A surface damage intelligent segmentation method based on multi-angle polarization imaging

Abstract

This invention belongs to the field of image processing technology, specifically relating to an intelligent surface damage segmentation method based on multi-angle polarization imaging. It includes the following steps: Step 1, continuous variable-angle polarization image acquisition: a rotating wedge prism beam deflection device, composed of a first wedge prism and a second wedge prism cascaded along the same optical axis, is used to construct a polarization image sequence containing N incident angle positions; Step 2, joint modeling of polarization bidirectional reflectance distribution functions: a polarization reflectance parameter feature map is constructed from the polarization reflectance parameter vectors of all pixels; Step 3, optimal transmission domain adaptive damage segmentation: using the polarization reflectance parameter distribution of a non-damaged standard sample as the domain alignment benchmark, pixel-level damage segmentation results of the surface of the workpiece under test are output. This invention effectively overcomes the problems of insufficient sensitivity of traditional single-angle illumination schemes for weak damage detection, sensitivity of end-to-end deep learning schemes to non-damage interference factors, and performance degradation in cross-batch workpiece detection.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and particularly relates to an intelligent sensing system for surface damage detection through image generation, specifically a surface damage intelligent segmentation method based on multi-angle polarization imaging. Background Technology

[0002] Surface damage detection is a crucial step in quality control during industrial manufacturing, widely used in the factory inspection of products such as metal parts, optical components, and semiconductor wafers. Traditional surface inspection methods primarily rely on ordinary bright-field imaging technology, which acquires grayscale images of the workpiece surface under fixed-angle white light illumination, followed by manual visual interpretation or image processing algorithms based on grayscale thresholds for damage identification. These methods are prone to masking damage information when dealing with smooth, specularly reflective metal surfaces due to saturation in highlight areas; and when facing workpiece surfaces with machining textures, texture interference and damage features are difficult to distinguish in the grayscale image, resulting in a high rate of missed detections of fine scratches and shallow corrosion pits.

[0003] In recent years, polarization imaging technology has been gradually introduced into the field of surface inspection. By setting polarization elements at the illumination or receiving end, information such as the degree of polarization and polarization azimuth angle of reflected light can be obtained, and damage enhancement can be achieved by utilizing the difference in polarization response between damaged and normal areas. However, existing polarization detection schemes typically use a single-angle illumination method with a fixed incident angle, which can only acquire polarization reflection information of the surface under a single geometric condition. Since different types of damage have different degrees of polarization response at different incident angles, single-angle illumination is difficult to simultaneously take into account the detection sensitivity of multiple damage types. In addition, most existing schemes directly input polarization images into deep learning segmentation networks for end-to-end processing, lacking modeling of the physical process of surface reflection, resulting in the network being sensitive to non-damage factors such as illumination fluctuations and camera gain drift, and having insufficient generalization ability. In actual production line applications, the polarization reflection characteristics of workpieces of the same material exhibit systematic drift due to differences in different production batches and surface treatment processes. The detection performance of the pre-trained segmentation model drops significantly after switching workpiece batches, requiring the re-collection of labeled samples and retraining of the model, which seriously restricts the industrial promotion of polarization detection technology. Summary of the Invention

[0004] The main objective of this invention is to provide a surface damage intelligent segmentation method based on multi-angle polarization imaging. This invention organically combines multi-angle polarization acquisition, physical parameter inversion and domain adaptive segmentation, effectively overcoming the problems of insufficient sensitivity of traditional single-angle illumination schemes for weak damage detection, sensitivity of end-to-end deep learning schemes to non-destructive interference factors, and performance degradation of cross-batch workpiece detection. It takes into account detection accuracy, generalization ability and engineering feasibility.

[0005] To address the aforementioned technical problems, this invention provides a smart surface damage segmentation method based on multi-angle polarization imaging, comprising the following steps: Step 1, Continuous variable angle polarization image acquisition: A rotating wedge prism beam deflection device is used, which consists of a first wedge prism and a second wedge prism cascaded along the same optical axis. By controlling the difference in rotational speed of the two wedge prisms, the incident angle of the illumination beam on the surface of the workpiece is continuously changed. At each incident angle position, a split-focus plane polarization camera is used to acquire reflected polarization images, forming a polarization image sequence containing N incident angle positions. Step 2, Joint modeling of polarization bidirectional reflectance distribution function: For each pixel in the polarization image sequence, calculate the normalized polarization component at each incident angle position, establish a polarization bidirectional reflectance distribution function model and fit it to obtain the polarization reflectance parameter vector, and construct the polarization reflectance parameter feature map from the polarization reflectance parameter vectors of all pixels. Step 3, Optimal Transmission Domain Adaptive Damage Segmentation: Using the polarization reflection parameter distribution of the undamaged standard sample as the domain alignment reference, the polarization reflection parameter feature map is mapped to the reference domain through the entropy regularization optimal transmission method to obtain the domain-aligned polarization reflection parameter feature map. The domain-aligned polarization reflection parameter feature map is input into the pre-trained fully convolutional damage segmentation network to output the pixel-level damage segmentation result of the surface of the workpiece to be tested.

[0006] Furthermore, in step 1, the first wedge prism and the second wedge prism have the same wedge angle and are driven to rotate around the optical axis by independent servo motors. By setting the angular velocity of the first wedge prism to be greater than that of the second wedge prism, the deflection angle of the illumination beam changes continuously with the relative rotation angle between the two wedge prisms. Within the preset acquisition range, a scanning segment with a continuously expanding incident angle is selected as the effective acquisition segment, so that the incident angle of the illumination beam on the surface of the workpiece to be measured gradually changes from the minimum incident angle to the preset maximum incident angle within the effective acquisition segment, and the scanning trajectory unfolds in a spiral.

[0007] Furthermore, in step 1, a linear polarizer and an achromatic quarter-wave plate are fixedly installed sequentially along the beam propagation direction in the illumination optical path. The fast axis direction of the achromatic quarter-wave plate forms a 45-degree angle with the transmission direction of the linear polarizer, so that the illumination beam is converted into circularly polarized light.

[0008] Furthermore, in step 1, the polarization sensor of the split-focus plane polarization camera integrates linear polarization microanalyzers in four directions (0°, 45°, 90°, and 135°) within each 2x2 superpixel unit; an angle trigger position is set on the servo motor encoder, and whenever the relative rotation angle between the first and second wedge prisms reaches a preset equally spaced angle position within the effective acquisition segment, a hardware trigger signal is sent to the split-focus plane polarization camera to acquire one frame of original polarization mosaic image.

[0009] Furthermore, for each frame of the original polarized mosaic image, using 2x2 superpixel units as the basic unit, full-resolution grayscale images of the four polarization directions (0°, 45°, 90°, and 135°) are recovered through edge-guided interpolation. The full-resolution grayscale images of the four polarization directions acquired at all incident angle positions within the effective acquisition segment are arranged in ascending order of incident angle to form a polarized image sequence.

[0010] Furthermore, in step 2, the process of calculating the normalized polarization component is as follows: For each incident angle position, the sum of the gray values ​​of the current pixel in the 0-degree direction and the 90-degree direction is taken as the total light intensity component; when the total light intensity component is less than the preset intensity threshold, the observation at the current incident angle position is marked as an invalid observation and excluded in the subsequent objective function calculation; when the total light intensity component is greater than or equal to the preset intensity threshold, the difference between the gray values ​​in the 0-degree direction and the 90-degree direction is divided by the total light intensity component to obtain the first normalized polarization component, and the difference between the gray values ​​in the 45-degree direction and the 135-degree direction is divided by the total light intensity component to obtain the second normalized polarization component; the above calculation is repeated for all N incident angle positions to obtain the effective observation sequence of the first normalized polarization component and the effective observation sequence of the second normalized polarization component of the current pixel.

[0011] Furthermore, in step 2, the polarization bidirectional reflection distribution function model represents the reflected light as the superposition of diffuse reflection and specular reflection components. The diffuse reflection component is determined by both the diffuse reflectance parameter and the surface refractive index parameter, and its polarization characteristics originate from the Fresnel transmission effect when light passes through the medium interface twice. The specular reflection component is determined by the specular reflectance parameter, surface roughness parameter, and surface refractive index parameter. The normal direction of the micro-surface element follows an isotropic distribution controlled by the surface roughness parameter, and the reflection on each micro-surface element follows Fresnel's reflection law. The model contains four parameters to be fitted: diffuse reflectance parameter, specular reflectance parameter, surface roughness parameter, and surface refractive index parameter. Physically feasible value ranges are set for each of the four parameters to be fitted as constraint boundaries. The fitting results of adjacent pixels or the prior values ​​corresponding to the material categories are used as the initial values ​​for iteration. The first and second normalized polarization components predicted by the polarization bidirectional reflectance distribution function model under the incident angle conditions corresponding to the effective observations are used as the model output. The sum of squared residuals between the model output and the effective observation sequence at all effective observation positions is used as the objective function. The Levenberg-Marquardt iterative optimization algorithm with constrained boundaries is used for optimization. The iteration stops when the change between two adjacent values ​​of the objective function is less than the preset convergence threshold. The optimal estimates of the four parameters to be fitted are obtained. The polarization reflectance parameter vector is formed by arranging the parameters in the order of diffuse reflectance, specular reflectance, surface roughness, and surface refractive index.

[0012] Furthermore, in step 3, a polarization image sequence of an undamaged standard sample of the same material type as the workpiece to be tested is pre-acquired. The standard sample polarization reflection parameter feature map is obtained by processing it according to the procedures of steps 1 and 2. The polarization reflection parameter vectors of all pixels in the standard sample form a domain-aligned reference sample set. In addition, a polarization image sequence of a training sample of the same material type containing damaged and undamaged regions and having pixel-level damage annotations is pre-acquired. The training sample polarization reflection parameter feature map is obtained by processing it according to the procedures of steps 1 and 2. A fully convolutional damage segmentation network is pre-trained on the polarization reflection parameter feature map of the training sample and the corresponding pixel-level damage annotations.

