A product defect detection method based on machine vision

Abstract

The application relates to the technical field of material physics detection, and discloses a product defect detection method based on machine vision, which comprises the following steps: a single pulse energy beam is applied by using a pulse radiation source, a polarization image sequence of a measured material is collected, and a phase delay amount is calculated; the sequence is divided into a first sampling interval dominated by heat conduction and a second sampling interval dominated by elastic relaxation; a background deformation component in the second sampling interval is extracted, and differential hedging is completed on the phase delay amount in the first sampling interval, so that macroscopic thermal expansion signals are filtered out, a local defect characteristic field is obtained, and internal defects are identified; the application uses physical parameters in a thermodynamic relaxation stage as a dynamic reference, solves a time domain aliasing problem of macroscopic deformation noise and microscopic thermal resistance signals under transient thermal excitation, realizes physical decoupling of subsurface micro-defect characteristics and background stress, improves the signal-to-noise ratio of internal stress quantification, and guarantees the physical quantization precision of semiconductor wafer deep layer micro-crack detection.

Description

Technical Field

[0001] This invention belongs to the field of materials physical testing technology, and in particular relates to a product defect detection method based on machine vision. Background Technology

[0002] Currently, in the quality control of semiconductor wafers and special optical glass, product defect detection is a core process to ensure yield. Existing technologies typically employ machine vision recognition, which uses a light source to illuminate the material under test and an image sensor to collect reflected or scattered light signals to identify scratches, chipping, or geometric anomalies on the material surface. The effectiveness of this detection mode depends on the optical contrast between the defect and the background; in essence, it uses the grayscale difference of image pixels to characterize the geometric morphology of the material surface. However, in high-end processes such as wafer manufacturing, the distribution of principal stress fields and subsurface damage within the material have a decisive impact on device performance. These latent defects often lack obvious optical scattering cross sections under normal lighting conditions, appearing optically invisible. This makes grayscale-based detection methods unable to detect the degradation of the material's internal physical properties. To achieve quantitative measurement of the internal stress state, the industry usually tries to introduce physical measurement methods such as holographic interferometry or ultrasonic scanning. However, such solutions are extremely sensitive to mechanical vibrations in industrial environments, and single-point sampling mechanisms are difficult to match the detection efficiency requirements of high-throughput production lines.

[0003] In existing schemes that introduce thermal excitation mechanisms to assist detection, new technical contradictions arise. When transient physical heat is injected into transparent, hard, and brittle materials, it induces macroscopic dynamic thermal expansion of the entire material. This macroscopic structural stress field generated by heating and the local thermal stress field caused by microscopic defects are highly superimposed in the time domain. Existing processing mechanisms usually treat the acquired polarization optical response as a single-dimensional attenuation process, lacking an effective path to decouple macroscopic deformation noise and microscopic lattice thermal resistance distortion signals at the physical mechanism level. This makes it difficult for the system to accurately extract intrinsic features reflecting the true mechanical state of the material when facing unsteady thermal boundary conditions. Hardware configuration affects system stability, and existing photoelastic detection methods have fundamental bottlenecks in dealing with analytical models of dynamic and complex working conditions. For example, publication number CN1 Chinese invention patent application 21558226A discloses a method and apparatus for detecting residual stress in birefringent crystals. It eliminates natural birefringence background and suppresses gradient lines by using multi-angle polarization sampling and spatial mean subtraction algorithm. The underlying logic of this technology is based on the static residual stress field with stable physical properties of the test object. The spatial mean compensation mechanism treats the background as a position-independent constant bias. In the dynamic detection scenario of pulsed thermal excitation, the test material generates an unsteady macroscopic thermal expansion signal that evolves with time. This noise is highly nonlinear in the time domain and its magnitude exceeds the local thermal resistance distortion caused by microscopic defects. Existing technologies lack a dynamic decoupling path for thermoelastic coupling modes and cannot separate transient high-frequency thermal stress and macroscopic structural warping signals in the time domain, causing microscopic defect features to be submerged in the unsteady background.

[0004] Therefore, the technical problem to be solved by this invention is how to achieve high signal-to-noise ratio quantitative measurement of latent defects inside transparent, hard and brittle materials by exploring the intrinsic physical laws of transient thermo-elastic-optical response while maintaining a universally accessible visual hardware architecture. Summary of the Invention

[0005] This invention provides a product defect detection method based on machine vision, comprising the following steps: Step 101: A single-pulse energy beam is projected onto the surface of the material under test using a pulsed radiation source to induce a thermoelastic effect inside the material under test and form a dynamic stress tensor field that evolves over time. Step 102: Capture the polarization image sequence of the material under test under the action of the dynamic stress tensor field using a time-series polarization array camera; Step 103: Perform polarization state calculation on the polarization image sequence to obtain the corresponding Stokes vector, and obtain the polarization phase delay sequence characterizing the birefringence intensity of the internal stress of the tested material based on the Stokes vector. Step 104: Divide the polarization phase delay sequence into a first time sampling interval and a second time sampling interval; the first time sampling interval is locked as the high-frequency phase fluctuation stage caused by thermal conduction, and the second time sampling interval is locked as the linear phase decay stage caused by the elastic relaxation of the structure of the material under test; Step 105: Calculate the average polarization phase delay in the second time sampling interval to establish the background deformation component, and use the background deformation component to differentially offset the polarization phase delay sequence in the first time sampling interval, thereby filtering out the pseudo-phase signal generated by the heating of the material under test and obtaining the local defect feature field. Step 106: Identify gradient abrupt change points in the local defect feature field to determine the internal defect physical parameters of the material under test.

