An industrial surface defect pure infrared light scanning detection system

Abstract

The application relates to the technical field of industrial nondestructive testing, and discloses an industrial surface defect pure infrared light scanning detection system which comprises optical acquisition, frequency domain equalization, energy modulation, ratio construction, matrix decomposition and defect output modules. The system acquires first and second light source reflection signals and dark field noise signals in time sequence; a frequency domain equalization module constructs a space-variable equalization kernel, filters the second light source reflection signal from which the dark field noise is removed to realize frequency alignment; a ratio construction module calculates dark field thermal drift variance to generate a dynamic sparse penalty parameter, and combines the aligned signal to construct differential ratio data to generate a two-dimensional observation matrix; a matrix decomposition module decomposes the observation matrix by using a space weight matrix and the penalty parameter to output a sparse anomaly matrix; and a defect output module binarizes the anomaly matrix and maps and outputs physical coordinates. The application adaptively suppresses detector thermal drift noise, offsets material emissivity interference, eliminates space spread difference, and improves defect extraction precision under a complex background.

Description

Technical Field

[0001] This invention relates to the field of industrial nondestructive testing technology, specifically to a pure infrared light scanning detection system for industrial surface defects. Background Technology

[0002] Industrial production lines impose stringent requirements on workpiece surface quality control during continuous manufacturing. Pure infrared scanning inspection systems are now widely used to detect physical morphological defects on surfaces of various materials. Infrared imaging technology captures the distribution of infrared energy reflected from the target surface, directly reflecting the three-dimensional undulations and geometric changes of the workpiece surface. This non-contact machine vision inspection method is well-suited to the high-paced operating environment of industrial sites, meeting the objective needs of non-destructive online quality inspection in large-scale manufacturing.

[0003] Existing infrared inspection equipment typically consists of an infrared imaging detector vertically mounted above the workpiece. Fixed infrared illumination sources are positioned on either side of the detector's field of view. To highlight minute three-dimensional geometric features, conventional solutions often employ asymmetric optical illumination structures. The two light sources project infrared beams onto the workpiece surface at a specific angle. The system sequentially acquires images of the workpiece surface reflection under different illumination conditions. The data processing stage usually involves directly performing a global subtraction or simple division operation on the image data generated by the two light sources. For complex background textures, some processing solutions introduce low-rank and sparse matrix factorization techniques to separate sparse points representing defects. Finally, the system binarizes outliers using a preset fixed grayscale threshold and outputs their coordinates.

[0004] However, existing infrared scanning detection technologies suffer from temperature rise due to continuous operation of the infrared detector, leading to increased internal dark currents and time-varying dark-field thermal drift noise. Fixed-parameter algorithm models are prone to divergence when environmental thermal noise dynamically increases. Asymmetrically arranged light sources on both sides exhibit objective differences in physical divergence angle mapping, resulting in misaligned spatial frequency responses in reflected images acquired from different illumination angles. Normal physical edges show differentiated optical broadening under different light source illumination, and the dark-field thermal drift noise generated during continuous operation of the infrared detector makes it difficult for fixed-parameter decomposition algorithms to converge or produces numerous false defect reports. Therefore, this invention provides a pure infrared light scanning detection system for industrial surface defects to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a pure infrared light scanning detection system for industrial surface defects. This system solves the problems of existing industrial infrared surface defect detection technologies, which struggle to effectively extract minute physical and geometric defect signals when dealing with drastic fluctuations in the absolute emissivity of workpiece surface materials and complex background textures. Furthermore, the asymmetry in the physical distribution of the light source can lead to inconsistent spatial frequency responses and edge artifacts. Additionally, the dark field thermal drift noise generated by the infrared detector during continuous operation can cause the fixed-parameter decomposition algorithm to fail to converge or generate a large number of false defect reports.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a pure infrared light scanning detection system for industrial surface defects, comprising:

[0007] The optical acquisition module is used to acquire the reflection signals of the first light source, the reflection signals of the second light source, and the dark field noise signal in a time-division multiplexed sequence.

[0008] The frequency domain equalization module is used to construct a spatially variable equalization kernel based on the divergence angle distribution, and to use the spatially variable equalization kernel to filter the denoised second light source reflection signal to generate a frequency-aligned second light source reflection signal.

[0009] An energy modulation module is used to extract a geometric gradient using the reflection signal of the first light source and the frequency-aligned reflection signal of the second light source to adjust the luminous energy.

[0010] The ratio construction module is used to calculate the thermal drift variance based on the dark field noise signal to generate dynamic sparse penalty parameters, and to construct differential ratio data splicing using the first light source reflection signal and the frequency-aligned second light source reflection signal to generate a two-dimensional observation matrix;

[0011] The matrix decomposition module is used to decompose the two-dimensional observation matrix using the spatial weight matrix of the divergence angle difference and the dynamic sparsity penalty parameter to output a sparse anomaly matrix.

