Quality management system for industrial products based on visual inspection

Abstract

This invention relates to the fields of machine vision and industrial quality inspection, specifically to an industrial product quality management system based on vision inspection, comprising: a light field modulation and acquisition unit, which controls a programmable fully polarized light source to illuminate a target object and generates a four-dimensional polarization raw data stream through a focal plane beam splitting micro-polarization array camera; a polarization space mapping unit, which converts the data stream into a polarization degree feature matrix and a polarization angle feature matrix; a polarization entropy gradient evaluation unit, which calculates the polarization angle Shannon entropy within a sliding window, generates a polarization high-entropy heatmap, and extracts polarization topological singularities by differential extraction with a baseline polarization manifold model; and a physical closed-loop and adaptive update unit, which outputs workpiece coordinates and shape classification instructions, performs physical control feedback, and extracts local Stokes vector features to update the baseline polarization manifold model.

Description

Technical Field

[0001] This invention relates to the field of machine vision and industrial quality inspection, specifically to an industrial product quality management system based on vision inspection. Background Technology

[0002] As an online quality inspection device, the industrial product quality management system based on vision inspection plays a crucial role in precision manufacturing, aerospace component processing, and final inspection of metal surfaces. Its detection accuracy directly affects the reliability of defect identification results and subsequent rejection control. Therefore, the key to ensuring the effectiveness of quality management is to stably acquire image information that can characterize the true state of the workpiece surface.

[0003] Existing detection methods have many problems. For example, they often use ordinary brightness imaging or color imaging to inspect the surface of highly reflective metal workpieces. Due to factors such as specular reflection, complex curvature changes and minute damage scale of the tested surface, it is easy to cause high light saturation, local detail obscuring, and confusion between normal surface texture and real defects. Especially in online inspection scenarios where workpieces are continuously transported, problems such as positional offset, inaccurate registration, and judgment lag can easily occur between different images, making it difficult to balance defect recognition accuracy, real-time performance, and production line closed-loop control requirements. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial product quality management system based on vision inspection, and to solve the following technical problems:

[0005] It avoids the problems of high light saturation and defect overwhelming that are prone to occur in conventional brightness imaging when inspecting highly reflective precision metal parts and complex curved surfaces. It is also easier to stably identify real micro-defects and distinguish between surface attachments and structural damage, thereby realizing an online quality inspection closed loop that directly drives production line control and adaptive model updates.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] The vision-based industrial product quality management system includes: a light field modulation and acquisition unit: controlling a programmable fully polarized light source to modulate the emission polarization state of the target object; capturing the light waves reflected by the target object through a focal plane beam-splitting micro-polarization array camera to generate a four-dimensional polarization raw data stream;

[0008] Polarization space mapping unit: In response to the four-dimensional polarization raw data stream, the computational processing unit configured by the polarization space mapping unit executes a Stokes decoding matrix to linearly map the four-channel light intensity to three Stokes parameters, and converts the four-dimensional polarization raw data stream into a polarization degree feature matrix and a polarization angle feature matrix.

[0009] Polarization entropy gradient evaluation unit: Set a sliding window on the polarization angle feature matrix, calculate the polarization angle Shannon entropy within the sliding window, and generate a polarization high-entropy heat map; perform difference analysis between the polarization high-entropy heat map and the preset baseline polarization manifold model to extract polarization topological singularities;

[0010] Physical closed-loop and adaptive update unit: Based on the polarization topological singularity and the pre-acquired spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, it generates two-dimensional physical coordinates and shape classification instructions for the workpiece coordinate system; outputs physical control feedback based on the shape classification instructions; extracts the local Stokes vector features corresponding to the polarization topological singularity, and updates the baseline polarization manifold model.

[0011] Furthermore, the method of controlling a programmable fully polarized light source to modulate the emission polarization state of the target object; and capturing the light waves reflected from the target object using a focal plane beam-splitting micro-polarization array camera to generate a four-dimensional polarization raw data stream includes:

[0012] The programmable fully polarized light source is activated to project a probe light field with a preset polarization state onto the moving target object;

[0013] The focal plane array camera is triggered to perform a single exposure; the light field data reflected from the surface of the target object is acquired.

[0014] Extract the zero-degree channel data, forty-five-degree channel data, ninety-degree channel data and one hundred and thirty-five-degree channel data that are integrated at the pixel level inside the focal plane beam-splitting micro-polarization array camera;

[0015] The zero-degree channel data, the forty-five-degree channel data, the ninety-degree channel data, and the one hundred and thirty-five-degree channel data are combined to generate the four-dimensional polarization raw data stream.

[0016] Furthermore, the computational processing unit includes a graphics processing unit tensor core, and the method of converting the four-dimensional polarization raw data stream into a polarization degree feature matrix and a polarization angle feature matrix by executing a Stokes decoding matrix through the computational processing unit in response to the four-dimensional polarization raw data stream includes:

[0017] The four-dimensional polarization raw data stream is input into the video memory of the graphics processor tensor core;

[0018] The total light intensity parameters and linear polarization component parameters corresponding to the four-dimensional polarization raw data stream are calculated in parallel.

[0019] Based on the total light intensity parameter and the linear polarization component parameter, the polarization state of the target object is reconstructed;

[0020] The reconstructed polarization states are mapped to a polarization degree feature matrix representing the degree of linear polarization and a polarization angle feature matrix representing the polarization angle, respectively.

[0021] Furthermore, the method of setting a sliding window on the polarization angle feature matrix and calculating the polarization angle Shannon entropy within the sliding window to generate a polarization high entropy heatmap includes: taking each pixel in the polarization angle feature matrix as the center pixel in turn;

[0022] Using the center pixel as a reference, establish a sliding window of a preset size; statistically analyze the probability distribution of polarization angle values ​​within the sliding window;

[0023] The local spatial information entropy is calculated based on the probability distribution and used as the polarization angle Shannon entropy.

[0024] Traverse all pixels in the polarization angle feature matrix, record the corresponding polarization angle Shannon entropy, and combine them to generate the polarization high entropy heatmap.

[0025] Furthermore, the method of extracting polarization topological singularities by performing differential analysis between the polarization high-entropy heatmap and the preset baseline polarization manifold model includes: loading the baseline polarization manifold model corresponding to the standard good product;

[0026] The local polarization angle Shannon entropy value in the polarization high-entropy heatmap is compared with the preset low-entropy threshold in the baseline polarization manifold model.

[0027] If the local polarization angle Shannon entropy value is higher than the preset low entropy threshold, it is determined that the corresponding region has undergone depolarization and multiple scattering, and the corresponding region is marked as the polarization topological singularity; if the local polarization angle Shannon entropy value is lower than or equal to the preset low entropy threshold, it is determined that the corresponding region is a continuous smooth surface and is retained.

[0028] Furthermore, after extracting the polarization topological singularity, the method further includes:

[0029] The mean polarization degree of the region corresponding to the polarization topological singularity in the polarization degree feature matrix is ​​extracted as the first feature value;

[0030] The spatial rate of change of the region corresponding to the polarization topological singularity in the polarization angle feature matrix is ​​calculated as the second feature gradient;

[0031] Obtain the preset depolarization threshold and the preset polarization angle abrupt change threshold;

[0032] If the first feature value is lower than the preset depolarization threshold, and the second feature gradient is lower than or equal to the preset polarization angle abrupt change threshold, then the polarization topological singularity is determined to be a surface attachment.

[0033] If the second feature gradient is higher than the preset polarization angle abrupt change threshold, and the first feature value is higher than or equal to the preset depolarization threshold, then the polarization topological singularity is determined to be a real structural damage; if the above conditions are not met, then the polarization topological singularity is determined to be an undefined anomaly.

