping relationship between the radial interference carrier frequency F and the axial depth of the defect; Where j is the pixel size of the dual-channel image sensor, The axial distance between the two focal points. The wavelength of the illumination light, This is the focal length of the objective lens group.

[0010] According to an embodiment of this application, the step of extracting speckle texture features from the first polarization channel image and the second polarization channel image to obtain the depth index of each defect includes: moving the standard sample sequentially along the axial direction to multiple preset depth positions; acquiring speckle images of the corresponding depth layer at each preset depth position and extracting speckle texture features; establishing a correspondence between each preset depth position and its speckle texture features; extracting speckle texture features from the local area where each defect is located to obtain measured speckle features; matching the measured speckle features of each defect with the speckle texture features at each preset depth position in the correspondence, and taking the preset depth position with the highest matching degree as the depth index.

[0011] According to an embodiment of this application, the step of performing a difference operation between the first polarization channel image and the second polarization channel image to obtain a spin differential image includes subtracting the gray values ​​of corresponding pixel positions in the first polarization channel image and the second polarization channel image to obtain a spin differential image; wherein, after the illumination light is reflected by the sample under test and modulated by the composite phase microstructure, the reflected light simultaneously contains a left-handed circularly polarized component and a right-handed circularly polarized component; the polarization shearing layer in the composite phase microstructure applies a linear phase tilt in a first direction to the left-handed circularly polarized component and applies a linear phase tilt in a second direction opposite to the first direction to the right-handed circularly polarized component, causing the diffuse spots of the same defect in the first polarization channel image and the second polarization channel image to produce radial shifts in opposite directions; under the condition of defect defocusing, the overlapping area of ​​the two diffuse spots changes with the axial depth of the defect; the spin differential image retains the spin differential contrast signal formed by the change of the overlapping area.

[0012] According to an embodiment of this application, obtaining the spin differential depth estimate z2 based on the spin differential image includes performing grayscale statistics on the local regions where each defect is located in the spin differential image, extracting the standard deviation of the grayscale values ​​of the pixels in the local regions, and determining the spin differential depth estimate z2 based on the standard deviation.

[0013] According to an embodiment of this application, the step of reconstructing the first polarization channel image and the second polarization channel image, outputting images of each depth layer, and obtaining the three-dimensional spatial coordinates (x, y, z) of the defect includes: determining an axial search range with the first depth coordinate z1 as the center and a preset error tolerance as the boundary; within the axial search range, modeling the first polarization channel image and the second polarization channel image as the sum of the convolution superposition of the defect distribution at each depth layer and the diffusion function of the corresponding depth point, plus a noise term. ,in, Corresponding to the first polarization channel image and the second polarization channel image; For depth Defect distribution at the location, Let z be the point spread function of the polarization channel at depth z. For noise terms, Represents a two-dimensional convolution; an initial depth value is constructed using the depth index and the spin difference depth estimate z2. By utilizing the sparsity constraint of the axial distribution of defects, a multi-channel collaborative optimization objective function is established: ,in, For dual-channel data fidelity, For axially sparse regularization terms, For depth prior constraints, and The regularization parameter is used; the multi-channel collaborative optimization objective function is solved using the alternating direction multiplier method, and the tomographic images of each depth layer are output. ; for the tomographic image Defect identification is performed to determine the depth layer position z of the defect and its lateral coordinates (x, y) within that depth layer. The three-dimensional spatial coordinates (x, y, z) of the defect are obtained by combining the lateral coordinates with the depth layer position.

[0014] A defect detection device based on the composite phase microstructure of the objective lens surface is used to implement the above-mentioned defect detection method. The defect detection device includes an illumination source, a coaxial beam splitter, an objective lens group, an aperture stop, a tube lens group, a polarizing beam splitter, and a dual-channel image sensor. The coaxial beam splitter is disposed between the illumination source and the objective lens group, and is used to reflect the illumination light emitted by the illumination source into the objective lens group, and transmit the reflected light after sample reflection into the tube lens group. The front focal point of the objective lens group coincides with the surface of the sample to be tested, and the rear focal point of the objective lens group coincides with the front focal point of the tube lens group. The focal point coincides with the photosensitive surface of the dual-channel image sensor; the aperture stop is located at the rear focal plane of the objective lens group; the polarizing beam splitter is located between the tube lens group and the dual-channel image sensor, used to spatially separate the left-hand circularly polarized component and the right-hand circularly polarized component modulated by the objective lens group and guide them to the two photosensitive channels of the dual-channel image sensor respectively; the two photosensitive channels of the dual-channel image sensor respectively receive the separated left-hand circularly polarized component and the right-hand circularly polarized component, and synchronously complete the photoelectric conversion; a composite phase microstructure is integrated on the lens surface of the objective lens group near the aperture stop.

[0015] According to an embodiment of this application, the composite phase microstructure includes a radial dual-focus phase modulation layer, a random phase scattering layer, and a polarization shearing layer arranged sequentially from bottom to top. The radial dual-focus phase modulation layer is used to radially divide the pupil into at least two concentric annular regions, each annular region having a different phase profile to introduce different defocus amounts, forming a dual focus with a fixed axial spacing. The random phase scattering layer is located on the radial dual-focus phase modulation layer, and its surface has a random concave-convex structure. The undulation height of the concave-convex structure is 0.1 to 2.0 times the wavelength of the illumination light, and the lateral feature size of the concave-convex structure is 0.2 to 0.8 times the wavelength of the illumination light. The polarization shearing layer is located on the random phase scattering layer and is used to apply radial linear phase tilts to the left-hand circularly polarized component corresponding to the first polarization channel and the right-hand circularly polarized component corresponding to the second polarization channel, respectively. The direction of the linear phase tilt applied to the left-hand circularly polarized component is opposite to the direction of the linear phase tilt applied to the right-hand circularly polarized component.