[0012] The defect output module is used to binarize and coordinate map the sparse anomaly matrix to output the surface defect detection results.

[0013] Preferably, the optical acquisition module includes an imaging receiving unit, a trigger control unit, and an asymmetric optical illumination unit:

[0014] The imaging receiving unit is vertically positioned above the workpiece being measured. The trigger control unit is coaxially connected to the mechanical transmission mechanism and outputs a trigger cycle signal and a rigid, constant exposure window time to the imaging receiving unit according to the equidistant physical movement distance.

[0015] The asymmetric optical illumination unit is arranged in parallel on both sides of the field of view of the imaging receiving unit, including a first light source and a second light source with a spatially gradient optical lens mounted on the surface. The first light source and the second light source generate a spatially asymmetric physical divergence angle mapping function on the pixel spatial coordinate axis.

[0016] The trigger control unit outputs a three-phase time-division multiplexing trigger timing control signal within the exposure window time, respectively driving the first light source to light up individually, the second light source to light up individually, and the two light sources to turn off synchronously, so as to control the imaging receiving unit to generate the first light source reflection signal, the second light source reflection signal, and the dark field noise signal in sequence without cross-boundary blurring.

[0017] Preferably, the frequency domain equalization module internally achieves frequency consistency reconstruction through the coordinated operation of a background denoising unit, an equalization kernel construction unit, and a convolutional filtering unit.

[0018] The background noise reduction unit uses the dark field noise signal at the same spatial coordinate position to directly cancel it out through a subtraction operation and outputs the first light source reflection signal and the second light source reflection signal after removing the dark field noise signal.

[0019] The equalization kernel construction unit extracts the absolute difference distribution between the first light source and the second light source on the physical divergence angle mapping function, and establishes a linear mapping relationship of local standard deviation in combination with the optical broadening mapping constant of the system, thereby dynamically constructing a one-dimensional spatially variable Gaussian equalization kernel for each pixel coordinate position.

[0020] The convolutional filtering unit applies the one-dimensional spatially variable Gaussian equalization kernel to the second light source reflection signal after removing the dark field noise signal, and performs a one-dimensional sliding window multiplication and addition operation pixel by pixel using a boundary extension mechanism in the direction of the pixel space coordinate axis, thereby eliminating the edge differences caused by the divergence angle asymmetry and driving the generation of the frequency-aligned second light source reflection signal.

[0021] Preferably, the ratio construction module is internally configured with a thermal drift evaluation unit to achieve adaptive adjustment of the penalty threshold:

[0022] The thermal drift evaluation unit uses a ring cache with a fixed row depth to continuously store the dark field noise signal of the historical scanning cycle, and extracts all dark field image data in the entire cache spatiotemporal domain to calculate the temporal thermal drift variance that deviates from the global statistical mean.

[0023] The thermal drift evaluation unit further performs a linear combination mapping between the calculated temporal thermal drift variance and the pre-calibrated basic penalty constant and regularization scaling factor, driving the generation of the dynamic sparse penalty parameters used to offset the dark flow artifacts of the high-temperature detector.

[0024] Preferably, the ratio construction module completes the cross-dimensional reorganization of data through its internal ratio calculation unit and cache splicing unit:

[0025] The ratio calculation unit uses the difference between the first light source reflection signal after removing the dark field noise signal at the same pixel position and the second light source reflection signal aligned with the frequency as the molecular data.

[0026] The ratio calculation unit uses the sum of the two sets of signals plus a normal number as the normalized denominator data, and constructs a single row of spatial reflection differential ratio data through division operation, so as to completely cancel the interference of the absolute emissivity constant of the material on the workpiece surface at the mathematical level.

[0027] The buffer stitching unit opens a sliding buffer inside and pushes the spatial reflection differential ratio data output line by line into the stack in sequence according to the scanning time. When the number of data lines reaches the preset horizontal cell depth, the two-dimensional observation matrix containing the planar topology is stitched together and the clear and reset operation is performed to maintain the data source replacement required for manifold optimization with non-overlapping block step size.

[0028] Preferably, the matrix factorization module combines a priori weight construction unit with a matrix weighting unit to inject physical prior knowledge into subsequent optimization:

[0029] The prior weight construction unit extracts the absolute difference between the physical divergence angle mapping functions of the first light source and the second light source and performs maximum value normalization, and then combines the weight amplification coefficient to generate a one-dimensional weight feature.