[0034] Furthermore, based on the polarization topological singularity and the pre-acquired spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, two-dimensional physical coordinates and morphological classification instructions for the workpiece coordinate system are generated.

[0035] The method of outputting physical control feedback based on the morphological classification instruction includes: extracting the two-dimensional pixel coordinates of the polarization topological singularity that is determined to be the real structural damage;

[0036] By combining the spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, the two-dimensional pixel coordinates are mapped to the two-dimensional physical coordinates of the target object in the workpiece coordinate system;

[0037] The maximum value of the polarization angle Shannon entropy in the region corresponding to the polarization topological singularity that is determined to be the real structural damage is extracted as the peak value of the polarization angle Shannon entropy.

[0038] Based on the peak value of the Shannon entropy at the polarization angle, the morphology classification instruction is generated; based on the two-dimensional physical coordinates of the workpiece coordinate system and the morphology classification instruction, a rejection signal for controlling the external pneumatic rejection device is generated and output.

[0039] Furthermore, the method for extracting the local Stokes vector features corresponding to the polarization topological singularity and updating the baseline polarization manifold model includes:

[0040] Extract optical response data from the region corresponding to the polarization topological singularity identified as the actual structural damage in the four-dimensional polarization raw data stream; calculate local Stokes vector features based on the optical response data;

[0041] The local Stokes vector features are output to an external edge server, which then adjusts the preset low-entropy threshold in the baseline polarization manifold model based on the local Stokes vector features, thereby completing the dynamic update of the baseline polarization manifold model.

[0042] The beneficial effects of this invention are:

[0043] 1. This invention uses a fully polarized light source and an array camera to acquire polarization data, which is then converted into a polarization degree and polarization angle matrix. By calculating the Shannon entropy of the local polarization angle and performing differential analysis with the baseline model, the saturation interference of highly reflective mirrors is effectively eliminated, and the normal physical gradient of the curved surface can be distinguished from the depolarization and disordered scattering caused by minor damage.

[0044] 2. This invention uses a single short exposure of a focal plane array micro-polarization array camera to instantly and synchronously capture data from four channels: 0 degrees, 45 degrees, and 90 degrees. This mechanism overcomes the positional misalignment and registration problems caused by time-division shooting in continuous conveyor mode, ensuring the absolute spatiotemporal consistency of polarization data of high-speed moving workpieces.

[0045] 3. This invention relies on the tensor core of the graphics processor to perform Stokes decoding in parallel, reducing the cumulative latency of the solution; at the same time, the system directly maps the extracted abnormal singularities to the physical coordinates of the workpiece, and generates hierarchical execution instructions based on the polarization entropy peak, realizing millisecond-level automated closed-loop control from visual detection to external pneumatic rejection.

[0046] 4. The system of this invention combines the characteristics of polarization degree and polarization angle abrupt change, which can accurately distinguish between cleanable attachments and real structural damage, and prevent excessive scrapping; it extracts the underlying Stokes features of real damage to the edge server, and only makes boundary fine-tuning on the baseline normal threshold, so that the system can automatically adapt to the drift of material reflective properties and ensure the stability of long-term online detection. Attached Figure Description

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] Figure 1 This is a schematic diagram of the modules of the vision-based industrial product quality management system provided in the embodiments of this application. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see Figure 1 The vision-based industrial product quality management system includes: a light field modulation and acquisition unit: controlling a programmable fully polarized light source to modulate the emission polarization state of the target object; capturing the light waves reflected by the target object through a focal plane beam splitting micro-polarization array camera to generate a four-dimensional polarization raw data stream;

[0051] Polarization Space Mapping Unit: In response to the four-dimensional polarization raw data stream, the computational processing unit configured by the polarization space mapping unit executes the Stokes decoding matrix to linearly map the four-channel light intensity to three Stokes parameters, and converts the four-dimensional polarization raw data stream into a polarization degree feature matrix and a polarization angle feature matrix.

[0052] Polarization entropy gradient evaluation unit: Set a sliding window on the polarization angle feature matrix, calculate the polarization angle Shannon entropy within the sliding window, and generate a polarization high-entropy heatmap; perform difference analysis between the polarization high-entropy heatmap and the preset baseline polarization manifold model to extract polarization topological singularities;

[0053] Physical closed-loop and adaptive update unit: Based on the polarization topological singularity and the pre-acquired spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, it generates two-dimensional physical coordinates and shape classification instructions for the workpiece coordinate system; outputs physical control feedback based on the shape classification instructions; extracts the local Stokes vector features corresponding to the polarization topological singularity and updates the baseline polarization manifold model.

[0054] This embodiment provides a quality management mechanism for online inspection of highly reflective precision metal parts. Specifically, the system is deployed at the final inspection station after the precision machining of aero-engine turbine blades. The blades are continuously fed into the inspection field of view by a conveying fixture. The surface to be tested has a mirror area, a complex curvature transition area, and a micron-level scratch-sensitive area. Conventional brightness imaging is prone to high light saturation and defect overwhelming in these areas. Therefore, this embodiment directly establishes the inspection link around polarization information.

[0055] Specifically, when the blade enters the shooting position, the light field modulation and acquisition unit drives the programmable fully polarized light source to output a detection light field with a set polarization state;

[0056] The programmable fully polarized light source includes an array of light-emitting elements and a polarization modulation component consisting of an electronically controlled liquid crystal variable phase retarder and a solid-state polarizer. The controller can dynamically switch the polarization state of the output probe light field by adjusting the driving voltage applied to the liquid crystal variable phase retarder.

[0057] After the light field of the probe illuminates the surface of the blade, the reflected light is captured by a focal plane array micro-polarization array camera in a single exposure. Since the camera's photosensitive surface integrates an array of micro-polarizers with different orientations, pixel data corresponding to four polarization directions can be obtained simultaneously within a single frame.

[0058] For ease of understanding, the neighborhood of the same pixel can be considered as consisting of four sub-pixels, denoted as the 0-degree, 45-degree, 90-degree, and 135-degree channels, respectively. If the four-channel sample values ​​of a certain local area are 120, 80, 40, and 70, then this set of values ​​does not only represent brightness, but also the response difference of that location in different polarization directions. The system aligns the four-channel responses of all locations in the entire image according to their spatial positions to form a four-dimensional polarization raw data stream.

[0059] The polarization space mapping unit sends the data stream to the computation processing unit for Stokes decoding; for illustrative purposes, assume that the sampling intensities of a pixel location on the polarization channels at 0 degrees, 45 degrees, 90 degrees, and 135 degrees are respectively... , , and Then the total light intensity parameter and the linear polarization component parameter can be obtained first;

[0060] The total light intensity reflects the overall light level at that location, while the two linearly polarized components reflect the polarization differences in different directions. Based on these quantities, the polarization state at that location can be reconstructed and further mapped to the degree of polarization and the polarization angle.

[0061] Specifically, the polarization state extracted in the above solution process is represented by a simplified local Stokes vector. Since the array camera only collects linearly polarized light in four directions, this vector does not contain circular polarization components. It consists only of a first Stokes parameter representing the total light intensity, a second Stokes parameter representing the difference between horizontal and vertical linear polarization, and a third Stokes parameter representing the difference between diagonal polarization. Correspondingly, the Stokes decoding matrix is ​​the transformation matrix that linearly maps the four-channel pixel response values ​​to the above three Stokes parameters.

[0062] If another pixel is close to saturation in total light intensity, but there are still significant differences in the response in different directions, its polarization degree and polarization angle can still retain usable information. This is the basis for this embodiment to continue working in the strong reflective area.