[0013] Furthermore, in step 3, the specific process of the entropy-regularized optimal transmission method is as follows: The polarization reflection parameter vectors of all pixels in the polarization reflection parameter feature map are used to form a test sample set; in the 4-dimensional space where the polarization reflection parameter vectors are located, the four coordinate axes of diffuse reflectance, specular reflectance, surface roughness, and surface refractive index are uniformly divided into M parts, forming M to the power of 4 grid cells; the number of samples falling into each grid cell in the domain-aligned reference sample set is counted, and divided by the total number of samples in the domain-aligned reference sample set to obtain the reference domain discrete distribution vector; the number of samples falling into each grid cell in the test sample set is counted, and divided by the total number of samples in the test sample set to obtain the test domain discrete distribution vector; the cost matrix is ​​constructed using the Euclidean distance between the centers of any two grid cells; based on the cost matrix and the preset entropy... The kernel matrix is ​​constructed using regularization coefficients. Each element in the kernel matrix is ​​equal to the negative value of the corresponding element in the cost matrix, divided by the entropy regularization coefficient, and then the natural exponent. All elements of the baseline domain scaling vector and the target domain scaling vector are initialized to 1. In each iteration, the baseline domain scaling vector is first updated by dividing the baseline domain discrete distribution vector element-wise by the product of the kernel matrix and the target domain scaling vector. Then, the target domain scaling vector is updated by dividing the target domain discrete distribution vector element-wise by the product of the transpose of the kernel matrix and the baseline domain scaling vector. Iteration stops when the maximum element change of both the baseline domain scaling vector and the target domain scaling vector between two adjacent iterations is less than a preset stopping threshold. The optimal transfer matrix is ​​constructed from the baseline domain scaling vector, the kernel matrix, and the target domain scaling vector. The row index of the optimal transfer matrix corresponds to the grid cell number of the target domain, and the column index corresponds to the grid cell number of the baseline domain.

[0014] Furthermore, in step 3, the process of performing centroid mapping for each polarization reflection parameter vector in the sample set to be tested is as follows: determine the grid cell number of the test domain to which the current polarization reflection parameter vector belongs, extract all elements of the row corresponding to that number from the optimal transfer matrix, divide the elements of that row by the sum of the elements of that row to obtain the transfer share vector, multiply each element in the transfer share vector by the coordinates of the reference domain grid cell center of the corresponding column index, and then sum them column by column to obtain the global mapping result; then calculate the offset of the current polarization reflection parameter vector relative to the center of the grid cell of the test domain to which it belongs, and superimpose the offset onto the global mapping result to obtain the mapped polarization reflection parameter vector; perform the above centroid mapping on all polarization reflection parameter vectors, and arrange the mapped polarization reflection parameter vectors according to their original spatial positions to form a domain-aligned polarization reflection parameter feature map.

[0015] This invention provides an intelligent surface damage segmentation method based on multi-angle polarization imaging, which has the following advantages: The invention achieves continuous scanning of the incident angle of the illumination beam through a rotating wedge prism beam deflection device. Compared with traditional fixed-angle polarization detection schemes, it can provide polarization reflection information covering multiple incident angles for each pixel on the surface of the workpiece, fully stimulating the differences in polarization response of different types of damage at different incident angles, and significantly improving the detection sensitivity of weak damage such as fine scratches and shallow corrosion pits. By employing joint modeling of the polarization bidirectional reflection distribution function, the original polarization image is transformed into a polarization reflection parameter feature map with clear physical meaning. Compared with the end-to-end scheme of directly inputting the polarization image into the deep learning network, the extracted physical parameters have stronger robustness to non-destructive interference factors such as illumination fluctuations and camera gain drift, reducing the segmentation network's dependence on the training data scale. An optimal transmission method based on entropy regularization is introduced to adaptively align the polarization reflection parameter distribution offset between the workpiece under test and the reference domain. This allows the pre-trained segmentation network to operate stably without retraining or fine-tuning for each batch of workpieces, solving the problem of significant performance degradation in existing polarization detection technologies when switching workpiece batches. This also reduces the deployment and maintenance costs of industrial production lines. The overall technical approach forms a closed chain from continuous variable-angle polarization acquisition to physical parameter inversion and then to adaptive domain segmentation. Each link supports the others, balancing detection accuracy, cross-batch generalization ability, and engineering feasibility. Attached Figure Description

[0016] Figure 1 A schematic diagram of the surface reflection principle of the polarization bidirectional reflection distribution function model provided in this embodiment of the invention; Figure 2 Multi-angle response curves and fitting comparison diagrams of normalized polarization components as a function of incident angle provided in embodiments of the present invention; Figure 3 The polarization reflection parameter feature map and damage segmentation result map are provided for embodiments of the present invention. Detailed Implementation

[0017] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0018] A surface damage intelligent segmentation method based on multi-angle polarization imaging includes the following steps: Step 1, Continuous variable angle polarization image acquisition: A rotating wedge prism beam deflection device is used, which consists of a first wedge prism and a second wedge prism cascaded along the same optical axis. By controlling the difference in rotational speed of the two wedge prisms, the incident angle of the illumination beam on the surface of the workpiece is continuously changed. At each incident angle position, a split-focus plane polarization camera is used to acquire reflected polarization images, forming a polarization image sequence containing N incident angle positions. Step 2, Joint modeling of polarization bidirectional reflectance distribution function: For each pixel in the polarization image sequence, calculate the normalized polarization component at each incident angle position, establish a polarization bidirectional reflectance distribution function model and fit it to obtain the polarization reflectance parameter vector, and construct the polarization reflectance parameter feature map from the polarization reflectance parameter vectors of all pixels. Step 3, Optimal Transmission Domain Adaptive Damage Segmentation: Using the polarization reflection parameter distribution of the undamaged standard sample as the domain alignment reference, the polarization reflection parameter feature map is mapped to the reference domain through the entropy regularization optimal transmission method to obtain the domain-aligned polarization reflection parameter feature map. The domain-aligned polarization reflection parameter feature map is input into the pre-trained fully convolutional damage segmentation network to output the pixel-level damage segmentation result of the surface of the workpiece to be tested.

[0019] Continuous variable-angle polarization image acquisition is the front-end data acquisition stage of the entire intelligent surface damage segmentation method. Its purpose is to provide polarization reflection information covering multiple incident angles for each pixel on the surface of the workpiece under test, thereby providing sufficient angle sampling data for subsequent joint modeling of polarization bidirectional reflection distribution functions. Compared with traditional fixed-angle illumination imaging schemes, continuous variable-angle illumination can excite the differences in polarization reflection of surface materials at different incident angles. Especially near large incident angles, the difference in polarization degree between the damaged area and the normal area is more significant, which is beneficial for distinguishing the physical parameters of the two types of areas in subsequent steps.

[0020] Continuous variable-angle polarization image acquisition includes the construction of the illumination optical path, the structure and operation of the rotating wedge prism beam deflection device, polarization state modulation, polarization image acquisition at the receiving end, and de-mosaic processing of the polarization mosaic image, which will be described in detail below.

[0021] The illumination source employs a high-power collimated LED or halogen lamp in conjunction with a collimating lens group to generate a nearly parallel illumination beam. The spectral range of the light source can be selected according to the material characteristics of the workpiece surface under test. For metal workpiece surfaces, a narrowband light source with a center wavelength in the range of 520 nm to 660 nm can obtain good polarization contrast; for dielectric workpiece surfaces, a broadband white light source is also applicable, but a narrowband filter must be used at the receiving end to eliminate the influence of dispersion on polarization measurement. The power of the light source must ensure that, under the preset maximum incident angle, the intensity of reflected light from the surface of the workpiece under test is still sufficient to keep the grayscale value of the split-plane polarization camera within the linear response range. In one optional embodiment, the light source uses an LED with a center wavelength of 630 nm and a full width at half maximum (FWHM) of 10 nm, with an output light power of 5 watts. After being shaped by an achromatic collimating lens group with a focal length of 100 mm, the output beam diameter is 25 mm, and the divergence angle is less than 0.5 degrees.

[0022] The rotating wedge prism beam deflection device consists of a first wedge prism and a second wedge prism cascaded along the same optical axis. Both wedge prisms use the same optical material and the same wedge angle; the optical material can be BK7 optical glass or fused silica. The size of the wedge angle determines the maximum deflection capability of a single wedge prism. In practical selection, if the wedge angle is too small, the incident angle scanning range is insufficient; if the wedge angle is too large, the aberrations and dispersion of the beam after passing through the prism increase. In one embodiment, both wedge prisms have a wedge angle of 10 degrees, a diameter of 30 mm, and are made of BK7 optical glass. In this case, the deflection angle of a single wedge prism for a 632.8 nm wavelength beam is approximately 5.2 degrees. After cascading the two wedge prisms, the total deflection angle range is from 0 degrees to approximately 10.4 degrees. The incident angle range on the surface of the workpiece can be covered from 0 degrees to approximately 60 degrees by adjusting the optical path geometry.