[0030] The prior weight construction unit expands and replicates the one-dimensional weight feature by equal value in the scanning row direction, thereby constructing a two-dimensional spatial weight matrix that perfectly matches the dimension of the two-dimensional observation matrix.

[0031] The matrix weighting unit receives the two-dimensional spatial weight matrix and the two-dimensional observation matrix generated by the front end, and generates a priori weighted observation matrix by performing element-wise Hadamard product operation between the two to numerically amplify and display the geometric scattering gradient of the real defect edge under asymmetric illumination.

[0032] Preferably, the matrix factorization module utilizes convex optimization iterative units to extract defects from complex backgrounds:

[0033] The convex optimization iterative unit calculates the arithmetic mean of the dynamic sparsity penalty parameters within the current matrix block as the global dynamic sparsity penalty parameters, and uses them as sparsity constraints. Combined with the prior weighted observation matrix, it constructs an augmented Lagrange objective function.

[0034] The convex optimization iterative unit uses the alternating direction multiplier method to iteratively update the low-rank background matrix and the sparse anomaly matrix in the augmented Lagrange objective function, and uses the singular value threshold shrinkage operator to solve the low-rank background matrix, and uses the soft threshold shrinkage operator combined with the global dynamic sparse penalty parameter to solve the sparse anomaly matrix.

[0035] The convex optimization iterative unit updates the iteration step size parameter monotonically according to the set amplification factor after each iteration. When the residual matrix reaches the preset residual convergence threshold or the rounds are exhausted, the iteration process is blocked, and the sparse anomaly matrix representing pure geometric mutation is separated from the aliased signal and output.

[0036] Preferably, the defect output module uses a threshold binarization unit to perform mathematical truncation determination of the morphology:

[0037] The threshold binarization unit extracts the statistical variance of the decomposition residual of the defect-free sample to adaptively generate a global hard threshold.

[0038] The threshold binarization unit iterates through the absolute values ​​of each element in the sparse anomaly matrix one by one, sets the coordinates that exceed the global hard threshold to a valid high level state, and otherwise sets them to a logic zero state, thereby converting the floating-point matrix into a binary defect mask that only contains morphological contour features.

[0039] Preferably, the defect output module utilizes the coordinate mapping unit to perform a spatial reconstruction process from relative pixels to global physical coordinates:

[0040] The coordinate mapping unit uses the eight-neighborhood connectivity algorithm to aggregate non-zero element connected regions in the binarized defect mask, removes pseudo-noise patches with pixel areas smaller than a preset area filtering threshold, and calculates the arithmetic mean of the row and column indices of each pixel for the effective non-zero element connected regions to generate the relative geometric centroid position coordinates.

[0041] The coordinate mapping unit obtains the total offset parameter of the global scan row recorded synchronously, and adds it to the relative row index in the relative geometric centroid position coordinates to transform it into an absolute row coordinate with global uniqueness.

[0042] The coordinate mapping unit integrates the horizontal spatial resolution, vertical pulse resolution, and the coordinate parameters of the initially calibrated physical origin on the workpiece of the system, and projects the absolute pixel index system in reverse to the physical coordinate system of the workpiece through linear transformation, thereby generating continuous horizontal and vertical physical coordinates without any breaks.

[0043] Preferably, the defect output module is equipped with an output execution unit at its end to achieve industrial interconnection execution:

[0044] The output execution unit will convert the generated horizontal physical coordinates, the vertical physical coordinates, the calculated physical area of ​​the connected domain with effective non-zero elements, and the corresponding timestamp parameters, and perform frame packing operations according to the preset industrial bus data structure.

[0045] The packaged industrial location guidance data is verified by the underlying data link layer and then pushed to the downstream external actuators through the industrial Ethernet interface to directly drive the external actuators to perform defective product rejection or laser marking processes with real physical location significance.

[0046] This invention provides a pure infrared light scanning detection system for industrial surface defects. It has the following beneficial effects:

[0047] 1. This invention constructs spatial reflection difference ratio data by dividing the difference between the reflection signal of a first light source at the same spatial location and the reflection signal of a second light source with aligned frequencies, using the difference as the numerator and the sum of the two sets of signals as the denominator. This process cancels out the interference of the absolute emissivity constant of the workpiece surface material. Simultaneously, the system utilizes dark field noise signals to calculate the temporal thermal drift variance to generate dynamic sparsity penalty parameters. This adaptively suppresses dark flow artifacts caused by continuous temperature changes in the infrared detector, reducing the false defect detection rate in complex background texture environments.