[0063] After obtaining the polarization degree feature matrix and the polarization angle feature matrix, the polarization entropy gradient evaluation unit no longer performs threshold segmentation directly on the brightness map, but instead calculates the local spatial information entropy on the polarization angle feature matrix; the polarization angle feature matrix can be understood as an angle map that records the change of polarization direction at each position;

[0064] If the surface is continuous and smooth, the change in polarization angle between adjacent pixels is usually gradual; if there are scratches, chipped edges or micro-pits, multiple scattering and depolarization will occur locally, resulting in disordered polarization angle distribution.

[0065] For ease of explanation, let's assume a certain The polarization angle values ​​within the window are mostly concentrated in Nearby, for example , , , If the probability distribution of the window is concentrated, the local entropy is low; if another The angles within the window are discretely distributed. , , , If there are multiple intervals, the probability distribution of the window will be discrete and the local entropy will be high; after the system traverses the entire polarization angle feature matrix, a high-entropy polarization heatmap will be formed.

[0066] The system performs differential analysis between the real-time polarization high-entropy thermal map and the pre-established baseline polarization manifold model. This baseline model can be obtained from standard good blades under the same work station, the same polarization illumination, and the same shooting parameters. Essentially, it records the range of polarization entropy distribution that is allowed to occur in each region of the normal curved surface.

[0067] If the real-time local entropy of a certain region is greater than the preset low entropy threshold of the corresponding position in the baseline model, then the region is extracted as a polarization topological singularity; here, a singularity refers to an actual physical region where abrupt changes occur in normal polarization continuity.

[0068] After extracting the singularity regions, the physical closed-loop and adaptive update units perform coordinate generation, shape determination and feedback control on these regions; the system maps the position of the singularity regions in the image to the workpiece coordinate system to form two-dimensional physical coordinates for execution by production line equipment; at the same time, the system outputs shape classification instructions based on the polarization response shape of the singularity regions, such as scratches, dents, attachments or undefined anomalies.

[0069] The external actuator completes the rejection, marking, or flow rejection according to the instruction; at the same time, the system also extracts the local Stokes vector features corresponding to the abnormal area and uses them as input for subsequent updating of the baseline polarization manifold model to adapt to the temporal feature drift of the batch material reflection characteristics.

[0070] Furthermore, to ensure consistency in terminology throughout the text, the above-mentioned morphological classification instructions adopt a hierarchical expression in this embodiment: the first layer is the defect attribute category, used to distinguish between surface attachments, real structural damage and undefined anomalies; the second layer is the damage subdivision label under the premise that it has been determined to be real structural damage, such as scratches, pits, chipped edges or depth damage levels.

[0071] Therefore, the terms "scratches," "dents," "attachments," or "undefined anomalies" mentioned above are combined descriptions for ease of understanding. Among them, attachments and undefined anomalies correspond to the first-level categories, while scratches, dents, etc., correspond to the second-level subdivision labels under actual structural damage. During online control, the system first determines whether to reject, clean and rework, or manually review based on the first-level categories. Then, when necessary, it combines the second-level subdivision labels to generate more specific process handling instructions, thus maintaining consistency with the classification logic in the subsequent subordinate implementation methods.

[0072] Furthermore, the updated baseline polarization manifold model here does not directly incorporate the actual structural damage area into the good product baseline, nor does it treat the abnormal area as a new normal template. Instead, it uses the local Stokes vector feature as one of the correction reference samples for the edge server to identify which polarization changes at the current workstation belong to stable process drift and which polarization changes belong to actual damage that should be continuously eliminated.

[0073] Specifically, the local Stokes vector features of the real structural damage area mainly serve to characterize the abnormal boundary and constrain the threshold adjustment. When updating, the edge server only performs partition fine-tuning on the allowable entropy range corresponding to the good product area and keeps the abnormal labels of the confirmed damage area from being absorbed into the baseline normal distribution, thereby avoiding the baseline model from being contaminated by defect samples.

[0074] Furthermore, under abnormal or boundary conditions, when the camera detects that the overall light intensity in a frame is lower than the minimum exposure threshold, it can be determined that the frame cannot form an effective polarization reconstruction. The system marks the frame as an invalid frame and requests the next frame to be sampled.

[0075] When a local area has missing channels, partial occlusion, or dead pixels, the system can interpolate based on adjacent valid pixels or directly mask the area to avoid misjudging equipment noise as a real defect. When the area of ​​the extracted high-entropy region exceeds the preset single defect area threshold, the system can further determine it as oil stains, liquid film, or large-area contamination and enter the manual review path instead of directly outputting a damage conclusion.

[0076] In the final inspection line of turbine blades, a blade numbered B017 enters the inspection station at a constant speed. Its leading edge polished area appears as a large area of ​​high light under a regular industrial camera. After adopting this embodiment, although the total light intensity in this area is relatively high, a stable polarization angle characteristic can still be obtained after polarization mapping.

[0077] At a position one-third of the chord length near the blade tip, the system detects a high-entropy peak region of approximately 0.18 mm, which forms a single polarization topological singularity after being differentiated from the baseline model. This singularity is mapped to a specified position in the blade workpiece coordinate system, outputting a morphological classification instruction corresponding to the actual structural damage. When the blade reaches the rejection position, the rejection device performs rejection and uploads the local Stokes vector features of the abnormal region to the edge server for subsequent baseline adjustment.

[0078] The purpose of this step is to transform the traditional defect identification that relies on brightness imaging into a detection and control closed loop that relies on polarization physical response, thereby achieving stable identification of micro-defects on highly reflective complex curved surfaces and enabling the detection results to directly drive production line quality management actions.

[0079] In a preferred embodiment of the present invention, the method of controlling a programmable fully polarized light source to modulate the emission polarization state of a target object; and capturing the light wave reflected by the target object through a focal plane beam-splitting micro-polarization array camera to generate a four-dimensional polarization raw data stream includes: activating the programmable fully polarized light source, projecting a probe light field with a preset polarization state onto the moving target object; and triggering the focal plane beam-splitting micro-polarization array camera to perform a single exposure.

[0080] Acquire light field data reflected from the surface of the target object; extract the zero-degree channel data, forty-five-degree channel data, ninety-degree channel data, and one hundred and thirty-five-degree channel data integrated at the pixel level inside the focal plane beam-splitting micro-polarization array camera; combine the zero-degree channel data, forty-five-degree channel data, ninety-degree channel data, and one hundred and thirty-five-degree channel data to generate a four-dimensional polarization raw data stream.

[0081] This embodiment provides a mechanism for acquiring raw polarization data in a single exposure. Specifically, in the above-mentioned final inspection scenario of the blade, if the method of switching polarizers through multiple time-division exposures is still used, although data in different polarization directions can theoretically be obtained, the blade is in a moving state and there will be positional offset between different frames, which will cause the data in the four directions to not strictly correspond to the same surface point, thus compromising the accuracy of subsequent polarization calculation.

[0082] Therefore, this embodiment introduces a focal plane beam splitting micro-polarization array camera with pixel-level integrated micro-polarization array, which can simultaneously acquire data in four polarization directions in a single exposure;

[0083] Specifically, the system first activates the programmable fully polarized light source according to the production line cycle and sets its emission polarization state to a preset value. This preset value can be selected according to the material to be tested. For example, linearly polarized probe light can be used for polished nickel-based alloy blades to enhance the polarization difference in the mirror reflection area.

[0084] When the conveyor encoder detects that the blade is in place, the controller sends a hard trigger signal to the camera, enabling the camera to complete a single image within a preset single trigger exposure time period. Since each micro-unit on the focal plane corresponds to a specific polarization direction, data from the zero-degree channel, forty-five-degree channel, ninety-degree channel, and one hundred and thirty-five-degree channel can be extracted after sampling.

[0085] To make it easier to understand, a certain camera setting can be configured. The four sub-pixels within a superpixel block correspond to the four polarization channels respectively; if the sampled values ​​of the superpixel block are 110, 78, 52, and 73 respectively, then this set of values ​​can be regarded as the instantaneous response of the same surface micro-region in the four polarization directions.