[0016] By adopting the above technical solution, a multifunctional composite phase microstructure is integrated on the objective lens surface. Combining circular polarization dual-channel acquisition and multi-channel collaborative sparse reconstruction, high resolution, large depth of field and high speed are achieved without reducing numerical aperture. At the same time, the three-dimensional position of defects can be accurately determined. Attached Figure Description

[0017] Fig. 1 This is a schematic diagram of the optical path of a defect detection device according to an embodiment of the present invention.

[0018] Fig. 2 This is a flowchart illustrating the steps of a defect detection method according to an embodiment of the present invention.

[0019] Figure label: 100. Illumination source; 200. Coaxial beam splitter; 300. Objective lens group; 400. Aperture stop; 500. Tube lens group; 600. Polarizing beam splitter; 700. Dual-channel image sensor. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0021] The following is in conjunction with the appendix Figs. 1-2 The specific embodiments of the present invention will be further described in detail below.

[0022] The present invention provides a defect detection method and a defect detection device based on the composite phase microstructure of the objective lens surface. The defect detection device is used to implement the defect detection method, and the defect detection method is used to determine the three-dimensional spatial coordinates (x, y, z) of the defect. The defect detection device includes an illumination source 100, a coaxial beam splitter 200, an objective lens group 300, an aperture stop 400, a tube lens group 500, a polarization beam splitter 600, and a dual-channel image sensor 700.

[0023] In some embodiments, the illumination source 100 is positioned at the front end of the optical path to emit illumination light. Specifically, the illumination source 100 may be a laser diode or a light-emitting diode, and the emission wavelength is preferably in the visible light band (such as 532nm or 633nm). A linear polarizer and a quarter-wave plate are sequentially arranged along the optical axis at the emission end of the light source. The linear polarizer is used to convert the illumination light emitted by the light source into linearly polarized light, and the quarter-wave plate is used to convert the linearly polarized light into circularly polarized light. At this time, the circularly polarized light can be regarded as the equal-amplitude superposition of the left-hand circularly polarized component and the right-hand circularly polarized component.

[0024] In some embodiments, a coaxial beam splitter 200 is positioned between the illumination source 100 and the objective lens group 300. It reflects the illumination light emitted by the illumination source 100 into the objective lens group 300 and transmits the reflected light after sample reflection into the tube lens group 500. Specifically, the coaxial beam splitter 200 employs a non-polarizing beam-splitting prism, placed obliquely at a 45° angle on the optical axis. Specifically, the beam-splitting surface of the coaxial beam splitter 200 forms a 45° angle with the optical axis. The side of the beam splitter closer to the illumination source 100 receives the illumination light, and the beam-splitting surface reflects the illumination light 90° before guiding it into the objective lens group 300 to illuminate the sample. Simultaneously, the beam-splitting surface allows the reflected light after sample reflection to pass through and enter the subsequent tube lens group 500 along the optical axis. The coaxial beam splitter 200 achieves coaxial superposition of the illumination and imaging optical paths, i.e., coaxial illumination, ensuring that the optical axes of the illumination and imaging optical paths coincide, avoiding shadow occlusion and parallax caused by off-axis illumination, thereby improving detection accuracy.

[0025] In some embodiments, the specific lens configuration and focusing imaging principle of the objective lens group 300 and the tube lens group 500 are well known to those skilled in the art and will not be described in detail here. The front focal point of the objective lens group 300 coincides with the surface of the sample to be tested, so that the two axial focal points of the inner and outer rings are focused at different depth positions of the sample to be tested. The rear focal point of the objective lens group 300 coincides with the front focal point of the tube lens group 500, and the rear focal point of the tube lens group 500 coincides with the photosensitive surface of the dual-channel image sensor 700. Specifically, the objective lens group 300 is located above the sample to be tested, and its front focal point is aligned with the sample surface to collect and collimate the reflected light from the sample surface into parallel light. The lens surface near the aperture stop 400 in the objective lens group 300 is integrated with a composite phase microstructure. The composite phase microstructure and the objective lens substrate can be processed by an integral molding process (such as grayscale lithography or electron beam etching), or fixed to the objective lens surface as an independent structural sheet by bonding. The objective lens group 300 and the tube lens group 500 together form a dual telecentric optical path, so that the light rays from different viewpoints on the object side are all incident parallel to the optical axis, and the light rays from different viewpoints on the image side are all emitted parallel to the optical axis. This maintains a constant lateral magnification throughout the entire field of view, eliminating perspective distortion and parallax. In addition, since the composite phase microstructure has rotational symmetry, its modulation of the illumination light does not introduce asymmetric aberrations, enabling uniform illumination. At the same time, the illumination light's outgoing path and the reflected light's return path pass through the same microstructure. The outgoing path modulation expands the effective illumination depth of field and generates a random illumination field distribution, while the return path modulation encodes the defect depth information into detectable interference carriers, speckle textures, and spin differential contrast.