[0048] 2. This invention extracts the absolute difference between the first and second light sources in their physical divergence angle mapping functions, and dynamically constructs a one-dimensional spatially variable Gaussian equalizer kernel by combining it with the system's optical broadening mapping constant. This equalizer kernel is then applied to the second light source's reflected signal after removing dark-field noise to perform sliding window convolution filtering, eliminating the edge physical broadening difference caused by light source asymmetry and achieving frequency alignment of the two signals.

[0049] 3. This invention injects prior physical knowledge into the matrix factorization algorithm during the defect separation stage, converting the absolute difference in the divergence angle of the light source into a two-dimensional spatial weight matrix and applying it to the two-dimensional observation matrix, thereby directionally amplifying the geometric scattering gradient of the edge of a real physical defect under asymmetric illumination. Combined with the aforementioned globally dynamic sparse penalty parameters as constraints, the objective function is iteratively updated using the alternating direction multiplier method. This accurately extracts the sparse anomaly matrix representing pure geometric abrupt changes from the low-rank background data, improving the accuracy of identifying and extracting minute physical defects. Attached Figure Description

[0050] Figure 1 This is a system architecture diagram of the present invention;

[0051] Figure 2 The image shows a comparison of the three-dimensional grayscale distribution of the original observation matrix and the reconstructed sparse anomaly matrix of the present invention, where (a) is the original observation matrix image and (b) is the reconstructed sparse anomaly matrix image.

[0052] Figure 3 This is a graph showing the dynamic sparsity penalty parameter of the present invention as a function of temperature drift.

[0053] Figure 4 This is a comparison chart of the ROC of the algorithm detection rate in one embodiment of the present invention.

[0054] Among them, 10 is the optical acquisition module; 20 is the frequency domain equalization module; 30 is the energy modulation module; 40 is the ratio construction module; 50 is the matrix decomposition module; and 60 is the defect output module. Detailed Implementation

[0055] The technical solutions in 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.

[0056] See attached document Figure 1 , Figure 1 This is a system architecture diagram according to an embodiment of the present invention. The present invention provides an industrial surface defect pure infrared light scanning detection system, including: an optical acquisition module 10, a frequency domain equalization module 20, an energy modulation module 30, a ratio construction module 40, a matrix decomposition module 50, and a defect output module 60.

[0057] The optical acquisition module 10 is positioned above the workpiece under test and is used to acquire the reflection signal of the first light source, the reflection signal of the second light source, and the dark field noise signal respectively within the constant camera exposure window according to the preset three-phase time-division multiplexing sequence.

[0058] The frequency domain equalization module 20 is connected to the optical acquisition module 10. It receives the reflection signal of the first light source, the reflection signal of the second light source, and the dark field noise signal. Based on the physical distribution parameters of the asymmetric divergence angle in the optical acquisition module 10, it generates a one-dimensional spatially variable Gaussian equalization kernel. The one-dimensional spatially variable Gaussian equalization kernel is used to perform local one-dimensional convolution filtering on the second light source reflection signal after removing the dark field noise signal, and outputs the frequency-aligned second light source reflection signal.

[0059] The energy modulation module 30 is connected to the frequency domain equalization module 20. It extracts the macroscopic geometric gradient of the current scanning line using the reflection signal of the first light source and the reflection signal of the second light source aligned with the frequency. It calculates the light source driving energy of the next scanning line based on the macroscopic geometric gradient and feeds the light source driving energy back to the optical acquisition module 10 to adjust the light emission energy of the light source in the subsequent scanning line.

[0060] The ratio construction module 40 is connected to the optical acquisition module 10 and the frequency domain equalization module 20 respectively. It calculates the time-series thermal drift variance based on the dark field noise signal to generate dynamic sparse penalty parameters. It calculates the spatial reflection differential ratio data using the reflection signal of the first light source and the reflection signal of the second light source aligned with the frequency, and fills the spatial reflection differential ratio data into the sliding buffer row by row to generate a two-dimensional observation matrix.

[0061] The matrix decomposition module 50 is connected to the ratio construction module 40. It receives the two-dimensional observation matrix and the dynamic sparse penalty parameter, and performs Lagrange multiplier iterative decomposition on the two-dimensional observation matrix using the spatial weight matrix of the asymmetric divergence angle mapping and the dynamic sparse penalty parameter. It separates and outputs the low-rank background matrix and the sparse anomaly matrix.

[0062] The defect output module 60 is connected to the matrix decomposition module 50. It receives the sparse anomaly matrix, applies a global hard threshold to the sparse anomaly matrix for binarization, extracts the coordinates of the defect connected components, and outputs the surface defect detection results.