[0006] Preferably, the process of dividing the time sampling interval into a first time sampling interval and a second time sampling interval in step 104 specifically involves: calculating the rate of change curve of the polarization phase delay sequence over time, and identifying the inflection point time when the rate of change curve changes from nonlinear disturbance to linear change; determining the sampling sequence before the inflection point time as the first time sampling interval, and determining the sampling sequence after the inflection point time as the second time sampling interval; wherein, the first time sampling interval corresponds to the transient distortion signal generated by the subsurface defect hindering heat flow conduction, and the second time sampling interval is used to characterize the background signal component caused by the warping of the macroscopic structure of the material under test.

[0007] Preferably, the parameters of the projected single-pulse energy beam in step 101 satisfy the following: the pulse width of the single-pulse energy beam is not greater than 100ms; the single-pulse energy beam is used to excite transient local stress field distortion in the subsurface region of the material under test, and the energy density of the single-pulse energy beam is lower than the damage threshold of the material under test.

[0008] Preferably, step 103 includes the following sub-steps: Step 1031: Calculate the polarization state vector distribution map of the surface of the material under test according to the component parameters in the Stokes vector; Step 1032: Extract the optical phase difference data corresponding to each pixel in the polarization state vector distribution map to obtain the polarization phase delay sequence.

[0009] Preferably, before obtaining the polarization phase delay sequence in step 103, the method further includes the following preprocessing steps: obtaining the initial polarization ground state matrix of the material under test when it is not excited by a pulsed radiation source; performing matrix subtraction operation on the corresponding pixels of the polarization image sequence and the initial polarization ground state matrix to filter out the residual stress noise inherent in the manufacturing process of the material under test.

[0010] Preferably, step 106 includes the following sub-steps: step 1061: mapping the local defect feature field to the depth space geometric coordinate system of the material under test; step 1062: calculating the physical depth and width parameters of the internal defect by detecting the phase gradient extremum points in the local defect feature field.

[0011] Preferably, the method for establishing the background deformation component in step 105 includes: extracting the linear regression slope of the polarization phase delay sequence within the second time sampling interval, and using the linear regression slope as a characteristic quantity characterizing the macroscopic mechanical energy degradation trend of the material under test.

[0012] Preferably, the process of filtering out pseudo-phase signals in step 105 further includes: inputting the background deformation component into a preset physical mode decoupling algorithm to calculate the dynamic phase offset curve generated by the overall thermal expansion of the material under test; and using the dynamic phase offset curve to perform real-time compensation on the sequence within the first time sampling interval.

[0013] Preferably, the material being tested is a transparent semiconductor wafer; the physical parameters of the internal defects include the geometry of microcracks located at the subsurface depth of the transparent semiconductor wafer.

[0014] Compared with existing technologies, the product defect detection method based on machine vision of this invention has the following advantages: 1. In product defect detection, by constructing a transient thermo-elastic-optical physical response detection mechanism, the detection object is transformed from the geometric morphology features of the material surface into the physical evolution process of the internal transient principal stress difference. Under this mechanism, the subsurface microcracks inside the tested material induce local thermal conductivity heterogeneity, resulting in a temporal abrupt change in the phase delay. This transforms the latent defects that originally did not have obvious geometric scattering cross sections under static optical conditions into quantifiable physical relaxation features. This detection method, which starts from the bottom of physics, bypasses the traditional visual dependence on the macroscopic morphology of defects and realizes the direct measurement of the internal mechanical health state of transparent, hard and brittle materials.

[0015] 2. By combining the relaxation time constant of the temporal phase delay with the second-order divergence judgment logic of the spatial dimension, this invention forms a space-time joint physical feature filtering system. This system utilizes the essential difference between local physical feature mutations and global thermal gradient smoothness to eliminate the interference of lighting fluctuations, surface dust, and uneven spatial energy distribution of thermal excitation sources on detection accuracy in industrial production environments. By extracting the degree of dispersion of intrinsic physical properties at each coordinate point as the judgment criterion, it ensures that the detection result depends only on the integrity of the internal structure of the material, thereby improving the measurement reliability of the system under complex thermal boundary conditions.