[0026] In some embodiments, the aperture stop 400 is positioned at the rear focal plane of the objective lens group 300 to filter out stray light. Specifically, the aperture stop 400 is a circular aperture, and its position coincides with the rear focal plane of the objective lens group 300. The aperture diameter of the aperture stop 400 determines the numerical aperture of the detection device. For example, in this embodiment, the numerical aperture NA = 0.45. The aperture stop 400 filters out stray light and higher-order diffraction light exceeding the designed numerical aperture, thereby improving image contrast. The tube lens group 500 is positioned behind the aperture stop 400, and its front focal point coincides with the rear focal point of the objective lens group 300, which in turn coincides with the photosensitive surface of the dual-channel image sensor 700. Specifically, the tube lens assembly 500 is composed of multiple achromatic lenses, which are used to converge the parallel light beam after passing through the aperture stop 400 to form an image, so that the parallel light emitted from the aperture stop 400 is focused onto the photosensitive surface of the dual-channel image sensor 700 after passing through the tube lens assembly 500.

[0027] In some embodiments, a polarizing beam splitter 600 is disposed between the lens assembly 500 and the dual-channel image sensor 700 to spatially separate the left-hand circularly polarized component and the right-hand circularly polarized component modulated by the objective lens assembly 300 and guide them respectively to the two photosensitive channels of the dual-channel image sensor 700. Specifically, the polarizing beam splitter 600 can be a polarizing beam splitter prism or a GranThompson prism in conjunction with a dual-sensor layout, or a micro-polarization array in conjunction with a single sensor for partitioned acquisition. A quarter-wave plate can be placed in front of the optical path of the polarizing beam splitter 600 to facilitate the beam splitting process. The incident surface of the polarizing beam splitter prism receives the light from the lens assembly 500, and its beam splitting surface separates the left-hand circularly polarized component and the right-hand circularly polarized component into two spatially separated beams according to their polarization states, which are then guided to the first photosensitive channel and the second photosensitive channel of the dual-channel image sensor 700, respectively. More specifically, the left-hand circularly polarized component passes through the first photosensitive channel to form the first polarization channel light, and the right-hand circularly polarized component passes through the second photosensitive channel to form the second polarization channel light.

[0028] In some embodiments, the two photosensitive channels of the dual-channel image sensor 700 respectively receive the separated left-hand circularly polarized component and right-hand circularly polarized component, and synchronously complete photoelectric conversion. Specifically, the dual-channel image sensor 700 consists of two CMOS or CCD image sensors with matched performance, such as two 2048×2048 pixel CMOS sensors with a pixel size of 3.45μm. These sensors respectively receive the first polarization channel light and the second polarization channel light separated by the polarization beam splitter 600, and synchronously complete photoelectric conversion and analog-to-digital conversion, outputting two digital images, which are respectively denoted as the first polarization channel image. With the second polarization channel image Dual-channel synchronous acquisition eliminates vibration and synchronization errors caused by mechanical scanning or time-division acquisition, ensuring that the two images strictly correspond to the same lighting conditions in time.

[0029] In some embodiments, a composite phase microstructure is integrated on the lens surface of the objective lens group 300 near the aperture stop 400. The composite phase microstructure includes a radially bifocal phase modulation layer, a random phase scattering layer, and a polarization shearing layer arranged sequentially from bottom to top. Specifically, in this embodiment, the composite phase microstructure is fabricated using a one-piece molding method. That is, the radially bifocal phase modulation layer is first fabricated on the substrate, and then random height undulations are superimposed on its surface to form a random phase scattering layer. A polarization shearing layer is then fabricated on the surface of the random phase scattering layer. This polarization shearing layer can be fabricated using micro / nano fabrication techniques, which will not be elaborated here. The radially bifocal phase modulation layer is used to radially divide the pupil into at least two concentric annular regions, each with a different phase profile to introduce different phase profiles. The focal length is used to form a double focal point with a fixed axial spacing. A random phase scattering layer is located on the radial double focal point phase modulation layer, and its surface has a random uneven structure. The height of the uneven structure is 0.1 to 2.0 times the wavelength of the illumination light, and the lateral feature size of the uneven structure is 0.2 to 0.8 times the wavelength of the illumination light. A polarization shearing layer is located on the random phase scattering layer and is used to apply radial linear phase tilts to the left-hand circular polarization component corresponding to the first polarization channel and the right-hand circular polarization component corresponding to the second polarization channel, respectively. The direction of the linear phase tilt applied to the left-hand circular polarization component is opposite to the direction of the linear phase tilt applied to the right-hand circular polarization component. Specifically, the radial double focal point phase modulation layer divides the pupil radially into two regions: an inner ring zone and an outer ring zone. Preferably, the inner ring zone occupies the central region of the pupil, for example, in radially normalized coordinates. The range is The outer ring occupies the outer region of the pupil, for example... Because the inner and outer rings have different effective pupil radii, their axial expansion rates of the defocus spread function differ. The outer ring has a larger radius, making it more sensitive to defocusing, and the blur spot expands faster with increasing depth. The inner ring has a smaller radius, expanding more slowly. The radially normalized coordinate is a dimensionless coordinate between 0 and 1, scaling the actual radius of the pupil (the circular aperture through which the beam passes) proportionally. The inner ring has a first phase profile. This contour is equivalent to introducing a first defocus amount, for example, the first defocus amount is... This allows the light passing through the inner ring to be focused at the first focal point along the axis. The outer ring band has a second phase profile. This contour is equivalent to introducing a second defocus amount, for example, the second defocus amount is... This allows the light transmitted through the outer ring to be focused at the second focal point along the axis. A fixed axial distance is formed between the two focal points. Two beams of light reflected from the same defect and transmitted through the inner and outer rings, respectively, are separated along this axial direction, resulting in an optical path difference related to the defect depth. These beams coherently superimpose at the point of lateral overlap on the image plane, forming a radial shearing interference carrier wave. This phase step is a rotationally symmetric concentric ring structure, which does not introduce any asymmetric phase tilt or pupil shift, thus preserving the dual-telecentric characteristic. The corresponding phase step height difference is 1-1.5. The radial dual-focal phase modulation layer, without reducing the numerical aperture of the system, extends the effective depth of field to 3-5 times that of a traditional dual-telecentric system through annular zone phase modulation on the same objective surface. The random phase scattering layer has a random uneven structure on its surface, which is continuously distributed across the entire pupil. The height distribution of the random uneven structure follows a zero-mean Gaussian statistical probability density function. for,