[0086] The system expands all superpixel blocks in the entire image using spatial alignment, forming four channel maps of the same size, and then combines them one by one according to their positions to form a four-dimensional polarization raw data stream; here, four-dimensional can be understood as a joint data organization form of spatial horizontal coordinate, spatial vertical coordinate, polarization direction index, and pixel intensity;

[0087] Furthermore, to avoid confusion in terminology, the acquisition device in this embodiment is uniformly referred to as a focal plane beam-splitting micro-polarization array camera throughout the text. It means that the same type of camera device integrates a focal plane pixel-level micro-polarizer array of 0 degrees, 45 degrees, 90 degrees and 135 degrees and can synchronously output multi-polarization channel data in a single exposure.

[0088] When the camera or the camera is mentioned in the following text, it is an abbreviation for the above-mentioned focal plane beam-splitting micro-polarization array camera, and not for other types of time-division switching polarization cameras or ordinary industrial cameras.

[0089] In one possible implementation, when a sudden increase in blade movement speed causes exposure ghosting to exceed a set threshold, the controller can shorten the exposure time and simultaneously increase the light source power; if the signal-to-noise ratio requirement still cannot be met, the current frame enters the waste frame recycling queue and is not involved in subsequent analysis.

[0090] If a polarization channel produces a large area of ​​abnormally low values ​​due to local contamination, micro-polarizer failure, or circuit fault, the system will first perform a channel self-check. When the abnormal area exceeds a preset proportion of the total sensor area, a device maintenance alarm will be triggered, and the camera will be suspended from participating in quality assessment. If only a few bad pixels exist, the median value of the neighborhood can be used to maintain data continuity.

[0091] On the same blade final inspection line, the surface reflectivity of a certain batch of workpieces is slightly higher than that of the previous batch; after the delivery encoder sends out the positioning pulse, the system drives the fully polarized light source to output linearly polarized probe light, and the camera obtains instantaneous images of four polarization directions in one exposure.

[0092] This allows the blade to continue moving forward during transport, with all four channels receiving data from the same moment, thus avoiding misalignment caused by time-division shooting; the subsequent processing unit can then perform unified polarization calculations on the blade's leading edge, back, and root transition zone.

[0093] The purpose of this step is to ensure that the moving target obtains spatiotemporally consistent multi-polarization data under single-frame conditions, thereby providing a reliable input basis for subsequent polarization parameter calculations.

[0094] In a preferred embodiment of the present invention, the computing processing unit includes a graphics processor tensor core. In response to the four-dimensional polarization raw data stream, the computing processing unit executes a Stokes decoding matrix to convert the four-dimensional polarization raw data stream into a polarization degree feature matrix and a polarization angle feature matrix. The method includes: inputting the four-dimensional polarization raw data stream into the video memory of the graphics processor tensor core.

[0095] The total light intensity parameter and linear polarization component parameter corresponding to the four-dimensional polarization raw data stream are computed in parallel; based on the total light intensity parameter and linear polarization component parameter, the polarization state of the target object is reconstructed; the reconstructed polarization state is mapped to a polarization degree feature matrix representing the degree of linear polarization and a polarization angle feature matrix representing the polarization angle, respectively.

[0096] This embodiment provides a polarization space mapping mechanism for real-time detection in pipelines. Specifically, although the aforementioned acquisition method can obtain four-directional polarization response, if a general-purpose central processing unit is still used for pixel-by-pixel serial calculation, at a rate of 60 frames per second or even higher, the entire line is prone to accumulated calculation delays, resulting in the blade reaching the rejection position but the judgment result not being output. Therefore, this embodiment directly sends the four-dimensional polarization raw data stream into the graphics processor's video memory and calls the tensor core to perform Stokes decoding matrix calculation in parallel.

[0097] Specifically, when a frame of four-dimensional polarization raw data stream enters the video memory, the system divides it into multiple parallel threads according to image blocks; each thread is responsible for one or more sets of superpixel data.

[0098] For ease of understanding, we will still use the four-channel values ​​of a single pixel location. , , , For example, the system first obtains the total light intensity parameter and the parameters of the two linear polarization components in parallel;

[0099] The total light intensity can be understood as the overall brightness base at that location. The two linear polarization components reflect the differences in the horizontal and vertical directions, and the differences in the oblique and anti-oblique directions, respectively. After obtaining these parameters, the linear polarization state at that location can be reconstructed.

[0100] The system maps the linear polarization state into two matrices that are more suitable for defect analysis; one is the polarization degree feature matrix, which is used to characterize the degree of polarization retention at that location; if a point still retains strong directional polarization after reflection, the value is high; if severe depolarization occurs, the value is low.

[0101] Another type is the polarization angle feature matrix, used to characterize the polarization direction; assuming pixels With adjacent pixels The polarization angles are respectively and If the adjacent pixels are respectively and If so, it indicates that there may be structural disturbances in the local area; by performing the same mapping on the entire frame image, the complete polarization degree feature matrix and polarization angle feature matrix are obtained;

[0102] Furthermore, a block-level parallel strategy can be adopted to improve throughput; for example, an image frame is divided into 16 processing blocks, each processing block independently completes four-channel reading, parameter calculation and matrix write-back; after all blocks are processed, a summarizing thread forms a unified output; since the relationship between the four channels and the Stokes parameters is fixed, the decoding matrix can be preset in the constant area of ​​video memory to reduce the overhead of repeated reading.

[0103] Furthermore, under abnormal or boundary conditions, when the total light intensity parameter at certain pixel locations is close to zero, the polarization degree calculation may result in an excessively small denominator.

[0104] At this point, the system can first determine that the location is an invalid point due to low illumination and mark it as invalid in the output matrix, and then stop calculating the degree of polarization and the polarization angle. If a certain processing block is not completed within the predetermined time limit due to memory congestion, the controller can prioritize outputting the intermediate results corresponding to the completed block and start downsampling and recalculation for the incomplete block to ensure that the overall delay of the whole frame is under control.

[0105] If the calculated polarization angle exceeds the preset angle range, periodic normalization is performed on it, for example, mapping it back. to The interval is used to avoid angle out-of-bounds errors in subsequent entropy calculations;

[0106] On the same blade final inspection line, a highly reflective stripe appeared in the middle of the back of a certain blade, and a large area on the ordinary brightness map was close to saturation; the system directly sent the four-channel polarization raw data of this frame into the graphics processor's video memory and solved it in parallel on the tensor core;

[0107] The results show that although the total light intensity corresponding to the highly reflective strip is relatively high, the polarization angle distribution remains continuous. However, in a small perturbation region at the edge of the strip, the polarization angle changes from... Mutation to These changes will be fully preserved in the polarization angle feature matrix, providing a basis for subsequent high-entropy region localization.

[0108] The purpose of this step is to complete the polarization state reconstruction and feature matrix generation in hardware parallelism, thereby providing a stable and computable polarization representation for the detection of defects on complex surfaces without sacrificing pipeline cycle time.

[0109] In a preferred embodiment of the present invention, the method of setting a sliding window on the polarization angle feature matrix and calculating the polarization angle Shannon entropy within the sliding window to generate a polarization high entropy heatmap includes: taking each pixel in the polarization angle feature matrix as the center pixel in turn; and establishing a sliding window of a preset size based on the center pixel.

[0110] The probability distribution of polarization angle values ​​within the sliding window is statistically analyzed; the local spatial information entropy is calculated based on the probability distribution, which is used as the Shannon entropy of the polarization angle; all pixels in the polarization angle feature matrix are traversed, and the corresponding Shannon entropy of the polarization angle is recorded, and the results are combined to generate a high-entropy polarization heatmap.