[0023] refer to Figure 1 The figure illustrates two reflection processes that occur after circularly polarized incident light illuminates the surface of a workpiece, and the formation mechanism of their polarization characteristics. The incident beam is aligned with the normal to the macroscopic surface. The angled direction illuminates the surface of the workpiece to be measured. Let be the angle of incidence. At the surface interface, the incident light is split into two parts. One part of the light undergoes Fresnel reflection directly on the surface micro-element, forming a specular reflection component. The micro-element is a microscopic model of the actual rough surface. Each micro-element is considered an optical plane at a local scale, and its normal direction does not completely coincide with the macroscopic surface normal, but rather obeys a parameter determined by the surface roughness. The isotropic statistical distribution of the control. The larger the value, the greater the dispersion of the normal direction of the micro-surface element, and the rougher the surface. The right side of the figure shows several micro-surface elements and their respective normal directions, visually demonstrating the random distribution characteristics of the normal directions. The reflection on each micro-surface element follows Fresnel's law of reflection, with different reflection coefficients for the s-polarization component perpendicular to the incident plane and the p-polarization component parallel to the incident plane; therefore, specularly reflected light exhibits a linear polarization state. The specular reflection component is determined by the specular reflectivity parameter. Surface roughness parameter and surface refractive index parameter The other portion of the light transmits through the surface interface into the interior of the medium, undergoes multiple random scatterings within the medium, and then exits from the surface, forming a diffuse reflection component. The tortuous path below the surface in the figure illustrates the volume scattering process of light inside the medium. Volume scattering itself tends to completely depolarize the polarization state of the light, but the light passes through the air-medium interface twice when entering and exiting the medium. Fresnel transmission occurs each time the light crosses the interface, and the transmission coefficients for the s-polarization component and the p-polarization component are different. The accumulation of these two Fresnel transmission effects causes the diffuse reflection light to exhibit a weak linear polarization state. The diffuse reflection component is determined by the diffuse reflectance parameter. and surface refractive index parameter The parameters were jointly determined. The figure also indicates the four parameters to be fitted in the polarization bidirectional reflectance distribution function model: diffuse reflectance parameter... Specular reflectivity parameter Surface roughness parameter and surface refractive index parameter When surface damage exists, the microstructure of the damaged area changes, causing the values ​​of these four parameters to deviate from the normal range. This deviation forms the physical basis for subsequent damage segmentation.

[0024] The first and second wedge prisms are mounted on independent hollow rotary tables, which are coaxially nested and each driven by a servo motor via synchronous belt or direct drive to rotate around the optical axis. Each servo motor is equipped with a high-resolution photoelectric encoder with an angular resolution of no less than 0.01 degrees, used to provide real-time feedback on the absolute rotation angle of each wedge prism. The two servo motors are coordinated and driven by the same motion controller, which can set the angular velocity and starting phase of each servo motor in real time.

[0025] The working principle of the rotating wedge prism beam deflection device is as follows. When the illumination beam passes sequentially through the first and second wedge prisms along the optical axis, each wedge prism applies a deflection to the beam. The direction of deflection is determined by the azimuth angle of the current wedge prism, and the amplitude of deflection is determined by both the wedge angle and the refractive index of the material. Let the first wedge prism be at time... The azimuth is The second wedge prism at time The azimuth is The deflection angle amplitude produced by each wedge prism when acting alone is Then, after passing through two wedge prisms, the resultant deflection angle of the illumination beam is... and the combined deflection azimuth angle This can be expressed as: the resultant deflection angle Synthetic deflection azimuth angle .in, The magnitude of the deflection angle produced by a single wedge prism on the illumination beam is given by the wedge angle. and the refractive index of the prism material at the working wavelength The specific relationship is determined jointly. ; and The first and second wedge prisms are respectively located at time... The rotation angle relative to the initial reference azimuth. The above relationship shows that the resultant deflection angle... It depends only on the relative angle between the two wedge prisms The cosine value is half of the value. When the relative rotation angle between the two wedge prisms is 0 degrees, the deflection effects of the two prisms are completely superimposed, and the combined deflection angle reaches its maximum value. When the relative rotation angle is 180 degrees, the deflection effects of the two prisms cancel each other out, and the combined deflection angle is 0 degrees.

[0026] In actual data acquisition, the incident angle of the illumination beam needs to gradually change from a small value to a large value over a period of time to obtain a polarization image sequence with the incident angle gradually expanding. To achieve this goal, the angular velocity of the first wedge prism is set to... The angular velocity of the second wedge prism is set to ,and Both are rotating at a constant speed. At this point, the relative angle between the two wedge prisms increases linearly with time, i.e. ,in Let be the angular velocity of the first wedge prism. The angular velocity of the second wedge prism, Let be the relative rotation angle between the two wedge prisms at the initial moment. Since the resultant deflection angle is proportional to the cosine of the relative rotation angle, within half a cycle as the relative rotation angle linearly changes from 180 degrees to 0 degrees, the resultant deflection angle continuously increases from 0 degrees to its maximum value. The incident angle of the illumination beam on the surface of the workpiece gradually expands from its minimum value to its maximum value. This half-cycle constitutes the effective acquisition segment. Simultaneously, due to the synthesized deflection azimuth angle... Related to the average of the two angular velocities, the deflection direction of the illumination beam continuously rotates within the effective acquisition range, thus the scanning trajectory of the illumination beam on the surface of the workpiece exhibits a spiral-like pattern. This can be achieved by adjusting the initial relative rotation angle. The value can control the starting position of the effective acquisition segment throughout the entire rotation cycle.

[0027] In one specific implementation, Set to 36 degrees per second. Set at 30 degrees per second, angular velocity difference The initial relative rotation angle is 6 degrees per second. The angle is set to 180 degrees. The time required for the relative angle to linearly decrease from 180 degrees to 0 degrees is 30 seconds. Within this 30-second effective acquisition period, the synthesized deflection angle continuously increases from 0 degrees to approximately 10.4 degrees. If the distance between the surface of the workpiece and the rotating wedge prism beam deflection device is 200 mm, the corresponding incident angle coverage range is approximately 0 to 55 degrees. In another optional implementation, the effective acquisition period can be shortened to 6 seconds by increasing the angular velocity difference to 30 degrees per second. This is suitable for scenarios with fast online detection cycles, but the dwell time at each incident angle position is shortened, requiring a higher frame rate camera or a higher power light source.

[0028] After the illumination beam is deflected by the rotating wedge prism beam deflector, it needs to pass through a polarization-modulated optical path before reaching the surface of the workpiece. In this path, a linear polarizer and an achromatic quarter-wave plate are fixedly installed sequentially along the beam propagation direction. The linear polarizer converts the unpolarized or partially polarized illumination beam into linearly polarized light, while the achromatic quarter-wave plate converts the linearly polarized light into circularly polarized light. The fast axis of the achromatic quarter-wave plate forms a 45-degree angle with the transmission direction of the linear polarizer. In this configuration, the electric vector of the linearly polarized light forms 45-degree components with both the fast and slow axes of the achromatic quarter-wave plate, which are then combined into circularly polarized light after a 1 / 4-wavelength phase delay. Circularly polarized light is used as the illumination polarization state because it is insensitive to the rotation of the illumination beam around its own optical axis. During the operation of the rotating wedge prism beam deflection device, although the deflection azimuth angle of the illumination beam continuously changes, the polarization state of the circularly polarized light remains unchanged under rotational symmetry. No additional polarization state modulation is introduced due to changes in the beam deflection azimuth angle, thus ensuring the consistency of the illumination polarization state at different incident angles. If linearly polarized light is used for illumination, the incident surface azimuth of the illumination beam on the workpiece surface changes accordingly when the beam deflection azimuth angle changes, resulting in inconsistent orientation of the illumination polarization state relative to the incident surface at different incident angles, increasing the complexity of subsequent polarization reflection modeling. An achromatic quarter-wave plate is selected to ensure that the delay deviation from 1 / 4 wavelength is minimized within the working wavelength range, resulting in a high ellipticity ratio for the output circularly polarized light. In one embodiment, the achromatic quarter-wave plate employs a super-achromatic design suitable for the visible light band of 400 nm to 700 nm, with a delay deviation of less than 1 nm across the entire band. The linear polarizer and achromatic quarter-wave plate are installed on the exit side of the rotating wedge prism beam deflection device, i.e., after the wedge prism group and before the surface of the workpiece to be measured. In another optional embodiment, the linear polarizer and achromatic quarter-wave plate can also be installed before the wedge prism group, i.e., between the light source and the first wedge prism, with the same effect. However, it should be noted that the birefringence of the wedge prism itself may have a slight influence on the polarization state. In this case, a low-stress optical glass material should be selected to reduce the birefringence effect.

[0029] The receiving end employs a focal plane polarization camera. The core component of this camera is a focal plane polarization sensor. On the sensor's photosensitive surface, a micro-linear polarization analyzer array is arranged in 2x2 superpixel units. Within each 2x2 superpixel unit, four sub-pixels cover micro-linear polarization analyzers in four transmission directions: 0°, 45°, 90°, and 135°. The 0° and 90° directions are orthogonal, as are the 45° and 135° directions. The micro-linear polarization analyzers typically utilize a metal nanowire grid structure, directly integrated above the sensor pixels, with an extinction ratio of at least 100:1. The sensor can simultaneously acquire image information from all four polarization analysis directions in a single exposure, eliminating the need for mechanically rotating polarizers or timing switches. Therefore, it can achieve transient polarization imaging during continuous changes in the incident angle of the illumination beam, avoiding motion artifacts or incident angle drift caused by time-division multiplexing. In one embodiment, the focal plane polarizing camera uses an industrial camera equipped with a Sony IMX250MZR sensor. The sensor has an effective pixel count of 2448 x 2048 and a pixel size of 3.45 micrometers. After de-mosaicing using 2 x 2 superpixel units, the effective resolution is 1224 x 1024, and the frame rate can reach 75 frames per second at full resolution. In another optional embodiment, an industrial camera equipped with a Sony IMX264MZR sensor can be used, with an effective pixel count of 2448 x 2048 and a pixel size of 3.45 micrometers. This global shutter is suitable for scenes requiring shorter exposure times to suppress motion blur.