[0016] 3. The introduced multi-window thermoelastic dynamic mode decoupling path separates the micro-lattice thermal conduction distortion from the macro-structure dynamic bending stress by dividing the physical signal during the thermal relaxation period. Using the physical parameters of the later mechanical elastic recovery dominant stage as a dynamic reference, the high-frequency signal of the earlier thermal conduction dominant stage is reverse-modulated, and the overall material thermal warping noise induced by the excitation source itself is removed. This deep linkage design based on the thermodynamic relaxation law of the material ensures that the micro-defect signal still has a high signal-to-noise ratio recognition under huge dynamic background stress, and reduces the dependence of the detection system on external precision vibration reduction and constant temperature environment. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the defect detection method for timing phase calculation of the present invention; Figure 2 This is a logical architecture diagram of the decoupling of the physical excitation of the thermoelastic effect and the defect characteristics of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0020] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0021] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] A machine vision-based product defect detection method includes the following steps: Step 101: A single-pulse energy beam is projected onto the surface of the material under test using a pulsed radiation source to induce a thermoelastic effect inside the material under test and form a dynamic stress tensor field that evolves over time. Step 102: Capture the polarization image sequence of the material under test under the action of the dynamic stress tensor field using a time-series polarization array camera; Step 103: Perform polarization state calculation on the polarization image sequence to obtain the corresponding Stokes vector, and obtain the polarization phase delay sequence characterizing the birefringence intensity of the internal stress of the tested material based on the Stokes vector. Step 104: Divide the polarization phase delay sequence into a first time sampling interval and a second time sampling interval; the first time sampling interval is locked as the high-frequency phase fluctuation stage caused by thermal conduction, and the second time sampling interval is locked as the linear phase decay stage caused by the elastic relaxation of the structure of the material under test; Step 105: Calculate the average polarization phase delay in the second time sampling interval to establish the background deformation component, and use the background deformation component to differentially offset the polarization phase delay sequence in the first time sampling interval, thereby filtering out the pseudo-phase signal generated by the heating of the material under test and obtaining the local defect feature field. Step 106: Identify gradient abrupt change points in the local defect feature field to determine the internal defect physical parameters of the material under test.

[0023] Preferably, the process of dividing the time sampling interval into a first time sampling interval and a second time sampling interval in step 104 specifically involves: calculating the rate of change curve of the polarization phase delay sequence over time, and identifying the inflection point time when the rate of change curve changes from nonlinear disturbance to linear change; determining the sampling sequence before the inflection point time as the first time sampling interval, and determining the sampling sequence after the inflection point time as the second time sampling interval; wherein, the first time sampling interval corresponds to the transient distortion signal generated by the subsurface defect hindering heat flow conduction, and the second time sampling interval is used to characterize the background signal component caused by the warping of the macroscopic structure of the material under test.

[0024] Preferably, the parameters of the projected single-pulse energy beam in step 101 satisfy the following: the pulse width of the single-pulse energy beam is not greater than 100ms; the single-pulse energy beam is used to excite transient local stress field distortion in the subsurface region of the material under test, and the energy density of the single-pulse energy beam is lower than the damage threshold of the material under test.

[0025] Preferably, step 103 includes the following sub-steps: Step 1031: Calculate the polarization state vector distribution map of the surface of the material under test according to the component parameters in the Stokes vector; Step 1032: Extract the optical phase difference data corresponding to each pixel in the polarization state vector distribution map to obtain the polarization phase delay sequence.

[0026] Preferably, before obtaining the polarization phase delay sequence in step 103, the method further includes the following preprocessing steps: obtaining the initial polarization ground state matrix of the material under test when it is not excited by a pulsed radiation source; performing matrix subtraction operation on the corresponding pixels of the polarization image sequence and the initial polarization ground state matrix to filter out the residual stress noise inherent in the manufacturing process of the material under test.

[0027] Preferably, step 106 includes the following sub-steps: step 1061: mapping the local defect feature field to the depth space geometric coordinate system of the material under test; step 1062: calculating the physical depth and width parameters of the internal defect by detecting the phase gradient extremum points in the local defect feature field.

[0028] Preferably, the method for establishing the background deformation component in step 105 includes: extracting the linear regression slope of the polarization phase delay sequence within the second time sampling interval, and using the linear regression slope as a characteristic quantity characterizing the macroscopic mechanical energy degradation trend of the material under test.

[0029] Preferably, the process of filtering out pseudo-phase signals in step 105 further includes: inputting the background deformation component into a preset physical mode decoupling algorithm to calculate the dynamic phase offset curve generated by the overall thermal expansion of the material under test; and using the dynamic phase offset curve to perform real-time compensation on the sequence within the first time sampling interval.

[0030] Preferably, the material being tested is a transparent semiconductor wafer; the physical parameters of the internal defects include the geometry of microcracks located at the subsurface depth of the transparent semiconductor wafer.