[0111] This embodiment provides an anomaly assessment mechanism based on the degree of local polarization angle dispersion; specifically, the polarization angle feature matrix alone is not enough to directly output the defect location, because the complex curved surface itself will also cause the polarization angle to change with the normal, resulting in parameter drift;

[0112] If the polarization angle is directly quantified by the first-order difference, normal curvature transitions may be misidentified as anomalies. Therefore, this embodiment introduces the Shannon entropy of the polarization angle within a sliding window to characterize the true anomaly by the degree of local order and disorder.

[0113] Specifically, the system sequentially uses each pixel in the polarization angle feature matrix as the center pixel and establishes a sliding window of a preset size centered on it; the window size can be set according to the defect scale, for example, for surface damage of 0.1 mm to 0.5 mm, the window size can be selected. or window;

[0114] Within each window, the system first performs binning statistics on the polarization angle; for example, ... to according to The interval is divided into 6 angular intervals, if a certain Nine pixel angles in the window fall at angles exceeding the preset first ratio. to The pixel angle falls within the range, and the proportion is lower than the preset second ratio. to If the 9 pixels fall within a range, it indicates that the window distribution is concentrated and the local entropy is low; if the 9 pixels are scattered across multiple ranges, it indicates that the window distribution is discrete and the local entropy is high.

[0115] For illustrative purposes, let's assume a sliding window as the first example. The nine polarization angles are , , , , , , , , If, after statistical analysis, almost all of them fall within the same angle range, then their local spatial information entropy can be considered low.

[0116] Let's assume a sliding window as a second example. The nine polarization angles are , , , , , , , , If the image is distributed across multiple regions, its local spatial information entropy will increase significantly. After the system repeats this process for all central pixels, it can form a polarization high-entropy heat map of the same or approximately the same size as the original image.

[0117] In this heatmap, continuous smooth areas usually show a low-value flat distribution, while damaged edges, scratch intersections, and micro-pit disturbance areas show local high-value peaks.

[0118] Furthermore, to avoid inconsistencies in calculations due to insufficient pixels at the window edges, mirroring, boundary duplication, or window shrinking can be used to handle the boundary region; to reduce statistical bias caused by angular periodicity, the edges can be aligned before binning. and The angle is used to merge the rings so that similar directions are not misclassified into two extreme intervals;

[0119] Furthermore, under abnormal or boundary conditions, when the number of valid pixels in a window is lower than a set ratio, for example, due to occlusion or bad pixels, only 4 out of 9 pixels have valid values, the entropy value of the window can be directly recorded as invalid, or the interpolation result of the neighboring window can be used to fill it, so as to avoid invalid data inducing false high entropy.

[0120] If the window size is set too small, causing noise to amplify, the system can automatically switch to a larger window for recalculation; if the window size is set too large, causing minor scratches to disappear on average, the system can perform a second local fine calculation in the suspicious area, forming a two-level structure of coarse screening plus fine inspection.

[0121] In the curvature transition region near the blade tip, the polarization angle changes continuously along the surface normal, although the value changes constantly. The areas within the window are still concentrated in a similar angle range, so they appear as low-value bands on the heatmap. Conversely, in a tensile area with a length of about 0.2 mm, there is multi-directional scattering inside, and the polarization angle distribution within the window is obviously discrete, forming a peak on the heatmap at the corresponding position. In this way, normal surface curvature and real defects can be effectively distinguished in the same polarization angle diagram.

[0122] The purpose of this step is to distinguish between the smooth polarization changes caused by the curved surface itself and the local polarization disorder caused by defects, thereby achieving stable enhancement of small abnormal regions.

[0123] In a preferred embodiment of the present invention, the method of extracting polarization topological singularities by performing differential analysis between the polarization high-entropy heat map and the preset baseline polarization manifold model includes: loading the baseline polarization manifold model corresponding to the standard good product;

[0124] The local polarization angle Shannon entropy value in the high-entropy polarization heatmap is compared with the preset low-entropy threshold in the baseline polarization manifold model.

[0125] If the local polarization angle Shannon entropy value is higher than the preset low entropy threshold, it is determined that the corresponding region has undergone depolarization and multiple scattering, and the corresponding region is marked as a polarization topological singularity; if the local polarization angle Shannon entropy value is lower than or equal to the preset low entropy threshold, it is determined that the corresponding region is a continuous smooth surface and is retained.

[0126] This embodiment provides a polarization singularity extraction mechanism based on the baseline of good products; specifically, judging anomalies directly based on high-value areas in a single frame heatmap may still be affected by the inherent structure of the workpiece, processing texture, or local normal boundary effects.

[0127] Especially in areas such as the blade tenon transition zone and the edge of the cooling hole, there may be a higher local entropy value than that of the large plane. In order to avoid misjudging normal structures that are permissible by the process as damage, this embodiment introduces a baseline polarization manifold model and performs differential analysis with the real-time thermal map.

[0128] Specifically, the system pre-collects a set of standard good blades and obtains polarization high-entropy thermal maps under the same light source polarization state, the same camera calibration, and the same workstation posture. These thermal maps are then aligned uniformly according to the workpiece coordinates to form a baseline polarization manifold model. This model can be understood as what kind of entropy distribution a normal blade should have at each position.

[0129] At the system data level, the baseline polarization manifold model is represented as a two-dimensional mesh feature matrix aligned with the topology of the standard workpiece surface. The coordinates of each node in the feature matrix correspond one-to-one with the workpiece coordinate system, and the node data records the average value of the polarization angle Shannon entropy and the tolerance radius obtained by multiple calibrations of the spatial position under normal conditions, thereby forming an allowable low entropy threshold distribution map.

[0130] For example, the baseline low entropy threshold corresponding to the polished surface in the middle of the blade may be 0.25, while the allowable low entropy threshold in the transition zone at the blade root may be 0.42, because the latter itself has more complex normal changes;

[0131] During real-time detection, the system compares the local entropy value of the current polarization high-entropy heatmap with the preset low-entropy threshold at the corresponding position of the baseline model point by point; for ease of explanation, assume the first example detection coordinates If the allowable threshold in the baseline is 0.30, and the currently measured entropy value is 0.22, then the area is determined to still be a continuous smooth surface.

[0132] Let's assume the second example detects coordinates. If the allowable threshold in the baseline is 0.35 and the current measured entropy value is 0.68, it indicates that the degree of polarization disorder at this location is much higher than the normal range. It can be determined that depolarization and multiple scattering enhancement have occurred, and the corresponding region is marked as a polarization topological singularity.

[0133] The differential analysis here can be either a simple threshold comparison or include recording the difference magnitude; if the difference between the real-time entropy value at a certain location and the baseline is larger, it indicates that the polarization anomaly at that location is stronger; the system can merge connected components of consecutive adjacent pixels that exceed the threshold to form candidate regions that are more in line with the physical defect morphology, rather than retaining scattered single points;

[0134] In one possible implementation, if the current workpiece posture has a slight translation or rotation error relative to the baseline model, the system can first perform rigid registration based on the workpiece's outer contour or marker points, and then perform entropy value comparison; if the registration fails, the frame enters the queue to be reshot.

[0135] If certain regions are already marked as undetectable areas in the baseline model, such as deep hole shadow areas or fixture occlusion areas, these regions can be skipped directly during real-time detection and will not participate in singularity judgment.

[0136] If the overall reflectivity of the material surface in the current batch undergoes a uniform drift, resulting in a slightly higher entropy value over a large area but without forming a local abrupt change, the system may temporarily not output a damage conclusion, but instead record it as a batch offset event to be corrected, to prevent the entire batch from being mistakenly rejected.