[0030] Synchronization between the image acquisition of the split-focus plane polarization camera and the beam deflection device of the rotating wedge prism is achieved through hardware triggering. Specifically, the azimuth angles of the two wedge prisms are read in real time from the encoders of the servo motors of the first and second wedge prisms. and The relative rotation angle is calculated in real time by the motion controller. Within the relative angle range of the effective acquisition segment, N equally spaced angle trigger positions are pre-set. Whenever the real-time calculated relative angle reaches one of the trigger positions, the motion controller sends a level transition signal to the focal plane polarization camera via a hardware trigger line. Upon receiving the trigger signal, the focal plane polarization camera immediately performs an exposure and acquires one frame of the original polarized mosaic image. Since each trigger position corresponds to a specific relative angle value, and the relative angle value uniquely determines the composite deflection angle and thus the incident angle, the incident angle corresponding to each frame is precisely known. In one embodiment, the relative angle range within the effective acquisition segment is 180 degrees to 0 degrees, and N is set to 60. Therefore, the relative angle interval between two adjacent trigger positions is 3 degrees, corresponding to an incident angle interval of approximately 1 degree. In another optional embodiment, N is set to 30, the interval between adjacent trigger positions is 6 degrees, and the incident angle interval is approximately 2 degrees. This is suitable for scenarios with lower angle sampling density requirements but a need for faster acquisition speed. The trigger delay between the motion controller and the focal plane polarization camera should be controlled within 100 microseconds to ensure that the angular deviation between the trigger position and the actual exposure position is less than 0.01 degrees.

[0031] The structure of each frame of the original polarized mosaic image is as follows: the sensor collects grayscale values ​​in only one polarization direction at each pixel location, and the entire image presents a 2x2 periodically arranged polarized mosaic pattern. To obtain a full-resolution grayscale image in each polarization direction, the original polarized mosaic image needs to be de-mosaiced. The de-mosaicing process uses 2x2 superpixel units as the basic processing unit, and the goal is to recover the full-resolution grayscale images for each of the four polarization directions (0°, 45°, 90°, and 135°) from the mosaic sampling. While directly using nearest neighbor or bilinear interpolation is simple, it introduces significant edge blurring errors in the estimation of polarization degree and polarization angle. Therefore, an edge-guided interpolation method is adopted, which detects the gradient direction of local pixel grayscale during the interpolation process and performs weighted interpolation along the direction with smaller grayscale changes, thereby reducing aliasing across edges while maintaining spatial resolution. Specifically, for each pixel location requiring interpolation, the absolute values ​​of the grayscale differences between adjacent pixels with the same polarization direction in the horizontal and vertical directions are calculated. The direction with the smaller grayscale difference is selected as the main interpolation direction, and linear interpolation is performed using the grayscale values ​​of adjacent pixels with the same polarization direction in the main interpolation direction. When the grayscale differences in the horizontal and vertical directions are similar, the average of the interpolation results in the two directions is taken. After de-mosaic processing, each frame of the original polarized mosaic image is decomposed into four full-resolution grayscale images, corresponding to polarization directions of 0 degrees, 45 degrees, 90 degrees, and 135 degrees, respectively.

[0032] refer to Figure 2 The figure contains two sub-figures, left and right. The vertical axis of the left sub-figure represents the first normalized polarization component. The vertical axis of the right subplot represents the second normalized polarization component. The horizontal axis of both sub-figures represents the angle of incidence, in degrees. Defined as the difference between the grayscale value at 0 degrees and the grayscale value at 90 degrees, divided by the total light intensity component. , Defined as the difference between the grayscale value at 45 degrees and the grayscale value at 135 degrees, divided by the total light intensity component. ,in This is the sum of the gray values ​​at 0 degrees and 90 degrees. Each sub-image contains two model prediction curves: one for the normal region and one for the damaged region. The normal region corresponds to one set of polarization reflection parameters, while the damaged region corresponds to a different set. The difference between these two sets of parameters results in significant differences in the shape and amplitude of the two curves. As the incident angle gradually increases from a small value to approximately 55 degrees, the normal region and the damaged region... and The differences between the response curves gradually increase because the difference between the s-component and p-component of the Fresnel reflection coefficient is more significant at large incident angles, and the polarization effect is more prominent. The region between the two curves is the region of polarization response difference caused by damage. The larger the area of ​​this difference region, the stronger the discrimination ability of subsequent parameter fitting and damage segmentation. The left sub-figure simultaneously marks valid and invalid observations. Valid observations come from the incident angle positions in the polarization image sequence determined by the validity of the total light intensity component. They are distributed near the model curve in the normal region and exhibit certain random fluctuations, reflecting the influence of sensor noise in actual measurements. Invalid observations correspond to incident angle positions where the total light intensity component is lower than the preset intensity threshold. These observations are marked as invalid due to their low signal-to-noise ratio and are excluded in the subsequent calculation of the objective function. The fitting process of the polarization bidirectional reflection distribution function model is to adjust the values ​​of the four parameters to be fitted so that the model prediction curve is as close as possible to the valid observations, minimizing the sum of squared residuals between the model output and the valid observation sequence at all valid observation positions.

[0033] Four full-resolution grayscale images with polarization directions are obtained by acquiring all N incident angle positions within the effective acquisition segment and performing de-mosaic processing. These images are then arranged in ascending order of incident angle to form a polarization image sequence. The polarization image sequence contains N incident angle positions, and each incident angle position contains four full-resolution grayscale images with polarization directions. Therefore, the entire polarization image sequence contains 4 x N grayscale images. In one embodiment, N is 60, so the polarization image sequence contains 240 grayscale images. Each grayscale image in the polarization image sequence is labeled with its corresponding incident angle value and polarization direction number, facilitating the subsequent step of reading the polarization grayscale value at each incident angle position pixel by pixel.

[0034] Throughout the continuous variable-angle polarization image acquisition process, the workpiece under test is fixed on a height-adjustable workpiece stage. The surface normal of the workpiece stage is aligned with the optical axis of the rotating wedge prism beam deflection device at zero deflection angle. A focal plane polarization camera is mounted directly above or to the side of the workpiece stage, with its lens optical axis pointing towards the surface of the workpiece. Both coaxial and off-axis geometric layouts are acceptable. In a coaxial layout, the illumination beam and the receiving beam share the same optical axis, requiring a beam splitter prism to separate them. In an off-axis layout, the illumination and receiving beams are positioned on opposite sides of the surface normal of the workpiece, maintaining a fixed geometric relationship between the illumination incident angle and the receiving observation angle. Each layout has its advantages and disadvantages: the coaxial layout is compact, but the beam splitter prism introduces light energy loss and additional polarization effects; the off-axis layout has a simple optical path and no beam splitting loss, but the receiving angle needs to be included in the geometric calculations when jointly modeling the polarization bidirectional reflection distribution function. In one specific embodiment, an off-axis layout is adopted, and the angle between the incident direction of the illumination beam and the normal of the workpiece stage surface is controlled by a rotating wedge prism beam deflection device to vary within the range of 0 to 55 degrees. The angle between the lens optical axis of the split-plane polarizing camera and the normal of the workpiece stage surface is fixed at 0 degrees, that is, the camera observes the surface of the workpiece to be measured vertically from directly above.

[0035] Joint modeling of polarization bidirectional reflectance distribution functions is a crucial step in the entire method, transforming raw polarized image data into physically meaningful surface feature representations. By physically modeling and parametrically inverting the polarization response of each pixel in the polarized image sequence at multiple incident angles, the surface optical state of each pixel can be reduced to a few physical parameters. These parameters directly reflect information such as the surface's diffuse reflection characteristics, specular reflection characteristics, micro-roughness, and material refractive index. Damaged regions exhibit systematic differences in polarization reflection behavior compared to normal regions due to altered surface microstructures. This difference is more stable and significant in the physical parameter space than in the original image grayscale space. Therefore, using physical parameters instead of the original grayscale as input to the subsequent segmentation network can significantly reduce the interference of non-destructive factors such as illumination fluctuations and camera gain drift on the segmentation results.

[0036] Solving for the normalized polarization components is the first step in the joint modeling of the polarization bidirectional reflection distribution function. For each pixel in the polarization image sequence, all N incident angle positions are traversed one by one. At each incident angle position, the gray values ​​of the current pixel in the 0-degree, 45-degree, 90-degree, and 135-degree directions are read from the corresponding full-resolution grayscale images in the four polarization directions. The gray value in the 0-degree direction is denoted as... The grayscale value at a 45-degree angle is denoted as The grayscale value in the 90-degree direction is denoted as The grayscale value in the 135-degree direction is denoted as ,in , , , These represent the grayscale response values ​​of the current pixel after passing through the linear polarization analyzers at the current incident angle position at 0 degrees, 45 degrees, 90 degrees, and 135 degrees, respectively.

[0037] First, calculate the total light intensity component. It is defined as the sum of the gray values ​​in the 0-degree direction and the gray values ​​in the 90-degree direction, i.e. Total light intensity component This characterizes the total intensity of reflected light received by the current pixel at the current incident angle, unaffected by the polarization analysis direction. The reason for using the sum of the 0° and 90° orthogonal directions to estimate the total light intensity is that the sum of the transmitted intensities of light in any polarization state along the two orthogonal polarization analysis directions is always equal to the total intensity of the incident light. This property is guaranteed by a corollary of Malus's law under orthogonal analyzer conditions. In practical applications, minor deviations caused by factors such as the limited extinction ratio of the micro-linear polarization analyzer and inconsistent pixel responses can be corrected through factory calibration.

[0038] The total light intensity component was calculated. Afterwards, its validity needs to be determined. When When the value is less than the preset intensity threshold, it indicates that the reflected light intensity of the current pixel at the current incident angle is too low, which may be caused by factors such as local surface shading, edge attenuation of the beam illumination range, or workpiece edge occlusion. Under such low light intensity conditions, the grayscale value is mainly dominated by sensor dark current noise and readout noise, and the normalized polarization component calculated accordingly will produce large numerical fluctuations, causing serious interference to subsequent parameter fitting. Therefore, when When the intensity is less than a preset intensity threshold, the observation at the current incident angle is marked as invalid and excluded from subsequent objective function calculations. The preset intensity threshold should take into account both the sensor's dark current level and the standard deviation of the readout noise. In one embodiment, the sensor's readout noise standard deviation is approximately 2.5 gray units, and the preset intensity threshold is set to 10 times this standard deviation, i.e., 25 gray units. In this case, the noise-induced deviation of the normalized polarization component does not exceed 0.08, and its impact on fitting accuracy is controllable. In another optional embodiment, the preset intensity threshold can be automatically determined based on calibration experiments: Dark-field images are acquired under no-illumination conditions, the mean and standard deviation of the entire image's grayscale are statistically analyzed, and the preset intensity threshold is set to the dark-field mean plus 8 times the standard deviation.