[0031] Example 1: In an industrial setting for screening subsurface microcracks in transparent semiconductor wafers, internal latent defects lack an optical scattering cross-section under static illumination. Furthermore, the overall dynamic thermal expansion noise induced by thermal excitation and the thermal stress field caused by local defects overlap in the time domain, preventing the visual inspection system from establishing a reliable mapping between image features and the internal physical and mechanical state of the material. A pulsed radiation source projects a single-pulse energy beam with a pulse width of no more than 100 ms onto the surface of the material under test. This energy beam induces a thermoelastic effect within the material and generates a dynamic stress tensor field that evolves over time. A time-series polarization array camera captures a multi-frame polarization image sequence of the material under test under the influence of the dynamic stress tensor field. The processor calculates the corresponding Stokes vector based on the polarization image sequence to obtain a polarization phase delay sequence characterizing the birefringence intensity of the internal stress of the material under test. The specific calculation formula satisfies... ;in, for The optical phase delay at time t. and Represent The second and third Stokes parameters obtained by time-mapping calculation.

[0032] The processor extracts the rate-of-change curve of the polarization phase delay sequence over time, identifies the inflection point where the curve transitions from nonlinear perturbation to linear change, and determines the first time sampling interval for the sampled sequence before this inflection point as the first time sampling interval for the transient distortion signal caused by subsurface defects hindering heat conduction. The second time sampling interval for the sampled sequence after this inflection point is determined as the second time sampling interval for the background signal of the overall material structure warping. The mean value of the polarization phase delay within the second time sampling interval is calculated to establish the background deformation component. This background deformation component is used to differentially offset the sequence within the first time sampling interval, filtering out spurious phase signals generated by the heating of the tested material and outputting a local defect feature field. This method is suitable for samples with complex initial warping or non-uniform thermal expansion. For materials with expansion characteristics, conventional static mean offsetting can easily produce baseline residual errors. Based on the objective physical law that the thermoelastic deformation of macroscopic structures exhibits approximately linear damping attenuation during the relaxation phase, the system extracts the linear regression slope of the polarization phase delay sequence within the second time sampling interval as a characteristic quantity characterizing the material's mechanical energy degradation trend. This extrapolation mechanism is input into the physical mode decoupling algorithm. The physical basis for this extrapolation mechanism lies in the fact that the thermal expansion of the macroscopic structure is driven by the overall thermal gradient between the wafer's bottom and surface, and relies on low-frequency bulk acoustic waves for propagation. Its structural thermal inertia is extremely large, and the corresponding response time constant is much larger than the evolution time of microscopic thermal resistance distortion caused by local micro-defects. Therefore, even in the transient high-frequency stage of the first time sampling interval, the thermal expansion is dominated by the overall heating of the material. The macroscopic warping component did not undergo a high-frequency jump with the transient energy injection, but instead maintained a low-frequency, slowly varying pattern at the same timescale as in the subsequent relaxation phase. This ensured that using the posterior slowly varying linear parameters to backfit the anteposterior macroscopic thermal expansion baseline was dynamically self-consistent, thus bridging the timescale gap between the microscopic and macroscopic worlds. The core processing logic is as follows: based on the linear regression slope and the initial phase value of the second time sampling interval, a dynamic phase offset curve spanning the first time sampling interval is constructed by back extrapolation, outputting a low-frequency background stress drift baseline characterizing the overall thermal expansion of the material; the polarization phase delay sequence within the first time sampling interval is compared with the dynamic phase offset curve at corresponding time nodes to complete real-time compensation and output a high-frequency background stress drift baseline. When implementing the modal decoupling algorithm, the signal-to-noise ratio local defect feature field contains two major processing logics: baseline fitting and vector difference. The processor reads the regression slope parameter representing the thermodynamic degradation trend and combines it with the global background bias value at the phase transition point to establish a univariate linear evolution equation with time as the independent variable to serve as the low-frequency approximation function. The algorithm system extracts the discrete sampling timestamp parameters corresponding to the first time sampling interval and substitutes them into the evolution equation for reverse point-by-point operation. This outputs a time-domain baseline compensation vector synchronized with the frame rate of the high-frequency transient image. In memory, the differences between the high-frequency transient sequence vector and the baseline compensation vector at each node are offset frame by frame, thus completing the decoupling operation without introducing new noise.The processor maps the local defect feature field and identifies gradient abrupt change points within it. It calculates the physical depth and width parameters of microcracks located at the subsurface depth of the transparent semiconductor wafer. This mechanism utilizes physical parameters from the thermodynamic relaxation stage as a dynamic benchmark to counteract the temporal aliasing interference between overall deformation noise and local thermal resistance signals under transient thermal excitation. This achieves physical decoupling between subsurface local defect features and background stress, enabling the output of mechanically vulnerable points within the material to be quantifiable physical parameter gradients. This establishes a technical route from surface geometry comparison to the determination of internal physical properties.