[0137] At the transition fillet near the mounting platform at the blade root, there is a medium brightness band in the real-time heat map; if viewed alone, it might be considered an anomaly, but after aligning with the baseline model, it was found that this location is already allowed to have a higher entropy value in standard good products, so it was preserved.

[0138] Meanwhile, in a small area at the leading edge of the leaf, the current entropy value significantly exceeds the corresponding baseline threshold and forms a local isolated peak. The system then marks it as a polarization topological singularity and proceeds to the subsequent classification process.

[0139] The purpose of this step is to filter out acceptable high-entropy regions caused by normal structures by performing position-related differential comparison with the polarization baseline of good products, thereby improving the accuracy of real anomaly extraction.

[0140] In a preferred embodiment of the present invention, after extracting the polarization topological singularity, the method further includes:

[0141] The mean polarization degree of the region corresponding to the polarization topological singularity in the polarization degree feature matrix is ​​extracted as the first feature value;

[0142] The spatial rate of change of the region corresponding to the polarization topological singularity in the polarization angle feature matrix is ​​calculated as the second feature gradient.

[0143] Obtain the preset depolarization threshold and the preset polarization angle abrupt change threshold;

[0144] If the first feature value is lower than the preset depolarization threshold and the second feature gradient is lower than or equal to the preset polarization angle abrupt change threshold, then the polarization topological singularity is determined to be a surface attachment.

[0145] If the second feature gradient is higher than the preset polarization angle abrupt change threshold, and the first feature value is higher than or equal to the preset depolarization threshold, then the polarization topological singularity is determined to be a real structural damage; if the above conditions are not met, then the polarization topological singularity is determined to be an undefined anomaly.

[0146] This embodiment provides a morphological discrimination mechanism for singularity regions. Specifically, extracting only polarization topological singularities is not enough, because in online production, the blade surface may simultaneously contain dust, polishing residue, coolant adhesion, and actual scratches. If these conditions are not distinguished, the system will treat cleanable contamination and irreparable damage together, resulting in excessive rejection. Therefore, this embodiment further combines the mean polarization degree and the spatial change rate of polarization angle for secondary judgment after singularity extraction.

[0147] Specifically, the system calculates the average polarization degree of each singularity region in the polarization degree feature matrix, which is used as the first feature value. The average polarization degree reflects the overall depolarization degree of the region. Under normal circumstances, loose attachments or dust particles will reduce the polarization retention ability of reflected light, so the first feature value is low.

[0148] The system calculates the spatial rate of change of the region in the polarization angle feature matrix as the second feature gradient; this quantity can be represented by the average absolute value, maximum value or local gradient intensity of the angle difference between adjacent pixels.

[0149] In a preferred embodiment of the present invention, the spatial change rate is calculated by using the average absolute value of the difference in polarization angle between the center pixel and its eight neighboring pixels, in order to quantify the severity of the sudden change in local polarization direction; real scratches usually cause polarization angle transitions with clear geometric boundaries, so the second feature gradient is too high;

[0150] For ease of understanding, let's assume a singularity region. The average degree of polarization is 0.12, and the spatial variation rate is 8; the preset depolarization threshold is 0.20, and the polarization angle abrupt change threshold is 15. If the combination of low polarization degree and non-drastic angle change is met, it can be identified as a surface deposit.

[0151] Re-hypothetical region The mean degree of polarization is 0.46, and the spatial variation rate is 28. Therefore... If the combination of characteristics of not low polarization degree and obvious angle change is met, it can be determined as real structural damage; if the region If the average polarization degree is 0.18 and the spatial variation rate is 22, indicating a mixed state, then it will not be forcibly classified for the time being, but will be marked as an undefined anomaly and transferred to manual review or subsequent multi-frame review.

[0152] Furthermore, the system can first perform connected component segmentation on the singularity region, and then count two features within each connected component to avoid single-point noise interference; for slender regions, the system can also count the rate of change of angle along the main direction and the vertical direction respectively to enhance the ability to identify directional damage such as scratches.

[0153] Furthermore, to avoid the polarization angle itself having... to The periodicity causes the spatial rate of change to be overestimated, so the system uses a periodically consistent angle difference rule when calculating the second feature gradient.

[0154] Specifically, the angle difference between any two adjacent pixels can be processed according to the minimum circumference difference, that is, the angle difference is taken as... and The smaller value in; thus, when adjacent pixels are respectively and When, its rate of spatial change is considered Instead This avoids misinterpreting the originally continuous polarization direction as a drastic change; thus, the second feature gradient more realistically reflects the local structural boundary rather than the angular coordinate representation itself.

[0155] In one possible implementation, when the singularity region is too small and contains only 1 or 2 effective pixels, the mean polarization degree and spatial change rate are unstable. In this case, the system can temporarily mark it as a minor anomaly and wait for repeated observations in adjacent frames before determining the category.

[0156] If both extremely low polarization degree and extremely high angular change rate appear in a region, it indicates that the region may be a composite anomaly, such as a droplet covering a scratch. In this case, it can be classified as an undefined anomaly to avoid incorrectly entering a single category. If a batch of workpieces has a uniform protective film residue on its surface, causing a large number of areas to decrease in polarization degree, the system can increase the depolarization threshold or suspend the automatic determination of attachments through batch rules to prevent batch misclassification.

[0157] In the same blade final inspection line, the system first extracts two singularities on the blade surface. The first singularity is located on the back of the blade near the trailing edge. The area is relatively scattered, the average polarization degree is significantly lower, and the polarization angle changes relatively gently. It is ultimately classified as coolant deposits and does not trigger scrapping. It is only sent to the cleaning and rework channel. The second singularity is located in the middle of the leading edge polishing zone. The polarization degree remains high, but the polarization angle changes sharply along the local boundary. It is classified as real structural damage and enters the rejection control path.

[0158] The purpose of this step is to use the combination of polarization retention capability and polarization direction change characteristics to distinguish between attached materials and actual damage, thereby reducing false rejection and improving the process feasibility of the detection results.

[0159] In a preferred embodiment of the present invention, two-dimensional physical coordinates and shape classification instructions for the workpiece coordinate system are generated by mapping the polarization topological singularity and the spatial calibration parameters of the pre-acquired focal plane beam-splitting micro-polarization array camera.

[0160] The physical control feedback based on morphological classification instruction output includes: extracting the two-dimensional pixel coordinates of polarization topological singularities that are determined to be real structural damage;

[0161] By combining the spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, the two-dimensional pixel coordinates are mapped to the two-dimensional physical coordinates of the target object in the workpiece coordinate system;

[0162] The maximum value of the polarization angle Shannon entropy in the region corresponding to the polarization topological singularity that is determined to be a real structural damage is extracted as the peak value of the polarization angle Shannon entropy.

[0163] Based on the peak value of the Shannon entropy at the polarization angle, a morphology classification instruction is generated; based on the two-dimensional physical coordinates of the workpiece coordinate system and the morphology classification instruction, a rejection signal for controlling the external pneumatic rejection device is generated and output.

[0164] This embodiment provides a closed-loop control mechanism from image judgment results to physical actions on the production line; specifically, the aforementioned steps can identify actual structural damage, but if the results remain at the image level, they cannot directly drive the production line actuators.

[0165] Especially under high-speed conveying conditions, if the workpiece coordinates and action instructions cannot be given in time, the detection results will not be able to serve as an effective triggering condition for physical control feedback. Therefore, in this embodiment, the pixel coordinates corresponding to the actual structural damage are mapped to the workpiece coordinate system, and the polarization angle Shannon entropy peak value is combined to generate a morphological classification instruction for controlling the external pneumatic rejection device.