[0030] Where h is the local height of the concave-convex structure. It is the root mean square of the height, and The wavelength of the illumination light is 0.1 to 0.5 times, and the maximum undulation height does not exceed 2.0 times the wavelength of the illumination light. The lateral feature size of the concave-convex structure is 0.2 to 0.8 times the wavelength of the illumination light, and the height values ​​of any adjacent concave-convex units are statistically uncorrelated, thus ensuring that the phase modulation of the incident wavefront is decorrelated on the lateral feature size scale, so that defects at different axial depths produce statistically distinguishable speckle patterns. The random phase scattering layer, based on the expanded depth of field of the dual-focus phase ladder, further utilizes full-pupil random scattering to generate depth-specific speckle textures, enabling precise depth positioning through statistical matching within a narrow window of the first depth coordinate z1. Simultaneously, this layer is superimposed on the same objective surface as the radial dual-focus phase modulation layer, eliminating the need for additional optical elements and avoiding the common problem of light energy distribution loss in multifocal imaging systems. The polarization shearing layer is a polarization-selective liquid crystal polymer grating, whose phase distribution satisfies ΦL(x,y)=αx, ΦR(x,y)= αx (or along the y-direction), where α is a shearing coefficient matched to the system magnification, ranging from 0.1 to 0.8 rad / mm, preferably 0.3 rad / mm. This layer applies a fixed and opposite linear phase tilt to the left-hand and right-hand circularly polarized components, causing a fixed lateral shift in the on-axis point spread function of the two channels. The amount of shift is independent of the defect depth. Under defocus conditions, the expansion of the blur spot causes this lateral shift to manifest as a depth-dependent change in the overlap area. That is, when focused, the blur spot size is small, and the lateral shift results in less overlap, a deep center depression, and high contrast in the differential image; when defocused, the blur spot size is large, there is more overlap, a shallow center depression, and low contrast in the differential image. The spin differential contrast is monotonically mapped to the defect axial depth z, thereby extracting the depth-related spin differential contrast signal through differential extraction. In addition, the differential operation cancels out the illumination intensity fluctuations, sample reflectivity changes and common-mode noise that occur synchronously in the two polarization channels, thereby improving the image contrast. Since the spin differential contrast only reflects the absolute value of the defocus, the same contrast may be produced when the defect is located at symmetrical depths on both sides of the focal point. Therefore, the spin differential depth estimate z2 is only used as a depth index within the unilateral axial search range defined by the first depth coordinate z1, which is used to eliminate symmetry ambiguity on both sides of the focal point.

[0031] The defect detection method provided in the embodiments of the present invention includes the following steps: S1. The sample to be tested is illuminated by illumination light to generate reflected light. The reflected light is modulated by the objective lens group 300, which integrates a composite phase microstructure, and then separated and collected by the polarization beam splitter 600 to obtain the first polarization channel image and the second polarization channel image. Specifically, dual-channel images are acquired simultaneously in a single exposure, eliminating the need for mechanical axial scanning, thereby improving the detection speed and eliminating vibration errors introduced by mechanical scanning. In addition, by modulating the illumination light through the composite phase microstructure, the information of the wafer is encoded in the first polarization channel image and the second polarization channel image, thereby achieving high-resolution detection with a large depth of field.

[0032] S2. Perform angular averaging on the first and second polarization channel images respectively to obtain a one-dimensional radial intensity distribution. Obtain the radial interference carrier frequency F based on the one-dimensional radial intensity distribution. Specifically, perform defect pre-positioning on the first and second polarization channel images to obtain the diffuse speckle corresponding to each defect. The diffuse speckle is the diffused light spot formed on the image plane after the defect is imaged by the optical system. Establish a polar coordinate system with the geometric center of the diffuse speckle corresponding to each defect as the origin. Integrate and average the image blocks along the circumferential direction to obtain a one-dimensional radial intensity distribution. Since speckle noise is random in the circumferential direction... The distribution, with its positive and negative fluctuations being uncorrelated, exhibits a statistical mean that tends to zero after integration along the circumference, thus being suppressed. In contrast, the axial shear interference carrier from the radial dual-focal phase ladder possesses rotational symmetry, resulting in consistent phase at all angles along the circumference. After integration, it is not canceled out but rather coherently enhanced, thereby converting the weak interference signal in the speckle field into the radial interference carrier frequency F. Due to the constant amplification of the dual-telecentric system and the rotational symmetry of the composite phase microstructure, the point spread function of the defect at any location within the entire field of view is consistent. This method of extracting the one-dimensional radial intensity distribution is universal across the entire field of view, eliminating the need to estimate the point spread function separately for different field points.