[0033] Example 2: In a test scenario targeting the quantification of subsurface microcrack depth on large-size silicon carbide wafers, mechanical vibrations and power frequency electromagnetic interference in industrial settings cause nonlinear distortion of weak photoelastic signals. A transient thermopolarization interferometry measurement platform was constructed, integrating a 1064nm pulsed radiation source and a time-series polarization array camera with an extinction ratio greater than 1000:1 and a sampling rate of 5000Hz. Test data were obtained from continuously acquired sensor sequences from this physical experimental platform. The pulse width of the single-pulse energy beam of the pulsed radiation source was set, taking into account both the thermal wave penetration depth and the temporal resolution of the transient stress peak. When the thermal diffusion time of the target defect feature was assessed to be on the order of milliseconds, the pulse width was reduced to improve the transient excitation energy change rate and prevent transient local thermal stress from being mixed in with the overall thermal expansion signal. Based on the correspondence between the thermal diffusivity of the material under test and the defect depth, the optimal pulse width for exciting the local stress gradient of a 25μm buried microcrack was calculated. For a duration of 20ms, a silicon carbide wafer with a pre-calibrated 25μm deep microcrack was selected as the sample. Gaussian white noise with a signal-to-noise ratio of 20dB was actively injected into the test circuit, and power frequency harmonic interference with a frequency of 50Hz was added. The control group was set to process polarization data using the full-time domain phase mean subtraction method, while the experimental group was set to process polarization data using the time domain dual-interval segmentation and differential offset method. The pulsed radiation source output a 20ms single-pulse energy beam, and the polarization array camera captured 1000 frames of polarization image sequence at a sampling rate of 2000Hz. The processor analyzed the polarization image sequence and output the Stokes vector of each pixel, and calculated the initial polarization phase delay sequence. The initial polarization phase delay sequence contained baseline drift and periodic high-frequency glitches. The rate of change curve of polarization phase delay over time was extracted, and the inflection point time of the rate of change curve from high-frequency violent fluctuation to smooth linear evolution was identified as 45.5ms.

[0034] Based on the inflection point time of 45.5ms, the sampling sequence from 0 to 45.5ms is divided into the first time sampling interval, and the sampling sequence from 45.5ms to 500.0ms is divided into the second time sampling interval. Within the second time sampling interval, the overall wafer structure elastic relaxation dominates the phase evolution. The processor calculates the average polarization phase delay in this interval as 0.12 rad, establishing the dynamic background deformation component. Regarding the data dimension transformation process of directly offsetting the dynamic characteristic field with a single DC scalar, its essence lies in the fact that the projected spot of the pulsed energy beam has uniformly covered the field of view on a macroscopic geometric scale, due to the thermoelasticity of the wafer bulk phase... Stress diffusion has a strong homogenizing effect in a very short time. This macroscopic background stress does not form a spatial gradient within the two-dimensional pixel field of view, but rather is superimposed on the entire signal as an isotropic, spatially flat distribution. Based on this, the system uses matrix broadcasting rules to copy and expand the obtained 0.12 rad pure scalar value into a constant background reference matrix that is completely consistent with the resolution of the time-series polarization image. This ensures that after dimensionality reduction, the point-by-point subtraction calculation of the corresponding matrix elements can be performed logically equivalently in both the spatial and temporal coordinate domains. The experimental group subtracted the 0.12 rad background stress point by point from the transient phase values ​​within the first time sampling interval. For the variable component, the control group calculated a phase mean of 0.18 rad across the entire time domain from 0 to 500.0 ms and applied global subtraction. In the experimental group, after differential offsetting, the local defect feature field exhibited a steep phase aberration at 22.5 ms with a peak value of 0.38 rad, increasing the local signal-to-noise ratio to 18.5 dB. In contrast, the control group's output phase aberration peak was only 0.15 rad, with a local signal-to-noise ratio degraded to 6.2 dB. The internal defect signal was obscured by the photoelastic noise generated by overall thermal warping. This process relied on the pure background component during the structural elastic relaxation period to drive inverse modulation, filtering out baseline drift and overall deformation noise. Boundary effect gradient tests were conducted on the pulse width parameter to monitor the evolution trend of the signal-to-noise ratio (SNR) of the local defect feature field. The pulse width of the single-pulse energy beam was increased from 20ms to 80ms. With the intensification of lateral heat diffusion, the peak phase delay of the local defect feature field decreased to 0.25rad, and the SNR remained above the effective identification threshold of 12.1dB. When the pulse width was set to exceed 100ms and extend to 120ms, the test data showed a nonlinear degradation inflection point. The dynamic range of the saturated polarization array camera was saturated with the overall stress tensor caused by the thermal expansion of the wafer phase, and the SNR of the local defect feature field dropped exponentially to 3.A 1dB decrease causes the system to lose the ability to extract gradient abrupt change points. The nonlinear decay trajectory verification pulse width parameter is limited to a range no greater than 100ms to maintain the temporal separation between the local thermal stress field and the overall thermal expansion field, avoiding mode aliasing failure caused by excessive energy injection. Polarization image sequences are acquired, Stokes vectors are analyzed, and temporal dual-interval segmentation and differential hedging are applied to suppress disturbances from the industrial environment and transient thermal warping of the material. A dynamic background elimination benchmark is established based on the material's internal mechanical relaxation characteristic parameters, constructing a monotonic correspondence between the physical size of microcracks and the polarization phase gradient, thus improving the quantitative accuracy of subsurface latent defect detection in transparent, hard, and brittle media.