[0039] when When the intensity is greater than or equal to a preset intensity threshold, the first normalized polarization component is calculated. and the second normalized polarization component The first normalized polarization component is calculated as follows: the difference between the gray value at 0 degrees and the gray value at 90 degrees is divided by the total light intensity component. The second normalized polarization component is calculated by dividing the difference between the grayscale value at 45 degrees and the grayscale value at 135 degrees by the total light intensity component. .in, It characterizes the degree of linear polarization asymmetry of the reflected light in the polarization direction from 0 degrees to 90 degrees. It characterizes the linear polarization asymmetry of the reflected light in the polarization direction from 45 degrees to 135 degrees. Together, they describe the complete directional and intensity information of the polarization state of the reflected light. and The values ​​of both are between -1 and +1. When the reflected light is completely unpolarized, both are 0. When the reflected light is completely linearly polarized, the sum of their squares is 1.

[0040] Using normalized polarization components and Using the normalized polarization component as the fitting target, rather than directly using the degree of polarization and polarization azimuth, has important numerical calculation considerations. The polarization azimuth exhibits periodic equivalence; for example, 0 degrees and 180 degrees physically correspond to the same polarization direction. If the polarization azimuth is directly used as the fitting target for residual calculation, discontinuous large residual jumps will occur when the azimuth approaches the periodic boundary, leading to unstable convergence of the iterative optimization algorithm. and As a linear combination of the sine and cosine of the polarization azimuth angle, it naturally eliminates the problem of angular periodicity. The residual function is continuous and smooth over the entire value range, ensuring the stable convergence of the subsequent Levenberg-Marquardt iterative optimization algorithm.

[0041] After repeating the above calculation process for all N incident angle positions, the first normalized polarization components corresponding to all incident angle positions that have passed the validity determination are arranged sequentially to form the effective observation sequence of the first normalized polarization component of the current pixel; the corresponding second normalized polarization components are arranged sequentially to form the effective observation sequence of the second normalized polarization component of the current pixel. The number of elements in each effective observation sequence is equal to the number of effectively observed incident angle positions, which does not exceed N and may vary depending on the pixel position. In one embodiment, N is 60, and the number of effective observations for most pixels is between 50 and 60. Only in the workpiece edge region, where the beam exceeds the workpiece range when the incident angle is large, the number of effective observations drops to about 30, which is still sufficient to support stable fitting of the four parameters.

[0042] After completing the normalized polarization component calculation, the next step is to establish and fit the polarization bidirectional reflectance distribution function model. The polarization bidirectional reflectance distribution function model describes the relationship between the polarization state of reflected light and the polarization state of incident light under given illumination and observation directions. In this method, the polarization bidirectional reflectance distribution function model represents surface reflected light as a superposition of diffuse reflection and specular reflection components. This decomposition is based on two fundamental physical processes of optical reflection from solid surfaces: one part of the incident light is directly reflected at the surface interface, forming the specular reflection component; the other part of the incident light is transmitted into the medium below the surface, undergoes multiple scatterings, and then exits from the surface, forming the diffuse reflection component. The polarization characteristics of the two components are generated by different mechanisms, therefore they need to be modeled separately.

[0043] The diffuse reflection component is determined by the diffuse reflectance parameter. and surface refractive index parameter Determined jointly. Among them... This represents the proportion of total energy of incident light that exits from the surface after being scattered by the volume. This represents the refractive index of the surface material of the workpiece under test at the working wavelength. The polarization characteristics of the diffuse reflection component do not originate from the volume scattering process itself, because the polarization state tends to be completely depolarized after light undergoes multiple random scatterings within the medium. The polarization characteristics of the diffuse reflection component originate from the Fresnel transmission effect when light passes through the medium interface twice: when incident light is transmitted from air into the medium, the transmission coefficients are different for different polarization directions; when it is transmitted back from the medium to air after volume scattering, there are also polarization-related transmission differences. The superposition effect of these two Fresnel transmissions makes the diffuse reflection light exhibit a weakly linearly polarized state, with the degree of polarization increasing with the incident angle, and the polarization direction perpendicular to the incident plane. For a given incident angle... ( (where is the angle between the illumination beam and the surface normal). The polarization characteristics of the reflected light from the diffuse component can be calculated by the difference between the s-component and p-component of the Fresnel transmission coefficient, where the s-component and p-component correspond to the polarization directions perpendicular to and parallel to the incident surface, respectively.

[0044] The specular reflection component is determined by the specular reflectivity parameter. Surface roughness parameter and surface refractive index parameter Determined jointly. Among them... This indicates the proportion of total energy reflected by a specular surface. This indicates the degree of deviation of the surface micro-element normal direction from the macroscopic surface normal. The micro-element model assumes that the actual rough surface is composed of a large number of tiny planes (micro-elements), each of which is an optical plane at a local scale, and reflections on it strictly obey Fresnel's law of reflection. The micro-element normal direction does not completely coincide with the macroscopic surface normal, but rather obeys a parameter determined by the surface roughness. The control is an isotropic distribution. The smaller the value, the more concentrated the micro-surface element normals are near the direction of the macro-normals, the smoother the surface, and the more concentrated the specular reflection; The larger the value, the more dispersed the distribution of the normals of the micro-surface elements, the rougher the surface, and the more diffuse the specular reflection. The reflection on each micro-surface element follows Fresnel's law of reflection, with different reflection coefficients for the s-component and p-component. Therefore, the specularly reflected light also exhibits a linear polarization state. Near the grazing angle (i.e., when the incident angle is close to 90 degrees), the difference between the s-component and p-component of the Fresnel reflection coefficient is the greatest, and the degree of polarization of the specularly reflected light increases significantly. This is the physical basis for why the polarization difference between the damaged area and the normal area is more prominent at large incident angles.

[0045] In a specific model implementation, the distribution of micro-surface element normals adopts the Beckmann distribution, and the probability density function of the Beckmann distribution is determined by the surface roughness parameter. and the zenith angle of the micro-element normal ( The Beckmann distribution is determined by the angle between the normal direction of the micro-facet and the normal of the macroscopic surface. It is widely used in optical scattering modeling because it can well describe the statistical properties of the micro-facet normals of randomly rough surfaces and contains only a single parameter. This helps control the total number of parameters to be fitted. In another optional implementation, the micro-element normal distribution can also adopt the GGX distribution. The GGX distribution decays more slowly at the tail end in the direction of large scattering angles than the Beckmann distribution, making it more suitable for describing rough surfaces with long trailing scattering characteristics. The choice between the two distributions does not affect the overall method flow; the differences lie only in fitting accuracy and the range of applicable materials.

[0046] The polarization bidirectional reflectance distribution function model contains four parameters to be fitted: diffuse reflectance parameter. Specular reflectivity parameter Surface roughness parameter and surface refractive index parameter Given a set of parameter values, the model can operate at any incident angle. The theoretical values ​​of the first and second normalized polarization components of the reflected light are predicted below. The prediction process is as follows: First, based on the incident angle... and surface refractive index parameter Calculate the s-components and p-components of the Fresnel reflection and transmission coefficients; then, based on the diffuse reflectance parameter... The polarization Mueller matrix elements for calculating the diffuse reflection component using Fresnel transmission coefficients; based on the specular reflectivity parameter. Surface roughness parameter The polarization Mueller matrix elements of the specular reflection component are calculated using the Fresnel reflection coefficient. The Mueller matrix elements of the two components are then added to obtain the Mueller matrix elements of the total reflection. Finally, the outgoing Stokes vector is calculated based on the incident Stokes vector and the total reflection Mueller matrix under circularly polarized illumination conditions. The theoretical predictions of the first and second normalized polarization components are extracted from the outgoing Stokes vector. Here, the Mueller matrix is ​​a 4x4 matrix describing how an optical element or reflecting surface transforms the incident polarization state to the outgoing polarization state; each matrix element is a function of the incident angle and material parameters.

[0047] Before starting the iterative fitting, it is necessary to set physically feasible value ranges for the four parameters to be fitted as constraint boundaries to prevent the parameter values ​​from drifting to physically unreasonable regions during the iteration process. Diffuse reflectance parameter The physically feasible range for reflectivity is 0 to 1, because reflectivity cannot be negative or exceed 1. (Specular reflectivity parameter) The physically feasible value range is also 0 to 1. Surface roughness parameter The physical feasibility value ranges from 0.01 to 1.0, with the lower limit set at 0.01 instead of 0 because when As the refractive index approaches zero, the micro-element distribution degenerates into a delta function, leading to singularities in numerical calculations. Surface refractive index parameter The physically feasible values ​​range from 1.0 to 3.0, covering common industrial materials from polymeric materials (refractive indices of approximately 1.4 to 1.6) to metal oxide surfaces (equivalent refractive indices of 2.5 and above). In an alternative implementation, if the material type of the workpiece being tested is known, the constraint boundaries can be further narrowed; for example, for aluminum alloy workpieces, the surface refractive index parameter can be... The value range is limited to between 1.0 and 2.0.

[0048] The selection of initial values ​​for iteration directly affects the convergence speed and global optimization capability of the Levenberg-Marquardt algorithm. If the initial value is too far from the true value, the algorithm may converge to a local minimum rather than the global optimum. In this method, two strategies are used to determine the initial values ​​for iteration. The first strategy is suitable for cases where the current pixel has adjacent pixels that have already been fitted: the fitting result of the adjacent pixels is directly taken as the initial value of the current pixel. This strategy utilizes the spatial continuity of the surface of industrial workpieces: the surface states of adjacent pixels are usually highly similar, so the fitting result of the previous pixel is a good initial estimate of the current pixel. The specific pixel scanning order is row-by-row scanning, processing from left to right within each row, and taking the fitting result of the pixel to its left as the initial value for the current pixel. The first pixel of each row takes the fitting result of the pixel in the same column of the previous row as the initial value. The second strategy is suitable for cases where the fitting of the first pixel in the first row or adjacent pixels fails: the prior value corresponding to the material category is used as the initial value. The prior value is obtained by performing a complete fitting on a standard planar sample of the same material category in advance and stored as a lookup table for online detection. In one embodiment, the prior initial value for aluminum alloy is: Take 0.3, Take 0.2, Take 0.15, Take 1.5.