[0033] S3. Obtain the first depth coordinate z1 of each defect based on the radial interference carrier frequency F. Specifically, use the monotonic mapping relationship between the radial interference carrier frequency F and the axial depth of the defect to calculate the coarse depth. Using the first depth coordinate z1 as a reference, compress the depth search that originally covered the entire axial range into a window near the first depth coordinate z1, thereby improving the defect location rate.

[0034] S4. Extract speckle texture features from the first polarization channel image and the second polarization channel image respectively, and obtain the depth index of each defect. The depth index is the depth layer identifier corresponding to the highest matching degree after matching the measured speckle texture features of the defect area with the depth.

[0035] S5. Perform a difference operation on the first polarization channel image and the second polarization channel image to obtain a spin difference image, and obtain the spin difference depth estimate z2 based on the spin difference image; specifically, the spin difference depth estimate z2, the first depth coordinate z1 and the depth index are derived from three independent physical mechanisms, that is, through three different criteria, the probability of misjudgment by a single criterion is reduced.

[0036] S6. The axial search range is limited by the first depth coordinate z1. The first polarization channel image and the second polarization channel image are reconstructed, and the images of each depth layer are output to obtain the three-dimensional spatial coordinates (x, y, z) of the defect. Specifically, the axial search range is limited by the first depth coordinate z1. The combination of depth index and spin differential depth estimation z2 improves the accuracy of defect positioning.

[0037] In some embodiments, angular averaging is performed on the first polarization channel image and the second polarization channel image respectively to obtain a one-dimensional radial intensity distribution. This includes pre-locating defects in the first polarization channel image and the second polarization channel image to obtain the diffuse spots corresponding to each defect; establishing a polar coordinate system with the geometric center of each diffuse spot as the origin, and integrating and averaging the two-dimensional local image patch along the circumferential direction to obtain the one-dimensional radial intensity distribution. Specifically, defect pre-location is performed through threshold segmentation, the details of which are not elaborated here. After determining the geometric center of the diffuse spot, within the radius... Local image patches are extracted within a pixel range, and integrated and averaged along the circumferential direction to obtain a one-dimensional radial intensity distribution. This one-dimensional radial intensity distribution is defined as... ,

[0038] in, It is the azimuth angle on the image plane, that is, the angle around the circumference with the geometric center of the diffuse spot as the origin; Before angular averaging, polar coordinates The grayscale value of the two-dimensional image at that location.

[0039] In some embodiments, obtaining the radial interference carrier frequency F based on the one-dimensional radial intensity distribution includes extracting the peak frequency after performing a one-dimensional Fourier transform on the one-dimensional radial intensity distribution, which is then used as the radial interference carrier frequency F. Specifically, since the one-dimensional radial intensity distribution is discretely sampled within a finite radial interval, directly performing a Fourier transform would cause spectral leakage. Therefore, it is necessary to first perform Hanning windowing on the one-dimensional radial intensity distribution, that is, multiply the one-dimensional radial intensity distribution by a Hanning window to smoothly attenuate both ends of the radial distribution to zero, suppressing spectral sidelobe leakage and ensuring the accuracy of peak frequency extraction. Subsequently, the one-dimensional Fourier transform converts the windowed radial intensity distribution from the spatial domain to the frequency domain. The axial shearing interference generated by the radial dual-focus phase modulation layer is cosine modulation in the spatial domain, and after conversion to the frequency domain, it is manifested as a clear spectral peak at the carrier frequency. The background energy is concentrated near the zero-frequency DC in the frequency domain. After eliminating the low-frequency DC component by bandpass filtering, the frequency corresponding to the maximum amplitude is located as the radial interference carrier frequency F. Frequency extraction can be completed with only one Fourier transform and a single peak detection, thereby simplifying the signal processing flow.