[0035] Example 3: This example combines Figures 1 to 2 This paper describes a product defect detection method based on machine vision, such as... Figure 1 The diagram illustrates the specific steps of a machine vision-based product defect detection method. The process begins in step 101, where a single-pulse energy beam is projected onto the surface of the material under test using a pulsed radiation source to induce a thermoelastic effect within the material and form a dynamic stress tensor field that evolves over time. The process then proceeds to step 102, where a time-series polarization array camera captures a sequence of polarization images of the material under test under the influence of the dynamic stress tensor field. Next, in step 103, the polarization state of the polarization image sequence is calculated to obtain the corresponding Stokes vector, and the polarization phase delay sequence, representing the intensity of birefringence stress within the material under test, is obtained based on the Stokes vector. Finally, in step 104, the polarization phase delay sequence is further processed... The sequence is divided into a first time sampling interval and a second time sampling interval. The first time sampling interval is locked to the high-frequency phase fluctuation stage caused by thermal conduction, and the second time sampling interval is locked to the phase linear decay stage caused by the elastic relaxation of the structure of the material under test. Then, the process transitions to step 105, where the average polarization phase delay in the second time sampling interval is calculated to establish the background deformation component. The background deformation component is then used to differentially offset the polarization phase delay sequence in the first time sampling interval, thereby filtering out the pseudo-phase signal generated by the heating of the material under test and obtaining the local defect feature field. Finally, in step 106, the gradient abrupt change points in the local defect feature field are identified to determine the internal defect physical parameters of the material under test.

[0036] like Figure 2As shown, the network architecture starts with the intrinsic stress background of the initial polarized ground state material as the source. Through the path of projecting a single-pulse energy beam, it drives the system state to transition to the thermoelastic effect physical excitation node in the dynamic stress tensor field state. Subsequently, relying on the temporal polarization image capture and solution process, the physical response is mapped and output as a polarization phase delay sequence stress birefringence intensity mapping node. On this basis, the system follows the branch of the inflection point time of the rate of change curve to enter the node dominated by thermal conduction in the first time sampling interval and the node dominated by elastic relaxation in the second time sampling interval of the time-domain segmented state. After that, along the logical evolution line that establishes the background deformation component offset, the data is sent to the local defect feature field background deformation component differential offset node. Then, by identifying gradient mutation points and determining actions, it extends to the internal defect physical parameter geometric size quantification and qualitative node. Finally, the state chain, along the closed loop path that completes the deep space geometric coordinate mapping, points back to the initial node, constructing a complete data decoupling and feature closed loop extraction framework.

[0037] Example 4: In industrial scenarios involving the inspection of semiconductor wafers made of multiple materials, the difference in thermal diffusivity of different substrate materials causes fluctuations in the evolution cycle of the internal thermal stress field. Fixed parameter sampling windows cannot adapt to the dynamically changing thermal properties of the materials, resulting in physical obstacles such as phase reference instability and blurred detection features. The processor acquires the thermal diffusivity parameters of the material under test. Calculate the pulse width of the single-pulse energy beam projected by the pulsed radiation source. The specific calculation formula satisfies ;in, The preset thermal influence depth feature length, The unit is cm² / s; for a substrate material with a thermal diffusivity of 0.82 cm² / s, the following is set... It is 0.12cm, calculated as follows The pulse width is 17.5 ms; the pulsed radiation source is based on this pulse width. An energy beam is projected to confine the thermoelastic effect to a subsurface depth range, preventing heat from dissipating disorderly into the bulk of the material. After acquiring the polarization phase delay sequence, the processor calculates the second-order time derivative curve of the polarization phase delay sequence and monitors the moment when its amplitude approaches zero and enters a preset environmental noise range, defining this moment as the inflection point time. The system uses this inflection point time as a basis. The system identifies the physical state phase transition point from transient nonlinear thermal expansion to quasi-linear structural elastic relaxation, dynamically divides the first time sampling interval into a second time sampling interval, and sets the focal plane coordinates of the polarization array camera as the reference XY plane to achieve the mapping from two-dimensional image to spatial coordinates. The normal direction of the thermal wave diffusion into the material is set as the Z-axis. Since the longitudinal propagation time of the thermal wave in the solid medium is directly related to the thermal diffusivity of the material, the system establishes a reconstruction algorithm based on the time-of-flight principle. The time node when the pixel point shows the phase extremum in the local defect feature field is converted into the physical depth value in the Z-axis direction according to the thermal diffusion equation. At the same time, by extracting the magnification calibration matrix of the camera optical lens system, the two-dimensional pixel coordinates in the XY plane are multiplied by the above matrix to project onto the actual physical surface of the material under test. Thus, a deterministic mathematical mapping model of two-dimensional polarization features and three-dimensional depth space geometric coordinate system is established.