[0063] This invention provides a method for detecting industrial surface defects using pure infrared light scanning, comprising the following steps:

[0064] S11, within the constant camera exposure window, the first light source reflection signal, the second light source reflection signal, and the dark field noise signal are collected respectively according to the preset three-phase time division multiplexing timing sequence;

[0065] S12, a one-dimensional spatially variable Gaussian equalization kernel is generated based on the physical distribution parameters of the asymmetric divergence angle. The one-dimensional spatially variable Gaussian equalization kernel is used to perform local one-dimensional convolution filtering on the second light source reflection signal after removing the dark field noise signal, and the frequency-aligned second light source reflection signal is output.

[0066] S13, extract the macroscopic geometric gradient of the current scan line using the reflection signal of the first light source and the reflection signal of the second light source aligned with the frequency, calculate the light source driving energy of the next scan line based on the macroscopic geometric gradient, and feed back the light source driving energy to adjust the light emission energy of the light source in the subsequent scan lines;

[0067] S14. Calculate the time-series thermal drift variance based on the dark field noise signal to generate dynamic sparse penalty parameters. Calculate the spatial reflection differential ratio data using the reflection signal of the first light source and the frequency-aligned reflection signal of the second light source. Then, fill the spatial reflection differential ratio data into the sliding buffer row by row to generate a two-dimensional observation matrix.

[0068] S15. The spatial weight matrix of the asymmetric divergence angle mapping and the dynamic sparse penalty parameter are used to perform Lagrange multiplier method iterative decomposition on the two-dimensional observation matrix to separate and output the low-rank background matrix and the sparse anomaly matrix.

[0069] S16 applies a global hard threshold to the sparse anomaly matrix for binarization, extracts the coordinates of the defect connected components, and outputs the surface defect detection results.

[0070] In this embodiment, to acquire the microscopic morphological features of the workpiece surface under test, the optical acquisition module 10 includes an imaging receiving unit. Preferably, the imaging receiving unit employs a short-wave infrared array camera, which can effectively suppress interference from visible light colors on the workpiece surface using a specific wavelength band, highlighting purely physical geometric defects. This camera is vertically fixed above the surface of the workpiece under test. A trigger control unit is provided in conjunction with the imaging receiving unit. This trigger control unit consists of a servo encoder and a field-programmable gate array (FPGA). The servo encoder is coaxially connected to the mechanical conveying mechanism carrying the workpiece under test, and outputs spatially equidistant pulse signals to the FPGA according to a fixed physical movement distance. After receiving the spatially equidistant pulse signals, the FPGA outputs a row trigger period signal and a constant camera exposure window time to the short-wave infrared array camera. .

[0071] Camera exposure window time The system maintains rigidity and stability throughout the continuous scanning process. By using encoder signals in combination with a constant exposure window, the physical spatial topological equidistance of the acquired two-dimensional image data in the transmission direction is guaranteed, thus avoiding image geometric stretching and distortion caused by dynamically changing the exposure time.

[0072] To convert minute changes in local surface curvature into directly detectable differences in reflected energy, the optical acquisition module 10 is equipped with an asymmetric optical illumination unit. The asymmetric optical illumination unit consists of a first light source and a second light source arranged parallel to each other on both sides of the camera's scanning field of view.

[0073] The first and second light sources are infrared light-emitting diode arrays. In this embodiment, the light-emitting surfaces of the first and second light sources are equipped with spatially graded optical lenses. Due to the curvature interference of the graded optical lenses, a non-uniform scattering state is generated when the infrared beam illuminates the surface of the workpiece under test. The physical divergence angle mapping function of the first light source on the pixel spatial coordinate axis perpendicular to the direction of movement of the workpiece under test is defined as... The physical divergence angle mapping function of the second light source on the same pixel space coordinate axis is defined as follows: The physical divergence angle is designed to range from 10 degrees to 60 degrees. The mapping function mentioned above can be obtained by fitting the multi-point light intensity distribution measurement using a standard diffuse whiteboard during the production line calibration stage.

[0074] To provide a physical reference for calculating the spatially variable point diffusion function in subsequent data processing, the optical lenses of the first and second light sources are designed with an asymmetric structure, ensuring that the physical divergence angle mapping function satisfies the spatial asymmetry distribution condition. In the formula, Here are the spatial coordinates of the photosensitive pixels of the shortwave infrared array camera, with values ​​ranging from... , This represents the total number of pixels distributed horizontally in the short-wave infrared linear array camera. For the underlying constant current drive circuit and heat dissipation backplane design of the infrared LED array, those skilled in the art can select mature commercial devices based on actual optical power requirements. The circuit topology and structural packaging are well-known technologies in this field and will not be elaborated upon here.

[0075] For a single line scan, the optical acquisition module 10 internally incorporates a time-division multiplexed acquisition unit, and the field-programmable gate array maintains a constant camera exposure window during the single line scan cycle. Internally, it outputs a three-phase time-division multiplexing trigger timing control signal. This trigger timing divides the constant camera exposure window into three consecutive phase intervals, corresponding to parameters as follows: , and The pulse width of the light emission.