[0049] The fitting process employs the Levenberg-Marquardt iterative optimization algorithm with constrained boundaries. The first and second normalized polarization components predicted by the polarization bidirectional reflectance distribution function model under the incident angle conditions corresponding to the effective observations are used as the model output. The objective function is the sum of squared residuals between the model output and the effective observation sequence at all effective observation locations. Let the number of incident angle locations for effective observations be... ( This represents the total number of incident angle positions that have passed the validity check for the current pixel. No more than ), objective function Defined as the sum of the squares of the differences between the model predictions and observed values ​​of the first normalized polarization component and the second normalized polarization component at all valid observation locations, i.e. ,in For the first The incident angle corresponding to each effective observation and These represent the first and second normalized polarization components predicted by the model under the current parameter values, respectively. and These are the first and second normalized polarization component effective observations, respectively, obtained from the polarization image sequence.

[0050] The Levenberg-Marquardt algorithm calculates the Jacobian matrix of the objective function with respect to the four parameters to be fitted in each iteration. Based on the Jacobian matrix and the current damping factor, it determines the parameter update direction and step size. When the updated parameter value exceeds the constraint boundary, it is truncated to the boundary value. Iteration stops when the change between two consecutive values ​​of the objective function is less than a preset convergence threshold. In one implementation, the preset convergence threshold is set to 1 x 10⁻⁸, and the maximum number of iterations is set to 200. Most pixels converge within 20 to 50 iterations. The initial value of the damping factor is set to 1 x 10⁻³, and it is adaptively adjusted during iteration according to the direction of change of the objective function: the damping factor is decreased when the objective function decreases to accelerate convergence, and increased when the objective function increases to enhance stability.

[0051] After fitting, the optimal estimates of the four parameters to be fitted are obtained. These are arranged in the order of diffuse reflectance, specular reflectance, surface roughness, and surface refractive index to form the polarization reflection parameter vector of the current pixel. This vector is a 4-dimensional vector, and each component has a clear physical meaning. The above operation is performed on all pixels on the surface of the workpiece under test, arranging the polarization reflection parameter vectors of all pixels according to their spatial positions to form a polarization reflection parameter feature map. The spatial resolution of the polarization reflection parameter feature map is the same as the full-resolution grayscale image after de-mosaicing, with 4 channels. In one embodiment, the resolution after de-mosaicing is 1224 x 1024, so the size of the polarization reflection parameter feature map is 1224 x 1024 x 4. The polarization reflection parameter feature map can be visualized as four single-channel grayscale images, corresponding to the diffuse reflectance parameter distribution map, specular reflectance parameter distribution map, surface roughness parameter distribution map, and surface refractive index parameter distribution map, respectively. In normal areas, the spatial distribution of the four parameters is relatively uniform; in damaged areas, the parameter values ​​show local anomalies, such as the surface roughness parameter in the scratched area. The diffuse reflectance parameter of the corrosion pit area is significantly higher than that of the surrounding normal area. and surface refractive index parameter At the same time, it deviates from the normal value.

[0052] Pixel-by-pixel fitting is computationally intensive. In one implementation, a 1224x1024 resolution image contains approximately 1.25 million pixels. Each pixel requires an average of 30 iterations, with each iteration involving approximately 200 floating-point operations, resulting in a total computation of approximately 7.5 billion floating-point operations. The processing time on a single-threaded CPU is approximately 120 seconds. GPU parallelization or multi-threaded CPU parallelization can reduce this time to less than 5 seconds. GPU parallelization maps the fitting task for all pixels to threads on the GPU, with each thread independently handling the fitting process for one pixel, leveraging the GPU's massive parallelism to process thousands of pixels simultaneously. In another alternative implementation, a lookup table acceleration method can be used: a dense grid is pre-built within the value range of four parameters. The model output is pre-calculated for each grid node and stored as a lookup table. During fitting, interpolation using the lookup table replaces the forward calculation of the model, reducing the single-pixel fitting time by approximately an order of magnitude.

[0053] Optimal transfer domain adaptive damage segmentation is the final step in the entire method. Its core function is to eliminate the polarization reflection parameter distribution shift caused by batch differences, surface treatment process fluctuations, and illumination condition drift between the workpiece under test and the training samples. In industrial online inspection scenarios, workpieces of the same material type exhibit minute but systematic differences in surface microstructure under different production batches and processing conditions. These differences are reflected in the polarization reflection parameter space as overall distribution translation, stretching, or distortion. If the polarization reflection parameter feature map is directly input into the pre-trained segmentation network, the distribution shift will cause the network to misclassify normal areas or miss damaged areas. The optimal transfer method finds an optimal mapping that maps the parameter distribution of the workpiece under test to the parameter distribution of the reference domain. This eliminates the global distribution shift while maintaining the relative structure within the parameter space, enabling the pre-trained segmentation network to work stably on different batches of workpieces without retraining.

[0054] Optimal transmission domain adaptive damage segmentation first requires establishing two types of preliminary datasets. The first type is the domain alignment benchmark sample set: Polarization image sequences of undamaged standard samples of the same material as the workpiece under test are pre-collected. These images are processed according to the continuous variable-angle polarization image acquisition process in step 1 and the joint modeling process of the polarization bidirectional reflection distribution function in step 2, obtaining the polarization reflection parameter feature map of the standard sample. The polarization reflection parameter vectors of all pixels in the polarization reflection parameter feature map of the standard sample constitute the domain alignment benchmark sample set. The selection criteria for undamaged standard samples are: using the same material and surface treatment process as the workpiece under test, and confirming no visible damage on the surface through manual visual inspection and high-magnification microscopy. All polarization reflection parameter vectors in the domain alignment benchmark sample set represent the polarization reflection characteristic distribution of a normal surface under ideal conditions, serving as the target distribution for subsequent optimal transmission. In one embodiment, the size of the undamaged standard sample is the same as that of the workpiece under test, the resolution of the acquired standard sample polarization reflection parameter feature map is 1224 x 1024, and the domain alignment benchmark sample set contains approximately 1.25 million 4D polarization reflection parameter vectors. In another alternative implementation, polarization reflection parameter feature maps can be collected from multiple non-destructive standard samples and merged to form a domain-aligned reference sample set, so as to more fully cover the variation range of normal surface parameter distribution.

[0055] The second type of preliminary data is the training sample set for the segmentation network. Additionally, polarization image sequences of training samples of the same material type, containing damaged and undamaged regions and with pixel-level damage annotations, are pre-collected. These are processed according to steps 1 and 2 to obtain the polarization-reflectance parameter feature map of the training samples. Experienced inspectors use annotation tools to label the damaged regions on the polarization-reflectance parameter feature map pixel by pixel as either damaged or normal, forming a pixel-level damage annotation mask. A fully convolutional damage segmentation network is pre-trained on the polarization-reflectance parameter feature map of the training samples and the corresponding pixel-level damage annotations. The input to the fully convolutional damage segmentation network is a 4-channel polarization-reflectance parameter feature map, and the output is a damage category label for each pixel. The network structure can adopt an encoder-decoder architecture, where the encoder extracts multi-scale features, and the decoder progressively upsamples to restore spatial resolution. In one implementation, the encoder uses a ResNet-34 backbone network, modifying the original 3-channel input to a 4-channel input to adapt to the polarization-reflectance parameter feature map; the decoder uses a feature pyramid structure, with the output resolution the same as the input. During training, a weighted combination of cross-entropy loss and Dice loss is used as the optimization objective to address the class imbalance problem where the proportion of damaged regions is usually much smaller than that of normal regions. The number of training samples needs to ensure the diversity of the types of damage to be detected. In one implementation, 20 training samples are used, each containing artificially created damage of different types and severity. In another optional implementation, the network structure can also use U-Net or SegFormer, without affecting the overall method flow.

[0056] It is important to note that the domain alignment benchmark sample set and the segmentation network training sample set are two independent datasets serving different functions. The domain alignment benchmark sample set consists only of undamaged standard samples, and its function is to provide the optimal target distribution for transmission; it does not participate in the training of the segmentation network. The segmentation network training sample set contains labeled data for both damaged and undamaged regions, and its function is to train the segmentation network to learn the discrimination boundary between damaged and normal surfaces. The reason for separating them is that the domain alignment benchmark requires a "pure" distribution, that is, it only contains the parameter distribution of normal surfaces, so as to establish an unbiased target distribution; while the training samples must contain damaged data for the network to learn to recognize damage.

[0057] In the online inspection phase, the polarization reflection parameter vectors of all pixels in the polarization reflection parameter feature map of the workpiece obtained in step 2 are used to construct the sample set. Next, a discrete distribution representation is constructed in the 4-dimensional space containing the polarization reflection parameter vectors. The four coordinate axes—diffuse reflectance, specular reflectance, surface roughness, and surface refractive index—are uniformly divided into M parts. Each coordinate axis is further divided into M intervals by M equally spaced dividing points. The intervals of the four coordinate axes are combined pairwise to form M to the power of 4 grid cells. The selection of the value of M requires a balance between resolution and computational cost: if M is too small, the distribution representation is coarse, with large differences within the grid and low mapping accuracy; if M is too large, the number of grid cells increases exponentially, leading to a sharp increase in the storage and computational overhead of the cost matrix and kernel matrix. In one implementation, M is set to 20, resulting in a total of 20 to the power of 4, or 160,000, and a cost matrix size of 160,000 x 160,000. Storing this in single-precision floating-point format requires approximately 95 gigabytes of memory, which is quite tight for a conventional server. Therefore, in practical implementations, the symmetry of the cost matrix can be utilized to store only the upper triangular portion, or a sparsity strategy can be employed to retain only the grid cells with non-zero occupancy. In another optional implementation, M is set to 10, resulting in a total of 10,000 grid cells. The cost matrix requires only approximately 400 megabytes, significantly improving computational efficiency and making it suitable for scenarios with high real-time requirements.