[0040] In some embodiments, obtaining the first depth coordinate z1 of each defect based on the radial interference carrier frequency F includes obtaining the first depth coordinate z1 of each defect based on the mapping relationship between the radial interference carrier frequency F and the axial depth of the defect; Wherein, the radial interference carrier frequency F is the digital spatial frequency in pixels on the image plane; j is the pixel size of the dual-channel image sensor 700. The axial distance between the two focal points. The wavelength of the illumination light, The objective lens group has a focal length of 300mm. Specifically, the coarse depth resolution is approximately... This compresses the axial search window to Within the range. The inner and outer ring bands of the radial dual-focus phase modulation layer have different effective pupil radii. At different axial depths of the defect, the radial scale difference of the out-of-focus spread function corresponding to the inner and outer ring bands changes accordingly, causing the radial interference carrier frequency F to have a monotonic mapping relationship with the axial depth of the defect. In some embodiments, speckle texture features are extracted from the first polarization channel image and the second polarization channel image to obtain the depth index of each defect. This includes moving the standard sample sequentially along the axial direction to multiple preset depth positions, acquiring speckle images of the corresponding depth layer at each preset depth position and extracting speckle texture features, establishing a correspondence between each preset depth position and its speckle texture features; extracting speckle texture features from the local regions where each defect is located to obtain measured speckle features; matching the measured speckle features of each defect with the speckle texture features at each preset depth position in the correspondence, and taking the preset depth position with the highest matching degree as the depth index. Specifically, in this embodiment, the standard sample uses a pinhole target, calibrating 101 depth layers at 2μm intervals within a range of ±100μm, constructing a correspondence between each preset depth position and its speckle texture features; calculating local binary mode statistics for the 64×64 pixel local regions where each defect is located to construct measured speckle features; and performing normalized cross-correlation matching between the measured speckle features and the speckle texture features at each preset depth position in the correspondence. Normalized cross-correlation matching is used to measure the similarity of the texture distribution between the measured speckle features and the speckle texture features at each preset depth position in the corresponding relationship. Since the overall brightness of speckles at different defect positions or different depth layers may vary due to the influence of illumination inhomogeneity, directly comparing absolute gray values ​​will mask the true texture similarity and lead to misjudgment. Therefore, the overall intensity of the measured speckle features is first normalized to the same benchmark, and then the overall intensity of the speckle texture features at each preset depth position in the corresponding relationship is normalized to the same benchmark to eliminate brightness differences. Then, the degree of conformity of the texture distribution is compared position by position. The degree of conformity of the texture distribution of the two is quantified into a cross-correlation number between 0 and 1. The closer the value is to 1, the more similar the texture distribution. The depth corresponding to the maximum value of the cross-correlation number is taken as the depth index, with a precision depth resolution of approximately ±2μm. The local binary mode statistics are generated by comparing the gray levels of the 3×3 neighborhood of each pixel within a 64×64 pixel local region. Using the center pixel as a threshold, pixels with gray levels higher than or equal to the center are recorded as 1, and those lower are recorded as 0, forming an 8-bit binary code with a value range of 0 to 255. The frequency of each code in the entire local image is counted to obtain a histogram, which constitutes the measured speckle characteristics. Because the local binary mode only compares the relative brightness of pixels within the neighborhood and does not depend on absolute gray values, it is robust to illumination non-uniformity, sample surface tilt, and global gray-level drift. Furthermore, because the dual telecentric optical path in the defect detection device has spatial invariance, the correspondence is universal across the entire field of view, eliminating the need for point-by-point recalibration.

[0041] In some embodiments, a spin differential image is obtained by performing a difference operation between the first polarization channel image and the second polarization channel image. This includes subtracting the gray values ​​of corresponding pixel positions in the first polarization channel image and the second polarization channel image to obtain the spin differential image. The illumination light, after being reflected by the sample under test and modulated by the composite phase microstructure, contains both left-handed and right-handed circularly polarized components. The polarization shearing layer in the composite phase microstructure applies a linear phase tilt in a first direction to the left-handed circularly polarized component and a linear phase tilt in a second direction opposite to the first direction to the right-handed circularly polarized component, causing the diffuse spots of the same defect in the first and second polarization channel images to undergo radial shifts in opposite directions. Under defect defocusing conditions, the overlapping area of ​​the two diffuse spots changes with the axial depth of the defect. The spin differential image retains the spin differential contrast signal formed by the change in the overlapping area.

[0042] In some embodiments, obtaining the spin differential depth estimate z2 from the spin differential image includes: performing grayscale statistics on the local regions where each defect is located in the spin differential image, extracting the standard deviation of the grayscale values ​​of the pixels in the local regions; and determining the spin differential depth estimate z2 based on the standard deviation. Specifically, before detection, calibration is required. The pinhole standard target is moved along the axial direction at fixed intervals (e.g., 2 μm) to different depth positions. At each position, a first polarization channel image and a second polarization channel image are simultaneously acquired. The grayscale values ​​of corresponding pixel positions in the two channel images are subtracted to obtain the spin differential image required for calibration. Grayscale statistics are performed on the local regions where the pinhole target is located in the spin differential image required for calibration, and the calibration standard deviation of the pixel grayscale values ​​in the local regions is extracted to establish a correspondence between the calibration standard deviation and the depth. During detection, grayscale statistics are performed on the local regions of each defect in the spin differential image, the standard deviation of the pixel grayscale values ​​in the local regions is extracted, and the standard deviation is compared with the calibration standard in the correspondence to obtain the spin differential depth estimate z2.