[0166] Specifically, the system first extracts two-dimensional pixel coordinates from the singularity region that is determined to be a real structural damage; these coordinates can be taken as the centroid of the region or as the pixel position where the entropy peak is located; the system calls the spatial calibration parameters of the camera to transform the pixel coordinates into two-dimensional physical coordinates in the workpiece coordinate system;

[0167] The calibration parameters here can be determined by the standard calibration plate, the fixture positioning reference, and the workpiece clamping posture. For ease of understanding, it is assumed that after calibration mapping, the pixels with horizontal and vertical coordinates of 520 and 310 in the image correspond to the positions of 12.4 mm and 36.8 mm in the blade workpiece coordinate system. These physical coordinates can then be used as the basis for the actuator to identify the location of the damage.

[0168] On the other hand, the system also extracts the maximum value of the polarization angle Shannon entropy in the actual structural damage region as a peak index; this peak value can be used to reflect the intensity of local polarization disorder caused by damage.

[0169] For example, if the peak value is lower than or equal to the preset severity level judgment threshold, a minor scratch instruction can be generated; if the peak value is higher than the preset severity level judgment threshold, a deep damage or edge chipping instruction can be generated; the specific interval division can be set according to the enterprise's quality standards; the controller encapsulates the workpiece coordinates and shape classification instructions into rejection signals, and sends execution control to the external pneumatic rejection device when the blade moves with the conveyor belt to the corresponding rejection position.

[0170] Furthermore, to maintain consistency with the preceding classification logic, the morphological classification instruction in this embodiment specifically refers to the generation of a second-level instruction for the severity of damage or the specific subtype of damage after the first-level category of actual structural damage has been determined in the previous step; surface attachments, actual structural damage, and undefined anomalies are used to complete the first-level category determination.

[0171] In this step, the instructions for minor scratches, deep scratches, and edge cracks obtained based on the peak value of the Shannon entropy of the polarization angle are used to further subdivide within the actual structural damage category in order to determine the priority of rejection, rework marking, or subsequent re-inspection. Therefore, this step does not overturn the category conclusion of the previous step, but adds a more granular process execution label on its basis.

[0172] Furthermore, to ensure accurate timing of actions, the system can combine the encoder pulse, workpiece spacing, and pneumatic response delay to perform feedforward compensation for rejection timing; if the same workpiece has multiple damage points, it can be uniformly rejected according to the most severe morphology classification instruction, or multi-coordinate information can be output for use by the back-end re-inspection station.

[0173] In one possible implementation, if the pixel coordinates are found to exceed the effective workpiece boundary when mapped to the workpiece coordinate system, it indicates that there may be an attitude recognition error or foreign object occlusion. The system can cancel the rejection action of the current coordinates and trigger a review. If the shape classification instruction does not reach the rejection threshold and only belongs to the reworkable range, the system can change to output a mark or diversion instruction instead of directly rejecting.

[0174] If instantaneous fluctuations in the conveying speed cause the time deviation of the expected arrival at the rejection position to exceed the allowable range, the controller can recalculate the trigger time based on the latest encoder feedback; if the action accuracy cannot be guaranteed within the safety window, the workpiece will be sent to the re-inspection channel to avoid mistakenly blowing adjacent good products.

[0175] During the inspection of blade B017, the system confirmed the existence of real structural damage in the leading edge region and mapped the pixel where its entropy peak was located to the horizontal coordinate of 12.4 mm and the vertical coordinate of 36.8 mm in the workpiece coordinate system; the polarization angle Shannon entropy peak of this region reached the set high-level range, thus generating a deep scratch command.

[0176] The controller combines the current conveying speed with the cylinder response time and sends a rejection signal before the blade reaches the rejection position. The pneumatic device blows the blade into the defective product collection tank, while other blades with only attached materials do not trigger this action.

[0177] The purpose of this step is to directly convert the abnormal information obtained from polarization detection into executable production line control commands, thereby achieving an integrated closed loop of online quality inspection and automatic rejection.

[0178] In a preferred embodiment of the present invention, the method for extracting the local Stokes vector features corresponding to the polarization topological singularity and updating the baseline polarization manifold model includes:

[0179] Optical response data of the region corresponding to the polarization topological singularity identified as real structural damage is extracted from the four-dimensional polarization raw data stream; local Stokes vector features are calculated based on the optical response data;

[0180] The local Stokes vector features are output to the external edge server, which then adjusts the preset low-entropy threshold in the baseline polarization manifold model based on the local Stokes vector features, thereby completing the dynamic update of the baseline polarization manifold model.

[0181] This embodiment provides a baseline adaptive update mechanism for long-term operation scenarios. Specifically, if the baseline polarization manifold model is fixed for a long time once it is established, the parameters will drift as batch materials, polishing processes, light source aging and lens contamination occur, and the model will gradually deviate from the actual production line state, eventually leading to an increase in false alarms or false alarms.

[0182] Therefore, after completing the physical removal, this embodiment does not terminate the data link, but further extracts the local Stokes vector features corresponding to the real structural damage area and sends them to the external edge server to adjust the preset low entropy threshold in the baseline model.

[0183] Specifically, the system first backtracks the optical response data of the actual structural damage area in the four-dimensional polarization raw data stream; since the pixel coordinates of this area have been determined in the aforementioned process, the four-channel raw response at the corresponding position can be directly extracted.

[0184] For example, the four-channel values ​​of three representative pixels in a certain damaged area are the first set of parameters 118, 82, 46, 71, the second set of parameters 122, 79, 43, 69, and the third set of parameters 115, 84, 48, 74. The system recalculates the local Stokes vector features of these original responses to describe the lower-level polarization state of the anomalous area.

[0185] These local Stokes vector features are packaged and sent to an external edge server; the edge server does not directly replace the online decision link, but serves as a model maintenance unit with a slower time scale, performing statistical analysis on abnormal and good samples accumulated over a period of time.

[0186] Here, the local Stokes vector features of the real structural damage area are not directly written into the good product baseline model, nor are they used as the target distribution of the normal surface. Instead, they are used as an exclusionary reference in the threshold update process. Specifically, the edge server uses these abnormal features to define the polarization anomaly boundaries that must not be absorbed by the baseline. Combined with the good product samples from the same period, rework and re-inspection results, or manual verification labels, the preset low entropy threshold of each workpiece area is adjusted under constraints.

[0187] If it is found that in a certain workstation area, among the recently acquired good samples exceeding the preset comparison threshold, all show a positive shift in overall entropy value without exceeding the preset tolerance range, and the local Stokes vector features of the confirmed damaged samples do not coincide with this drift pattern, then the edge server can appropriately increase the preset low entropy threshold for that area. Conversely, if the Stokes vector features of some damaged samples continuously show abnormalities prematurely in a certain area, and the original threshold is too high, leading to edge-related missed detections, then the corresponding threshold can be appropriately decreased. The updated baseline polarization manifold model is then transmitted back to the online system for subsequent workpiece inspection.

[0188] For ease of understanding, model updates can be viewed as partitioned adjustments rather than global replacements; for example, if the surface texture of the blade leading edge polishing zone changes slightly due to a new polishing wheel, the edge server can adjust the low entropy threshold of that region only from 0.30 to 0.34, while the threshold of the blade root region remains unchanged.

[0189] This avoids the overall model being skewed by the characteristics of a single region; moreover, the above adjustment corresponds to the correction of the allowable entropy range of the good product region, rather than rewriting the identified real structural damage region as a normal region. Therefore, the updated model still maintains its rejection of defects.

[0190] In one possible implementation, if the number of abnormal samples uploaded in a certain period of time is insufficient, the edge server can maintain the original threshold unchanged to prevent a small number of occasional samples from causing model overfitting.

[0191] If there are obvious acquisition failure features in the uploaded local Stokes vector features, such as four channels being abnormally low at the same time or large areas being saturated, these samples will be marked as abnormal samples and will not participate in the model update. If the difference between the old and new baseline models exceeds the preset safety boundary, the update action will not take effect directly, but will enter the manual review or gray release process to verify its stability on a small batch of workpieces first.