[0058] The discrete distribution vector of the reference domain is obtained by dividing the number of samples falling into each grid cell in the statistical domain-aligned reference sample set by the total number of samples in the reference sample set. Similarly, the discrete distribution vector of the test domain is obtained by dividing the number of samples falling into each grid cell in the test sample set by the total number of samples in the test sample set. Both the reference domain discrete distribution vector and the test domain discrete distribution vector are one-dimensional vectors with a length equal to the total number of grid cells. Each element represents the proportion of samples in the corresponding grid cell, and the sum of all elements is 1. The cost matrix is ​​constructed using the Euclidean distance between the centers of any two grid cells. The cost matrix is... Line 1 Column element equals the first The center coordinates of the first grid cell and the first The Euclidean distance between the center coordinates of each grid cell is given by the 4D coordinates formed by the midpoint values ​​of the grid cell across the four coordinate axes. The cost matrix is ​​a symmetric matrix with diagonal elements of 0.

[0059] The Sinkhorn iterative algorithm is employed to solve the entropy-regularized optimal transport problem. The purpose of introducing entropy regularization is to transform the original optimal transport problem from a linear programming problem into a rigorously convex optimization problem that can be solved iteratively with high efficiency, while also making the obtained transport scheme smoother and avoiding the excessive sensitivity of sparse transport schemes to discretization errors. This is based on the cost matrix and preset entropy regularization coefficients. Construct the kernel matrix, where A positive real number that controls the strength of entropy regularization. The nth kernel in the kernel matrix. Line 1 The element of the column is equal to the negative value of the corresponding element in the cost matrix divided by . Then take the value of the natural index. The smaller the value, the closer the optimal transmission scheme is to the exact solution without regularization, but the slower the iterative convergence and the worse the numerical stability. The larger the value, the smoother the transmission scheme but the further it deviates from the exact solution. In one implementation, The value is set to 0.05 times the median of all non-zero elements in the cost matrix, which corresponds to approximately 0.01 when M is 10. In another alternative implementation, the optimal value can be adaptively selected by comparing the segmentation accuracy after domain alignment across several candidate values. value.

[0060] The specific process of Sinkhorn iteration is as follows: Initialize the baseline domain scaling vector. and the scaling vector of the domain to be measured All elements are 1, where Length and The length of each value is equal to the total number of grid cells. In each iteration, the base domain scaling vector is first... Updated to element-wise division of the discrete distribution vector of the reference domain by the kernel matrix and the scaling vector of the domain to be measured. The product of the two domains is then scaled to the vector of the domain under test. Updated to element-wise division of the discrete distribution vector of the test domain by the kernel matrix transpose and the scaling vector of the reference domain. The product of two vectors. Here, element-wise division refers to dividing the corresponding elements of the two vectors; the product of a matrix and a vector refers to standard matrix multiplication. Iterate until the reference domain is scaled between two adjacent iterations. and the scaling vector of the domain to be measured The iteration stops when the maximum change in any element is less than a preset stopping threshold. In one implementation, the preset stopping threshold is set to 1 x 10^-6, and convergence typically occurs within 50 to 200 iterations. After convergence, the vector is scaled by the reference domain. Kernel matrix and scaling vector of the domain under test Constructing the optimal transfer matrix ,in The Line 1 Column element equals The The nth element multiplied by the kernel matrix Line 1 Column elements multiplied by The Element. Optimal transfer matrix. The row index corresponds to the grid cell number of the domain to be tested, and the column index corresponds to the grid cell number of the reference domain. Each element represents the probabilistic mass transmitted from the corresponding test domain grid cell to the corresponding reference domain grid cell.

[0061] After constructing the optimal transfer matrix, barycentric mapping is performed on each polarization reflection parameter vector in the test sample set, mapping it from the test domain space to the reference domain space. The specific process of barycentric mapping is as follows: First, the grid cell number of the test domain to which the current polarization reflection parameter vector belongs is determined, that is, determining which interval the 4-dimensional vector falls into on each coordinate axis, thus locating a unique grid cell. Then, from the optimal transfer matrix... Extract all elements from the row corresponding to the given number, and divide the row's elements by the sum of its elements to obtain the transmission share vector. Theoretically, the sum of the row's elements equals the probability value of the corresponding grid cell in the discrete distribution vector of the test domain. Dividing by the sum of the row's elements normalizes the transmission share into probability weights, allowing subsequent weighted averaging calculations to be performed under normalization. Multiply each element in the transmission share vector by the center coordinates of the corresponding column index of the reference domain grid cell, and sum the results column by column to obtain the global mapping result. The physical meaning of the global mapping result is: the weighted average coordinates of the target location after the probability quality in the current test domain grid cell is optimally distributed to the reference domain grid cells.

[0062] refer to Figure 3 The figure contains six sub-figures, labeled (a) to (f). Sub-figures (a) to (d) respectively show the spatial distribution of the four parameters obtained through joint modeling using the polarization bidirectional reflectance distribution function on the surface of the workpiece under test. Sub-figure (a) shows the diffuse reflectance parameter. Distribution map, normal areas in the map The values ​​are concentrated around 0.35 and are relatively evenly distributed spatially, indicating the damaged area. The value rose to the range of 0.55 to 0.60, which is in stark contrast to the normal range. The physical reason for the increased value is that the proportion of volume-scattered emitted light increases after the surface microstructure of the damaged area is destroyed. Subfigure (b) shows the specular reflectivity parameter. Distribution map, normal area The values ​​are concentrated around 0.20, indicating the damaged area. The value decreases to the range of 0.08 to 0.12 because the regularity of the surface microstructure in the damaged area is disrupted, resulting in reduced specular reflection energy. Subfigure (c) shows the surface roughness parameter. Distribution map, normal area The value is approximately 0.12, and the surface is relatively smooth; the damaged area The value significantly increased to the range of 0.35 to 0.45, directly reflecting the increased surface micro-roughness caused by damage such as scratches and corrosion pits. Among the four parameters, The surface is most sensitive to surface damage, exhibiting the highest contrast between the damaged and normal areas. Subplot (d) shows the surface refractive index parameter. Distribution map, normal area The values ​​are concentrated around 1.50, indicating the damaged area. As the value increases to the range of 1.75 to 1.80, changes in the surface oxide layer or material exposure lead to alterations in the equivalent refractive index. Two typical damage morphologies can be identified in all four parametric distribution plots: linear scratches extending diagonally and circular corrosion pits distributed at different locations. Subplot (e) shows the diffuse reflectance parameters before and after domain alignment. The distribution comparison histogram. The horizontal axis is... The vertical axis represents the probability density. The figure shows that the measured domain exhibits an overall offset relative to the reference domain before alignment, with their distribution peak positions not coinciding. This domain offset originates from the systematic drift of polarization reflection parameters under different batches of workpieces or different lighting conditions. After entropy-regularized optimal transmission domain alignment, the distribution peak positions of the measured domain move closer to the reference domain, and their distribution patterns become consistent, indicating that domain alignment effectively eliminates the global distribution offset. Sub-figure (f) shows the pixel-level segmentation results of the domain-aligned polarization reflection parameter feature map by the fully convolutional damage segmentation network. The damaged and normal regions are clearly separated in the figure, and the shapes of linear scratches and circular corrosion pits are consistent with the positions and shapes of the parameter anomaly regions in sub-figures (a) to (d). A narrow transition band exists at the segmentation boundary, reflecting the physical characteristic of the polarization reflection parameter gradually transitioning from anomaly to normal values ​​at the damage edge. The figure labels the types of three typical damages, including obliquely extending scratches and two corrosion pits of different sizes, verifying that the segmentation network can simultaneously identify two different types of surface damage: linear and planar.

[0063] The global mapping result reflects the mapping relationship at the grid cell level. However, different polarization reflection parameter vectors falling within the same grid cell of the test domain have different positions within the cell. If all are mapped to the same global mapping result, the spatial differences within the cell will be lost. To preserve this difference, the offset of the current polarization reflection parameter vector relative to the center of its respective grid cell of the test domain is calculated, and the offset is superimposed on the global mapping result to obtain the mapped polarization reflection parameter vector. This process assumes that within the local range of a single grid cell, the optimal transport mapping is approximately a translation transformation, and the relative positional relationship of each sample within the cell remains unchanged before and after mapping. This approximation has a small error when the grid cell size is small (i.e., the M value is large), and is acceptable under engineering accuracy requirements.

[0064] The above-described barycentric mapping is performed on all polarization reflection parameter vectors. The mapped polarization reflection parameter vectors are then arranged according to their original spatial positions to form a domain-aligned polarization reflection parameter feature map. The domain-aligned polarization reflection parameter feature map has the same spatial resolution and number of channels as the original polarization reflection parameter feature map, but the polarization reflection parameter vector of each pixel has been mapped from the test domain space to the reference domain space. The domain alignment operation does not change the spatial arrangement relationship between pixels; it only adjusts the position of each pixel in the 4-dimensional parametric space.

[0065] A fully convolutional damage segmentation network is pre-trained using a domain-aligned polarization reflectance parameter feature map as input to a 4-channel image. The network outputs a damage category label for each pixel in the domain-aligned polarization reflectance parameter feature map, resulting in pixel-level damage segmentation with the same spatial resolution as the input. Since the domain alignment operation pulls the polarization reflectance parameter distribution of the workpiece under test back to a reference domain space similar to the training samples, the segmentation network can directly apply the discrimination boundaries learned during training to the domain-aligned test data, thus adapting to the detection tasks of different batches of workpieces without retraining or fine-tuning.