[0043] In some embodiments, the first polarization channel image and the second polarization channel image are reconstructed to output images of each depth layer and obtain the three-dimensional spatial coordinates (x, y, z) of the defect. This includes determining the axial search range centered on the first depth coordinate z1 and bounded by a preset error tolerance. The first depth coordinate z1 is used as the center because it is directly calculated from the radial interference carrier frequency F through a monotonic mapping relationship. Although the resolution is limited, it locks the true depth of the defect within a narrow window. Using a preset error tolerance, such as ±15μm, as the boundary compresses the search, which would normally cover the entire axial depth range, into this narrow window, thereby improving detection efficiency. Within the axial search range, the first polarization channel image and the second polarization channel image are modeled as the sum of the convolution superposition of the defect distribution at each depth layer and the corresponding depth point spread function, plus the noise term. ,in, Corresponding to the first polarization channel image and the second polarization channel image, when it corresponds to the first polarization channel image, it is actually... When it corresponds to the second polarization channel image, it is actually ; For depth Defect distribution at the location, Let z be the point spread function of the polarization channel at depth z. For noise terms, The term "two-dimensional convolution" represents the modeling of the acquired image as a superposition of the convolutional superposition of the defect distributions at each depth layer and the corresponding depth point spread functions. This is because, under bicentric imaging conditions, the grayscale distribution on the image plane is physically equal to the linear superposition of the diffuse speckles modulated by the system from defects at each depth layer on the image plane; the point spread function... The phase modulation originates from the composite phase microstructure (including the dual-focus defocus of the radial dual-focus phase modulation layer, the speckle modulation of the random phase scattering layer, and the linear phase tilt of the polarization shearing layer). Due to the constant amplification of the dual-telecentric system and the rotational symmetry of the composite phase microstructure, the point spread function remains consistent across the entire field of view and does not change with the lateral field of view position. Therefore, unlike ordinary optical systems, it is not necessary to estimate or update the point spread function separately for different field points. An initial depth value is constructed using the depth index and the spin differential depth estimate z2. Initial depth value The weighted result of the depth index and the spin differential depth estimate z2 (which is well known to those skilled in the art and will not be elaborated here) is used as the center, and within the axial search window defined by the first depth coordinate z1, the weight distribution is assigned to each depth layer, where the initial depth value The tensor is a three-dimensional tensor with a Gaussian weight distribution along the depth dimension, used to guide the subsequent sparse reconstruction algorithm to converge quickly to the ground truth. Utilizing the sparsity constraint of the defect's axial distribution, a multi-channel collaborative optimization objective function is established: ,in, This is the dual-channel data fidelity term, i.e., the square of the L2 norm. This is the axially sparse regularization term, i.e., the L1 norm. This is a depth prior constraint term used to fuse coarse-depth, fine-depth fingerprints and spin-difference estimates, thereby guiding the solution towards the true defect depth during reconstruction optimization. and Here is the regularization parameter, where The value range is 0.01-0.1, preferably 0.05. The value range is 0.05-0.5, preferably 0.1; the multi-channel collaborative optimization objective function is solved by the alternating direction multiplier method, and the tomographic images of each depth layer are output. ; for tomographic images Defect identification is performed to determine the depth layer location z of the defect and its lateral coordinates (x, y) within that depth layer. The three-dimensional spatial coordinates (x, y, z) of the defect are obtained by combining the lateral coordinates and the depth layer location. Specifically, within the axial search range, the continuous depth is discretized into several depth layers, and each layer is assigned an axial depth coordinate z. The tomographic image reconstructed from each layer is then used to output the tomographic image. Defect identification is performed. If a layer contains a preserved connected defect region, the preset axial depth coordinate z corresponding to that layer is the depth layer location where the defect is located. For the tomographic image... Thresholding segmentation and connected component labeling are performed layer by layer. By setting an adaptive threshold (based on the mean and standard deviation of grayscale within the layer, for example, a value of 90), pixels above the adaptive threshold are labeled as candidate defects. Then, connected defect regions are extracted through 8-connected component labeling, and the centroid coordinates of each connected component are calculated as the lateral coordinates (x, y). Combined with the depth layer position z corresponding to the tomographic image, the three-dimensional spatial coordinates (x, y, z) of the defect are obtained. Unlike the depth blurring and artifact problems caused by single-channel reconstruction or simple superposition of two channels in the prior art, this step improves the robustness of axial depth detection and suppresses the occurrence of depth position blurring during the detection process by using three channels: coarse depth, fine depth, and spin difference.

[0044] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A defect detection method based on composite phase microstructures on the objective lens surface, characterized in that, Includes the following steps: The sample to be tested is illuminated with illumination light to generate reflected light. The reflected light is modulated by an objective lens group with an integrated composite phase microstructure and then separated and collected by a polarization beam splitter to obtain a first polarization channel image and a second polarization channel image. The first polarization channel image and the second polarization channel image are respectively subjected to angular averaging to obtain a one-dimensional radial intensity distribution, and the radial interference carrier frequency F is obtained based on the one-dimensional radial intensity distribution. The first depth coordinate z1 of each defect is obtained according to the radial interference carrier frequency F; The speckle texture features of the first polarization channel image and the second polarization channel image are extracted respectively to obtain the depth index of each defect; Perform a difference operation between the first polarization channel image and the second polarization channel image to obtain a spin difference image, and obtain the spin difference depth estimate z2 based on the spin difference image; The axial search range is defined by the first depth coordinate z1. The first polarization channel image and the second polarization channel image are reconstructed, and the images of each depth layer are output to obtain the three-dimensional spatial coordinates (x, y, z) of the defect.

2. The defect detection method according to claim 1, characterized in that, The step of performing angular averaging on the first polarization channel image and the second polarization channel image respectively to obtain a one-dimensional radial intensity distribution includes, Defect pre-positioning is performed on the first polarization channel image and the second polarization channel image to obtain the diffuse spots corresponding to each defect; A polar coordinate system is established with the geometric center of each of the diffuse spots as the origin. The two-dimensional local image patch is integrated and averaged along the circumferential direction to obtain the one-dimensional radial intensity distribution.

3. The defect detection method according to claim 1, characterized in that, The step of obtaining the radial interference carrier frequency F based on the one-dimensional radial intensity distribution includes: The peak frequency is extracted after performing a one-dimensional Fourier transform on the one-dimensional radial intensity distribution and used as the radial interference carrier frequency F.

4. The defect detection method according to claim 1, characterized in that, The step of obtaining the first depth coordinate z1 of each defect based on the radial interference carrier frequency F includes, The first depth coordinate z1 of each defect is obtained based on the mapping relationship between the radial interference carrier frequency F and the axial depth of the defect; Where j is the pixel size of the dual-channel image sensor. The axial distance between the two focal points. The wavelength of the illumination light, This is the focal length of the objective lens group.