[0076] Within the first phase interval, the field-programmable gate array outputs a first drive pulse to illuminate the first light source, simultaneously controlling the short-wave infrared linear array camera to capture the light source reflection signal from the surface of the workpiece being measured. Within the second phase interval, the field-programmable gate array outputs a second driving pulse to illuminate the second light source, controlling the short-wave infrared linear array camera to capture the reflected signal from the second light source. Within the third phase interval, the field-programmable gate array outputs a cutoff signal to simultaneously extinguish the first and second light sources, controlling the short-wave infrared linear array camera to capture dark field noise signals in a dark environment without active illumination. The reflection signals from the first and second light sources, and the dark field noise signal are transmitted and latched separately in independent data buffer registers within the field-programmable gate array (FPGA). To ensure the rigid constraints of spatial sampling, the sum of the durations of the three phase intervals in the three-phase time-division multiplexing triggering sequence must be strictly less than the constant camera exposure window time, i.e., the logical boundary conditions must be met. This prevents pixel-crossing exposure blur caused by high-speed workpiece movement.

[0077] In this embodiment, the dark field noise signal is acquired in real time through the third phase interval. This signal fully includes the thermal drift inside the detector and the ambient background radiation value. Since the subsequent processing module dynamically cancels out the dark field noise signal, the system does not have stringent control requirements for temperature fluctuations and background illuminance in the test environment, and can maintain normal operation in a typical industrial workshop room temperature environment. Regarding the specific high-frequency clock division logic and multi-channel synchronous trigger code implementation inside the field-programmable gate array, those skilled in the art can use hardware description languages ​​for conventional configuration. Its underlying digital logic design is well-known in the field and will not be elaborated upon here.

[0078] In this embodiment, the frequency domain equalization module 20 receives raw three-phase time-division multiplexed data from the optical acquisition module 10, which is used to eliminate the local spatial frequency response differences caused by the asymmetric optical divergence angle of the physical hardware in the low-level computing stage. To implement this data processing, the frequency domain equalization module 20 is internally divided into three logical execution branches: a background denoising unit, an equalization kernel construction unit, and a convolution filtering unit.

[0079] The background noise reduction unit performs background noise cancellation calculations on the first and second light source reflection signals using the dark field noise signal. Specifically, the dark field noise signal at the same photosensitive pixel spatial coordinate position is directly subtracted from the corresponding first and second light source reflection signals, outputting the first light source reflection signal after removing the dark field noise signal. and the second light source reflection signal after removing dark field noise. The corresponding background noise cancellation formulas are as follows:

[0080] ;

[0081] ;

[0082] In the formula, The signal reflected from the first light source. The signal reflected by the second light source. This is a dark field noise signal. For the spatial coordinates of the photosensitive pixel, The coordinates are for the scan row. The above subtraction operation effectively eliminates DC bias interference, thus preserving the reflected energy excited solely by external light source illumination.

[0083] After the aforementioned denoising process, the mathematical calculation problems caused by inherent hardware asymmetry still need to be addressed. Due to the spatial asymmetry in the physical divergence angles of the first and second light sources, the point spread function generated by the same micro-region on the surface of the workpiece under different light source illuminations is not the same, resulting in differences in the edge sharpness of the reflected signal. Without intervention, this inconsistency in spatial frequency response will cause edge halo artifacts or even division-to-zero anomalies in the subsequent ratio calculation stage. Therefore, an equalization kernel construction unit is used to extract the absolute difference in the physical divergence angles of the first and second light sources to construct a one-dimensional spatially variable Gaussian equalization kernel. Specifically, the absolute difference function of the physical divergence angle is calculated. And combined with the system's optical stretching mapping constant Establish the mathematical mapping relationship of local standard deviation Generate a one-dimensional spatially variable Gaussian equalization kernel corresponding to the spatial coordinates of the photosensitive pixel. The specific one-dimensional spatially variable Gaussian equilibrium kernel formula is as follows:

[0084] ;

[0085] In the formula, The relative cell index within the local one-dimensional sliding window is limited to the range of [−K, K]. K is the physical pixel width parameter of one side of the local one-dimensional sliding window. Its value is usually set according to the maximum physical divergence angle difference. As a preferred method, this parameter can be set to an integer between 3 and 15 pixels. Let be the physical divergence angle mapping function of the first light source. Let be the physical divergence angle mapping function for the second light source. Optical broadening mapping constant. The empirical parameters pre-calibrated for the system are determined by calculating the diameter of the standard point light source diffusion spot during the equipment calibration stage, in conjunction with the specific focal length lens used on site. The effective value range is set between 0.5 and 2.5.