[0192] After the turbine blade final inspection line has been running for a week, the system has uploaded multiple batches of local Stokes vector features of real structural damage areas.

[0193] Edge server analysis revealed that with the introduction of a new batch of materials, the polarization response in the normal region of the blade back changed slightly, causing a small number of false alarms in this region in the online system. Based on this, the server only adjusted the low-entropy threshold corresponding to this region and issued a new baseline model. After the update, the online system significantly reduced the false alarms in this region while maintaining its sensitivity to real scratches.

[0194] The purpose of this step is to enable the baseline polarization manifold model to dynamically adapt to changes in the actual optical conditions of the production line, thereby maintaining detection stability and judgment consistency during long-term operation.

[0195] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An industrial product quality management system based on vision inspection, characterized in that, include: Optical field modulation and acquisition unit: controls a programmable fully polarized light source to modulate the emission polarization state of the target object; The light waves reflected from the target object are captured by a focal plane array camera to generate a four-dimensional polarization raw data stream. Polarization space mapping unit: In response to the four-dimensional polarization raw data stream, the computational processing unit configured by the polarization space mapping unit executes a Stokes decoding matrix to linearly map the four-channel light intensity to three Stokes parameters, and converts the four-dimensional polarization raw data stream into a polarization degree feature matrix and a polarization angle feature matrix. Polarization entropy gradient evaluation unit: Set a sliding window on the polarization angle feature matrix, calculate the polarization angle Shannon entropy within the sliding window, and generate a polarization high-entropy heat map; perform difference analysis between the polarization high-entropy heat map and the preset baseline polarization manifold model to extract polarization topological singularities; Physical closed-loop and adaptive update unit: Based on the polarization topological singularity and the pre-acquired spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, it generates two-dimensional physical coordinates and shape classification instructions for the workpiece coordinate system; outputs physical control feedback based on the shape classification instructions; extracts the local Stokes vector features corresponding to the polarization topological singularity, and updates the baseline polarization manifold model.

2. The industrial product quality management system based on vision inspection according to claim 1, characterized in that, The programmable fully polarized light source modulates the emission polarization state of the target object. The methods for capturing the light waves reflected from the target object using a focal plane array camera to generate a four-dimensional polarization raw data stream include: The programmable fully polarized light source is activated to project a probe light field with a preset polarization state onto the moving target object; The focal plane array camera is triggered to perform a single exposure; the light field data reflected from the surface of the target object is acquired. Extract the zero-degree channel data, forty-five-degree channel data, ninety-degree channel data and one hundred and thirty-five-degree channel data that are integrated at the pixel level inside the focal plane beam-splitting micro-polarization array camera; The zero-degree channel data, the forty-five-degree channel data, the ninety-degree channel data, and the one hundred and thirty-five-degree channel data are combined to generate the four-dimensional polarization raw data stream.

3. The industrial product quality management system based on vision inspection according to claim 1, characterized in that, The computational processing unit includes a graphics processing unit tensor core. The method of converting the four-dimensional polarization raw data stream into a polarization degree feature matrix and a polarization angle feature matrix by executing a Stokes decoding matrix through the computational processing unit in response to the four-dimensional polarization raw data stream includes: The four-dimensional polarization raw data stream is input into the video memory of the graphics processor tensor core; The total light intensity parameters and linear polarization component parameters corresponding to the four-dimensional polarization raw data stream are calculated in parallel. Based on the total light intensity parameter and the linear polarization component parameter, the polarization state of the target object is reconstructed; The reconstructed polarization states are mapped to a polarization degree feature matrix representing the degree of linear polarization and a polarization angle feature matrix representing the polarization angle, respectively.

4. The industrial product quality management system based on vision inspection according to claim 3, characterized in that, The method of setting a sliding window on the polarization angle feature matrix and calculating the polarization angle Shannon entropy within the sliding window to generate a polarization high-entropy heatmap includes: taking each pixel in the polarization angle feature matrix as the center pixel in turn. Using the center pixel as a reference, establish a sliding window of a preset size; statistically analyze the probability distribution of polarization angle values ​​within the sliding window; The local spatial information entropy is calculated based on the probability distribution and used as the polarization angle Shannon entropy. Traverse all pixels in the polarization angle feature matrix, record the corresponding polarization angle Shannon entropy, and combine them to generate the polarization high entropy heatmap.

5. The industrial product quality management system based on vision inspection according to claim 4, characterized in that, The polarization high-entropy heatmap is compared with the preset baseline polarization manifold model to extract polarization topological singularities. The method of extracting polarization topological singularities includes: loading the baseline polarization manifold model corresponding to the standard good product. The local polarization angle Shannon entropy value in the polarization high-entropy heatmap is compared with the preset low-entropy threshold in the baseline polarization manifold model. If the local polarization angle Shannon entropy value is higher than the preset low entropy threshold, it is determined that the corresponding region has undergone depolarization and multiple scattering, and the corresponding region is marked as the polarization topological singularity; if the local polarization angle Shannon entropy value is lower than or equal to the preset low entropy threshold, it is determined that the corresponding region is a continuous smooth surface and is retained.

6. The industrial product quality management system based on vision inspection according to claim 5, characterized in that, After extracting the polarization topological singularity, the process also includes: The mean polarization degree of the region corresponding to the polarization topological singularity in the polarization degree feature matrix is ​​extracted as the first feature value; The spatial rate of change of the region corresponding to the polarization topological singularity in the polarization angle feature matrix is ​​calculated as the second feature gradient; Obtain the preset depolarization threshold and the preset polarization angle abrupt change threshold; If the first feature value is lower than the preset depolarization threshold, and the second feature gradient is lower than or equal to the preset polarization angle abrupt change threshold, then the polarization topological singularity is determined to be a surface attachment. If the second feature gradient is higher than the preset polarization angle abrupt change threshold, and the first feature value is higher than or equal to the preset depolarization threshold, then the polarization topological singularity is determined to be a real structural damage; if the above conditions are not met, then the polarization topological singularity is determined to be an undefined anomaly.

7. The industrial product quality management system based on vision inspection according to claim 6, characterized in that, Based on the polarization topological singularity and the pre-acquired spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, the workpiece coordinate system two-dimensional physical coordinates and shape classification instructions are generated. The method of outputting physical control feedback based on the morphological classification instruction includes: extracting the two-dimensional pixel coordinates of the polarization topological singularity that is determined to be the real structural damage; By combining the spatial calibration parameters of the focal plane beam-splitting micro-polarization array camera, the two-dimensional pixel coordinates are mapped to the two-dimensional physical coordinates of the target object in the workpiece coordinate system; The maximum value of the polarization angle Shannon entropy in the region corresponding to the polarization topological singularity that is determined to be the real structural damage is extracted as the peak value of the polarization angle Shannon entropy. Based on the peak value of the Shannon entropy at the polarization angle, the morphology classification instruction is generated; based on the two-dimensional physical coordinates of the workpiece coordinate system and the morphology classification instruction, a rejection signal for controlling the external pneumatic rejection device is generated and output.

8. The industrial product quality management system based on vision inspection according to claim 7, characterized in that, The methods for extracting the local Stokes vector features corresponding to the polarization topological singularity and updating the baseline polarization manifold model include: Extract optical response data from the region corresponding to the polarization topological singularity identified as the actual structural damage in the four-dimensional polarization raw data stream; calculate local Stokes vector features based on the optical response data; The local Stokes vector features are output to an external edge server, which then adjusts the preset low-entropy threshold in the baseline polarization manifold model based on the local Stokes vector features, thereby completing the dynamic update of the baseline polarization manifold model.

Images