[0066] In one optional implementation, when the damage type includes multiple categories (such as scratches, corrosion pits, coating peeling, etc.), the output layer of the fully convolutional damage segmentation network is set to multi-class output, with the number of output channels equal to the total number of damage categories plus 1 (normal categories occupy 1 channel), and the category label of each pixel is taken from the category corresponding to the channel with the largest response. In another optional implementation, if only the difference between damage and normal needs to be distinguished, the output layer is set to single-channel binary output, with pixels having an output value greater than 0.5 marked as damage, and pixels with an output value less than or equal to 0.5 marked as normal.

[0067] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A surface damage intelligent segmentation method based on multi-angle polarization imaging, characterized in that, Includes the following steps: Step 1, Continuous variable angle polarization image acquisition: A rotating wedge prism beam deflection device is used, which consists of a first wedge prism and a second wedge prism cascaded along the same optical axis. By controlling the difference in rotational speed of the two wedge prisms, the incident angle of the illumination beam on the surface of the workpiece is continuously changed. At each incident angle position, a split-focus plane polarization camera is used to acquire reflected polarization images, forming a polarization image sequence containing N incident angle positions. Step 2, Joint modeling of polarization bidirectional reflectance distribution function: For each pixel in the polarization image sequence, calculate the normalized polarization component at each incident angle position, establish a polarization bidirectional reflectance distribution function model and fit it to obtain the polarization reflectance parameter vector, and construct the polarization reflectance parameter feature map from the polarization reflectance parameter vectors of all pixels. Step 3, Optimal Transmission Domain Adaptive Damage Segmentation: Using the polarization reflection parameter distribution of the undamaged standard sample as the domain alignment reference, the polarization reflection parameter feature map is mapped to the reference domain through the entropy regularization optimal transmission method to obtain the domain-aligned polarization reflection parameter feature map. The domain-aligned polarization reflection parameter feature map is input into the pre-trained fully convolutional damage segmentation network to output the pixel-level damage segmentation result of the surface of the workpiece to be tested.

2. The method according to claim 1, characterized in that, In step 1, the first wedge prism and the second wedge prism have the same wedge angle and are driven by independent servo motors to rotate around the optical axis. By setting the angular velocity of the first wedge prism to be greater than that of the second wedge prism, the deflection angle of the illumination beam changes continuously with the relative rotation angle between the two wedge prisms. Within the preset acquisition range, a scanning segment with a continuously expanding incident angle is selected as the effective acquisition segment. This allows the incident angle of the illumination beam on the surface of the workpiece to gradually change from the minimum incident angle to the preset maximum incident angle within the effective acquisition segment, and the scanning trajectory unfolds in a spiral.

3. The method according to claim 1 or 2, characterized in that, In step 1, a linear polarizer and an achromatic quarter-wave plate are fixedly installed sequentially along the beam propagation direction in the illumination optical path. The fast axis direction of the achromatic quarter-wave plate forms a 45-degree angle with the transmission direction of the linear polarizer, so that the illumination beam is converted into circularly polarized light.

4. The method according to claim 2, characterized in that, In step 1, the polarization sensor of the split-focus plane polarization camera integrates linear polarization microanalyzers in four directions (0°, 45°, 90°, and 135°) within each 2x2 superpixel unit. An angle trigger position is set on the servo motor encoder. Whenever the relative rotation angle between the first and second wedge prisms reaches a preset equally spaced angle position within the effective acquisition segment, a hardware trigger signal is sent to the split-focus plane polarization camera to acquire one frame of original polarization mosaic image.

5. The method according to claim 4, characterized in that, For each frame of the original polarized mosaic image, using 2x2 superpixel units as the basic unit, full-resolution grayscale images of the four polarization directions (0°, 45°, 90°, and 135°) are recovered through edge-guided interpolation. The full-resolution grayscale images of the four polarization directions acquired at all incident angle positions within the effective acquisition segment are arranged in ascending order of incident angle to form a polarized image sequence.

6. The method according to claim 5, characterized in that, In step 2, the process of solving the normalized polarization component is as follows: for each incident angle position, the sum of the gray value of the current pixel in the 0-degree direction and the gray value in the 90-degree direction is taken as the total light intensity component. When the total light intensity component is less than the preset intensity threshold, the observation at the current incident angle position is marked as an invalid observation and excluded from the subsequent objective function calculation; When the total light intensity component is greater than or equal to the preset intensity threshold, the difference between the gray value in the 0-degree direction and the gray value in the 90-degree direction is divided by the total light intensity component to obtain the first normalized polarization component, and the difference between the gray value in the 45-degree direction and the gray value in the 135-degree direction is divided by the total light intensity component to obtain the second normalized polarization component. Repeat the above calculation for all N incident angle positions to obtain the effective observation sequence of the first normalized polarization component and the effective observation sequence of the second normalized polarization component of the current pixel.

7. The method according to claim 6, characterized in that, In step 2, the polarization bidirectional reflection distribution function model represents the reflected light as the superposition of diffuse reflection and specular reflection components. The diffuse reflection component is determined by the diffuse reflectance parameter and the surface refractive index parameter, and its polarization characteristics originate from the Fresnel transmission effect when light passes through the medium interface twice. The specular reflection component is determined by the specular reflectance parameter, surface roughness parameter, and surface refractive index parameter. The normal direction of the micro-facet follows an isotropic distribution controlled by the surface roughness parameter, and the reflection on each micro-facet follows Fresnel's reflection law. The model contains four parameters to be fitted: diffuse reflectance parameter, specular reflectance parameter, surface roughness parameter, and surface refractive index parameter. Physically feasible value ranges are set for each of the four parameters to be fitted as constraint boundaries, with adjacent... The fitting result of the pixel or the prior value corresponding to the material category is used as the initial value for iteration. The first and second normalized polarization components predicted by the polarization bidirectional reflectance distribution function model under the incident angle condition corresponding to the effective observation are used as the model output. The sum of squared residuals between the model output and the effective observation sequence at all effective observation positions is used as the objective function. The Levenberg-Marquardt iterative optimization algorithm with constrained boundaries is used for optimization. The iteration stops when the change between two adjacent values ​​of the objective function value is less than the preset convergence threshold. The optimal estimates of the four parameters to be fitted are obtained. The polarization reflectance parameter vector is formed by arranging the parameters in the order of diffuse reflectance, specular reflectance, surface roughness, and surface refractive index.

8. The method according to claim 1, characterized in that, In step 3, a polarization image sequence of a non-damaged standard sample of the same material type as the workpiece to be tested is pre-acquired. The standard sample polarization reflection parameter feature map is obtained by processing it according to the process of step 1 and step 2. The polarization reflection parameter vectors of all pixels in the standard sample form a domain-aligned reference sample set. In addition, a polarization image sequence of a training sample of the same material type containing damaged and non-damaged regions and with pixel-level damage annotations is pre-acquired. The training sample polarization reflection parameter feature map is obtained by processing it according to the process of step 1 and step 2. A fully convolutional damage segmentation network is pre-trained on the polarization reflection parameter feature map of the training sample and the corresponding pixel-level damage annotations.

9. The method according to claim 8, characterized in that, In step 3, the specific process of the entropy regularization optimal transmission method is as follows: The polarization reflection parameter vectors of all pixels in the polarization reflection parameter feature map are used to form a test sample set; in the 4-dimensional space where the polarization reflection parameter vectors are located, the four coordinate axes of diffuse reflectance, specular reflectance, surface roughness, and surface refractive index are uniformly divided into M parts, forming M to the power of 4 grid cells; the number of samples falling into each grid cell in the domain-aligned reference sample set is counted, and divided by the total number of samples in the domain-aligned reference sample set to obtain the reference domain discrete distribution vector; the number of samples falling into each grid cell in the test sample set is counted, and divided by the total number of samples in the test sample set to obtain the test domain discrete distribution vector; the cost matrix is ​​constructed using the Euclidean distance between the centers of any two grid cells; a kernel matrix is ​​constructed based on the cost matrix and a preset entropy regularization coefficient, where each element in the kernel matrix is ​​equal to the negative value of the corresponding element in the cost matrix divided by the entropy regularization coefficient and then taking the natural exponent value. Initialize all elements of the reference domain scaling vector and the test domain scaling vector to 1; In each iteration, the reference domain scaling vector is first updated to the reference domain discrete distribution vector by dividing the kernel matrix and the scaling vector of the domain to be tested element by element, and then the scaling vector of the domain to be tested is updated to the domain to be tested discrete distribution vector by dividing the kernel matrix transpose and the scaling vector of the reference domain element by element. The iteration stops when the maximum element change of both the reference domain scaling vector and the test domain scaling vector between two consecutive iterations is less than a preset stopping threshold. The optimal transfer matrix is ​​constructed from the reference domain scaling vector, the kernel matrix, and the test domain scaling vector. The row index of the optimal transfer matrix corresponds to the grid cell number of the test domain, and the column index corresponds to the grid cell number of the reference domain.

10. The method according to claim 9, characterized in that, In step 3, the process of performing centroid mapping for each polarization reflection parameter vector in the test sample set is as follows: determine the grid cell number of the test domain to which the current polarization reflection parameter vector belongs, extract all elements of the row corresponding to that number from the optimal transfer matrix, divide the elements of that row by the sum of the elements of that row to obtain the transfer share vector, multiply each element in the transfer share vector by the coordinates of the reference domain grid cell center of the corresponding column index, and sum them column by column to obtain the global mapping result; then calculate the offset of the current polarization reflection parameter vector relative to the center of the grid cell of the test domain to which it belongs, and superimpose the offset onto the global mapping result to obtain the mapped polarization reflection parameter vector; perform the above centroid mapping on all polarization reflection parameter vectors, and arrange the mapped polarization reflection parameter vectors according to their original spatial positions to form a domain-aligned polarization reflection parameter feature map.

Images