[0038] The processor maps a local defect feature field and extracts the maximum polarization phase gradient value in the local defect feature field. The physical width of subsurface microcracks is calculated based on the physical size mapping operator. The calculation formula satisfies ;in, Based on the scaling factor calibrated using the photoelastic constant of the tested material, this physical model relies on the intrinsic properties of the material to eliminate signal aliasing caused by macroscopic thermal expansion. It transforms the extracted optical feature distribution into specific physical geometric dimensions, quantifies the physical parameters of microcracks within media with different thermophysical properties, and during gradient abrupt change point extraction, the system traverses all pixels in the local defect feature field. It applies a differential filtering operator to the phase delay within adjacent coordinates to solve for the first-order spatial derivative to generate a feature gradient matrix. To determine mathematically significant abrupt changes, the system uses the sensor's background noise level measured under no-load conditions. The standard deviation is set to set a screening threshold. In the gradient matrix, all zero-crossing coordinates whose amplitude exceeds the screening threshold and whose second-order spatial divergence shows a sign reversal are searched. The system identifies these zero-crossing coordinates as mutation centers and extracts the maximum first-order derivative value distributed in their neighborhood as the quantitative basis for determining the distortion intensity of the internal defect boundary. Based on the physical law of heat transfer where the diffusion depth of transient heat waves in a solid medium is proportional to the square root of time, the processor extracts the time nodes corresponding to the maximum polarization phase gradient in the local defect feature field and calculates the physical depth parameters of the subsurface microcracks. The physical depth calculation formula satisfies... In logical relationships, Characterizes the physical depth of internal microcracks. Represents the thermal diffusivity of the tested material. The parameter indicating the time delay from the start of single-pulse energy beam injection to the point where the polarization phase gradient of the corresponding pixel reaches its maximum value. This is a dimensionless system calibration constant used to eliminate geometric errors introduced by the non-uniformity of the spatial distribution of the pulsed radiation source spot. The specific value is determined by scanning a pre-fabricated calibration block with a known standard depth microcrack and deriving it back into the formula. The legal value is between 0.85 and 1.15. The numerical boundary of this calibration constant directly corresponds to the physical failure limit of the system under specific optical hardware conditions. When the value of this constant derived from the calibration block falls below 0.85, it indicates that the projected spot has a severe collapse of radiation energy in the central region and causes abnormal lateral backflow of heat flow in the surrounding area, causing the physical premise of one-dimensional depth heat diffusion to completely collapse. If the derived value exceeds the upper limit of 1.15, it reflects that the scattering at the outer edge of the beam is too intense. At this time, the intensity of the lateral parasitic birefringence signal will completely mask the intrinsic extreme value brought about by the depth feature. Therefore, any setting that exceeds the reasonable range of this constant is regarded as the spatial distribution of the excitation source is unreliable and will directly trigger the system's underlying pre-alignment intervention alarm.

[0039] Example 5: In the pre-calibration process of multi-batch semiconductor wafer inspection, the fluctuation of material composition causes the stress birefringence coefficient to shift, resulting in measurement error of physical size. The system acquires a set of standard wafers with an internal preset width of 10μm to 50μm. The processor drives the pulse radiation source to project a single pulse energy beam with a pulse width of 20ms onto the standard wafers, and the time-series polarization array camera captures the corresponding polarization image sequence.

[0040] The processor processes the polarization image sequence to obtain the local defect feature field and extracts the polarization phase gradient maxima sequence corresponding to each standard microcrack. ,in The processor assigns index numbers to standard microcracks and summarizes reference values ​​for the physical width of gradient microcracks. With polarization phase gradient maximum sequence The scaling factor is calculated using the least squares method. , proportional conversion factor The calculation formula is as follows: ;in, This is the scaling factor. For the first A reference value for the physical width of a standard microcrack. To extract the corresponding polarization phase gradient maxima, the processor uses this scaling factor. Write to memory and, when inspecting the batch of wafers under test, according to the scaling factor The extracted phase feature values ​​are converted into physical geometric parameters to eliminate quantization bias caused by fluctuations in material composition.

[0041] Example 6: In the initialization scenario of deploying the detection system or replacing the load-bearing component of the material under test, the phase component generated by the surface unevenness of the load-bearing component will suppress the subsurface defect characteristics. The system drives the pulse radiation source to project a single pulse energy beam in the no-load state, and the time-series polarization array camera captures and solves the background phase distribution matrix. The processor calculates the background phase distribution matrix. Standard deviation of phase fluctuation of each pixel And based on the background phase distribution matrix The statistical characteristics determine the initial detection sensitivity threshold of the system. Initial detection sensitivity threshold The calculation formula is as follows: ;in, This is the initial detection sensitivity threshold. This is a preset proportional coefficient. The standard deviation of the pixel phase fluctuation is used to establish a quantization benchmark for the inherent background noise of the system, providing a ground state reference for the differential offsetting step in the subsequent detection process, and eliminating signal offset caused by differences in hardware environment.