[0086] Using the dynamically generated equalization kernel, the convolutional filtering unit performs local one-dimensional convolutional filtering on the second light source reflection signal after removing dark field noise, thereby eliminating edge halo artifacts and outputting a frequency-aligned signal. This process involves sliding a local one-dimensional sliding window pixel-by-pixel along the pixel space coordinate axes, multiplying the signals of each neighboring pixel within the window by the weight value of the equalization kernel at the corresponding relative pixel index position, and then summing the results. The specific local one-dimensional convolutional filtering formula is as follows:

[0087] ;

[0088] In the formula, The second light source reflection signal is aligned with the output frequency. The coordinates of the photosensitive pixel; The coordinates of the scan row; Spatial coordinate neighborhood offset position The second light source reflection signal after removing dark field noise signal, This refers to the one-dimensional spatially variable Gaussian equilibrium kernel weight value corresponding to the relative cell index position.

[0089] To avoid the algorithmic dead zone problem of coordinate out-of-bounds errors in the sliding window at image edges, the system performs one-dimensional boundary extension of the input signal by mirroring edge pixels before performing the convolution operation. With this spatially variable convolution processing mechanism, the reflected signals corresponding to the first and second light sources achieve strict physical alignment in the spatial frequency domain. For the multiply-accumulator calls and pipeline resource allocation within the field-programmable gate array (FPGA) to implement this one-dimensional convolution operation, those skilled in the art can deploy it using existing parallel architectures for digital signal processing. The underlying register allocation and timing control logic are well-known technologies in the field and will not be elaborated upon here.

[0090] In this embodiment, the ratio construction module 40 is mainly responsible for eliminating global emissivity interference and completing the data dimension conversion, thereby providing a reliable two-dimensional input matrix for subsequent manifold optimization algorithms.

[0091] In actual industrial environments, during continuous operation, the dark current of infrared detectors inevitably drifts as the equipment operating temperature increases. Fixed-parameter manifold optimization algorithms struggle to adapt to this dynamic change in background thermal noise, easily leading to the system extracting false defect features. Therefore, the ratio construction module 40 incorporates a thermal drift feedback mechanism, utilizing a depth of... A circular buffer continuously stores historical dark field noise signals and calculates the temporal thermal drift variance of the current scan cycle in real time. Specifically, the system extracts all dark field image data within the buffer cycle and calculates its fluctuation degree from the global statistical mean. The corresponding formula for temporal thermal drift variance is:

[0092] ;

[0093] To avoid insufficient historical data during the initial system startup scan leading to array index issues The algorithm dead zone problem of out-of-bounds access needs to be addressed before execution. Within each scan cycle, the missing historical rows in the circular buffer are represented by the dark field noise signal of the first row. Perform copying and filling. In the formula, The temporal thermal drift variance of the current scan line. The line depth of the circular cache. This represents the total number of pixels in the horizontal direction for a shortwave infrared array camera. This represents the maximum boundary value of the bias index for historical data rows within the circular cache. This represents the maximum index value of the lateral photosensitive pixel spatial coordinates of a shortwave infrared array camera. This represents the dark field noise signal of the historical scan lines. It is the global statistical mean of the dark field noise signal in the spatiotemporal domain of the ring buffer, and its specific calculation method is the arithmetic mean of the values ​​of all dark field pixels in the buffer.

[0094] As a preferred approach, to balance statistical significance and system memory overhead, cache depth... Set to an integer between 100 and 500 rows. After obtaining the time-series thermal drift variance, the system linearly maps it to the dynamic sparsity penalty parameters required by the manifold optimization model. The corresponding mapping formula is:

[0095] ;

[0096] In the formula, Based on the penalty constant, This is the regularization scaling factor. The above two parameters, combined with a standard smoothed sample, are determined by calculating the stable extreme value range of pure background data during the equipment preheating and calibration phase. In typical industrial infrared detection scenarios, the basic penalty constant... The valid value range is between 0.01 and 0.1, and the regularization scaling factor is... The value ranges from 0.5 to 2.0. This closed-loop feedback mechanism enables the system to automatically increase the penalty threshold for subsequent matrix decomposition when the detector's thermal noise increases, ensuring the robustness of the algorithm.