5. The defect detection method according to claim 1, characterized in that, The speckle texture features are extracted from the first polarization channel image and the second polarization channel image respectively to obtain the depth index of each defect. include, The standard sample is moved sequentially along the axis to multiple preset depth positions. At each preset depth position, a speckle image of the corresponding depth layer is acquired and speckle texture features are extracted to establish the correspondence between each preset depth position and its speckle texture features. Extract speckle texture features from the local areas where each defect is located to obtain the measured speckle features; The measured speckle features of each defect are matched with the speckle texture features at each preset depth position in the correspondence, and the preset depth position with the highest matching degree is taken as the depth index.

6. The defect detection method according to claim 1, characterized in that, The step of performing a difference operation between the first polarization channel image and the second polarization channel image to obtain a spin difference image includes: Subtract the gray values ​​of corresponding pixel positions in the first polarization channel image from those in the second polarization channel image to obtain the spin difference image; The illumination light, after being reflected by the sample under test and modulated by the composite phase microstructure, contains both left-handed and right-handed circularly polarized components. The polarization shearing layer in the composite phase microstructure applies a linear phase tilt in a first direction to the left-handed circularly polarized component and a linear phase tilt in a second direction opposite to the first direction to the right-handed circularly polarized component, causing the diffuse spots of the same defect in the first polarization channel image and the second polarization channel image to produce radial shifts in opposite directions. Under the condition of defect defocusing, the overlapping area of ​​the two diffuse spots changes with the axial depth of the defect. The spin differential image retains the spin differential contrast signal formed by the change of the overlapping area.

7. The defect detection method according to claim 6, characterized in that, The step of obtaining the spin difference depth estimate z2 based on the spin difference image includes, Gray-level statistics are performed on the local regions where each defect is located in the spin difference image, and the standard deviation of the gray-level values ​​of the pixels in the local regions is extracted. The spin difference depth estimate z2 is determined based on the standard deviation.

8. The defect detection method according to claim 1, characterized in that, The first polarization channel image and the second polarization channel image are reconstructed to output images of each depth layer, thereby obtaining the three-dimensional spatial coordinates (x, y, z) of the defect. include, With the first depth coordinate z1 as the center and the preset error tolerance as the boundary, the axial search range is determined; Within the axial search range, the first polarization channel image and the second polarization channel image are respectively modeled as the sum of the convolution superposition of the defect distribution at each depth layer and the diffusion function at the corresponding depth point, plus the noise term: in, Corresponding to the first polarization channel image and the second polarization channel image; For depth Defect distribution at the location, Let z be the point spread function of the polarization channel at depth z. For noise terms, Represents two-dimensional convolution; An initial depth value is constructed using the depth index and the spin differential depth estimate z2. ; By utilizing the sparsity constraint of the axial distribution of defects, a multi-channel collaborative optimization objective function is established: in, For dual-channel data fidelity, For axially sparse regularization terms, For depth prior constraints, and For regularization parameters; The multi-channel collaborative optimization objective function is solved using the alternating direction multiplier method, outputting tomographic images of each depth layer. ; For the tomographic image Defect identification is performed to determine the depth layer position z of the defect and its lateral coordinates (x, y) within that depth layer. The three-dimensional spatial coordinates (x, y, z) of the defect are obtained by combining the lateral coordinates with the depth layer position.

9. A defect detection device based on the composite phase microstructure of an objective lens surface, characterized in that, For implementing the defect detection method according to any one of claims 1-8, the defect detection device includes an illumination source, a coaxial beam splitter, an objective lens group, an aperture stop, a tube lens group, a polarizing beam splitter, and a dual-channel image sensor. The coaxial beam splitter is disposed between the illumination source and the objective lens group, and is used to reflect the illumination light emitted by the illumination source into the objective lens group, and transmit the reflected light after being reflected by the sample into the tube lens group. The front focal point of the objective lens group coincides with the surface of the sample to be tested, the rear focal point of the objective lens group coincides with the front focal point of the tube lens group, and the rear focal point of the tube lens group coincides with the photosensitive surface of the dual-channel image sensor. The aperture stop is located at the rear focal plane of the objective lens group; The polarization beam splitter is disposed between the objective lens group and the dual-channel image sensor, and is used to spatially separate the left-hand circular polarization component and the right-hand circular polarization component modulated by the objective lens group and guide them to the two photosensitive channels of the dual-channel image sensor respectively. The two photosensitive channels of the dual-channel image sensor are respectively responsible for receiving the separated left-hand circular polarization component and right-hand circular polarization component, and synchronously complete the photoelectric conversion. The objective lens group has a composite phase microstructure integrated on the lens surface near the aperture stop.

10. The defect detection device according to claim 9, characterized in that, The composite phase microstructure includes a radial dual-focus phase modulation layer, a random phase scattering layer, and a polarization shearing layer arranged sequentially from bottom to top. The radial dual-focus phase modulation layer is used to radially divide the pupil into at least two concentric annular regions, each annular region having a different phase profile to introduce different defocus amounts, forming a dual focus with a fixed axial spacing. The random phase scattering layer is located on the radial dual-focus phase modulation layer, and its surface has a random uneven structure. The height of the uneven structure is 0.1 to 2.0 times the wavelength of the illumination light, and the lateral feature size of the uneven structure is 0.2 to 0.8 times the wavelength of the illumination light. The polarization shearing layer is located on the random phase scattering layer and is used to apply radial linear phase tilts to the left-hand circular polarization component corresponding to the first polarization channel and the right-hand circular polarization component corresponding to the second polarization channel, respectively, and the direction of the linear phase tilt applied to the left-hand circular polarization component is opposite to the direction of the linear phase tilt applied to the right-hand circular polarization component.

Images