[0042] When the system is applied to batches of wafers with different surface features, the processor initiates a sampling frequency calibration program. By traversing sampling frequencies within the range of 1000Hz to 5000Hz, the signal-to-noise ratio of the local defect feature field at each frequency is monitored. The processor identifies the signal-to-noise ratio based on the changing trend. The critical frequency point where the curve enters a plateau and no longer increases with increasing sampling frequency is determined as the optimal sampling frequency. Among them, signal-to-noise ratio The calculation formula is as follows: ;in, For signal-to-noise ratio, For defective signal power, For background noise power, use this optimal sampling frequency The system drives a time-series polarization array camera to capture images, enabling it to capture transient characteristics of the thermoelastic effect while meeting data processing bandwidth limitations, and maintaining the quantification accuracy of physical parameters of subsurface microcracks.

[0043] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A product defect detection method based on machine vision, characterized in that, Includes the following steps: Step 101: A single-pulse energy beam is projected onto the surface of the material under test using a pulsed radiation source to induce a thermoelastic effect inside the material under test and form a dynamic stress tensor field that evolves over time. Step 102: Capture the polarization image sequence of the material under test under the action of the dynamic stress tensor field using a time-series polarization array camera; Step 103: Perform polarization state calculation on the polarization image sequence to obtain the corresponding Stokes vector, and obtain the polarization phase delay sequence characterizing the birefringence intensity of the internal stress of the tested material based on the Stokes vector. Step 104: Divide the polarization phase delay sequence into a first time sampling interval and a second time sampling interval; the first time sampling interval is locked as the high-frequency phase fluctuation stage caused by thermal conduction, and the second time sampling interval is locked as the linear phase decay stage caused by the elastic relaxation of the structure of the material under test; Step 105: Extract the linear regression slope of the polarization phase delay sequence in the second time sampling interval, use the linear regression slope as a feature quantity to characterize the macroscopic mechanical energy degradation trend of the material under test to establish the background deformation component, and use the background deformation component to differentially offset the polarization phase delay sequence in the first time sampling interval to filter out the pseudo phase signal generated by the heating of the material under test and obtain the local defect feature field. Step 106: Map the local defect feature field to the depth space geometric coordinate system of the material under test, identify gradient abrupt change points by detecting the phase gradient extreme points in the local defect feature field, and calculate the physical depth and width parameters of the internal defects to determine the physical parameters of the internal defects of the material under test.

2. The product defect detection method based on machine vision according to claim 1, characterized in that, The process of dividing the time sampling interval into the first time sampling interval and the second time sampling interval in step 104 is as follows: calculate the rate of change curve of the polarization phase delay sequence over time, and identify the inflection point time when the rate of change curve changes from nonlinear disturbance to linear change; determine the sampling sequence before the inflection point time as the first time sampling interval, and determine the sampling sequence after the inflection point time as the second time sampling interval; wherein, the first time sampling interval corresponds to the transient distortion signal generated by the subsurface defect hindering heat flow conduction, and the second time sampling interval is used to characterize the background signal component caused by the warping of the macroscopic structure of the material under test.

3. The product defect detection method based on machine vision according to claim 1, characterized in that, In step 101, the parameters for projecting a single-pulse energy beam are: the pulse width of the single-pulse energy beam is not greater than 100 ms; the single-pulse energy beam is used to excite transient local stress field distortion in the subsurface region of the material under test, and the energy density of the single-pulse energy beam is lower than the damage threshold of the material under test.

4. The product defect detection method based on machine vision according to claim 1, characterized in that, Step 103 includes the following sub-steps: Step 1031: Calculate the polarization state vector distribution map of the surface of the material under test based on the component parameters in the Stokes vector; Step 1032: Extract the optical phase difference data corresponding to each pixel in the polarization state vector distribution map to obtain the polarization phase delay sequence.

5. The product defect detection method based on machine vision according to claim 1, characterized in that, Before obtaining the polarization phase delay sequence in step 103, the method further includes the following preprocessing steps: obtaining the initial polarization ground state matrix of the material under test when it is not excited by a pulsed radiation source; performing matrix subtraction operation on the corresponding pixels of the polarization image sequence and the initial polarization ground state matrix to filter out the residual stress noise inherent in the manufacturing process of the material under test.

6. The product defect detection method based on machine vision according to claim 1, characterized in that, The process of filtering out pseudo-phase signals in step 105 also includes: inputting the background deformation component into a preset physical mode decoupling algorithm to calculate the dynamic phase offset curve generated by the overall thermal expansion of the material under test; and using the dynamic phase offset curve to perform real-time compensation on the sequence within the first time sampling interval.

7. The product defect detection method based on machine vision according to claim 1, characterized in that, The material under test is a transparent semiconductor wafer; the physical parameters of internal defects include the geometry of microcracks located at the subsurface depth of the transparent semiconductor wafer.

Images