[0097] To eliminate the influence of the absolute emissivity of different material regions on the surface of the workpiece on defect judgment, the ratio construction module 40 performs a division operation using the first light source reflection signal (after removing dark field noise) and the second light source reflection signal (aligned with frequency) to construct spatial reflection difference ratio data. During the calculation, the difference between the two sets of independent light source reflection signals at the same photosensitive pixel spatial coordinate position is used as the numerator, and the sum of the two sets of signals is used as the denominator for normalization. The corresponding spatial reflection difference ratio formula is:

[0098] ;

[0099] In the formula, This is the spatial reflectance differential ratio data output for the current scan line. The first light source reflection signal after removing dark field noise signal, The second light source reflection signal is frequency-aligned. It is a very small positive number, and its value is set to 10. -5The magnitude of this constant is significant. The main purpose of introducing this constant is to avoid algorithmic dead-zone collapse due to division by zero when the infrared reflected energy received by the detector approaches zero. Through the aforementioned ratio construction, the material albedo constant of the workpiece surface is canceled out in the division operation, making the output data sensitive only to local geometric gradient scattering induced by asymmetric light sources.

[0100] Because the manifold optimization algorithm requires the input data to have a two-dimensional matrix topology, while the output of the front-end optical acquisition module 10 is a one-dimensional line scan sequence, the buffer stitching unit further performs cross-dimensional data recombination operations. The system internally allocates a row depth of... The sliding buffer, during continuous operation, will calculate and output single-line spatial reflection differential ratio data line by line. The data is continuously pushed into the sliding buffer in chronological order of scan time. To ensure the integrity of the matrix dimensions, the initial scan... The system maintains data accumulation and temporarily refrains from outputting to subsequent stages. When the number of data rows stacked in the sliding buffer reaches... At that time, a complete two-dimensional observation matrix is ​​generated by splicing the data for subsequent manifold optimization. This two-dimensional observation matrix Spatially it is presented as containing Columns and A set of rows of data. A two-dimensional observation matrix. Once generated, the data will be transferred to the GPU memory resources of the computing core for processing. Simultaneously, the sliding buffer actively discards the oldest historical row data at a preset single-row step size and moves in the most recently acquired scan row data, thereby maintaining the pipeline's continuous and uninterrupted operation.

[0101] As a preferred method, buffer depth Set to the total horizontal pixels of the camera Equal values ​​result in the final generated two-dimensional observation matrix. The structure is square, ensuring a balanced iterative convergence speed for subsequent matrix factorization algorithms across both row and column dimensions. For the direct memory access controller configuration and read / write pointer management involved in the underlying cache concatenation unit, those skilled in the art can deploy it using a standard circular queue data structure. Its underlying memory scheduling logic is well-known in the field and will not be elaborated upon here.

[0102] In this embodiment, the matrix decomposition module 50 is used to extract microscopic defect signals from observation data containing complex backgrounds. At the underlying logical architecture level, the matrix decomposition module 50 is internally divided into a priori weight construction unit, a matrix weighting unit, and a convex optimization iteration unit.

[0103] In practical inspection scenarios, the contrast of defect manifestation in different areas of the workpiece surface varies physically under asymmetric divergence angle illumination. To compensate for and utilize this physical characteristic at the algorithm level, the prior weight construction unit calculates and generates a two-dimensional spatial weight matrix based on the extreme prior distribution of the asymmetric divergence angle of the light source. Specifically, the physical divergence angle mapping functions of the first and second light sources are extracted, the absolute value of their difference is calculated, and normalization is performed. Since the illumination distribution of the workpiece under test remains consistent along the conveying direction, this one-dimensional feature is equivalently extended along the scanning line direction, and the formula for the generated two-dimensional spatial weight matrix is:

[0104] ;

[0105] In the formula, For the generated two-dimensional spatial weight matrix, For the spatial coordinates of the photosensitive pixel, The coordinates of the scan row. Let be the physical divergence angle mapping function of the first light source. Let be the physical divergence angle mapping function of the second light source. This is the maximum scalar value of the absolute difference function across the entire cell coordinate axis. This is the weighting amplification factor.

[0106] As a preferred method, the weight amplification factor The effective value range is set between 0.5 and 2.0, and its specific value is determined by comparing the local signal-to-noise ratio of typical defect areas during the calibration stage before the equipment is put into production.

[0107] After obtaining the two-dimensional spatial weight matrix, the matrix weighting unit fuses it with the two-dimensional observation matrix input from the front end. This fusion process essentially involves performing a Hadamard product operation on the two-dimensional spatial weight matrix and the two-dimensional observation matrix to generate a priori weighted observation matrix. The corresponding Hadamard product formula is:

[0108] ;

[0109] In the formula, For the generated prior weighted observation matrix, The two-dimensional observation matrix output by the ratio construction module 40, The Hadamard product represents the element-wise multiplication of a matrix. It is a two-dimensional spatial weight matrix.

[0110] For the weighted data stream, the convex optimization iterative unit uses the augmented Lagrange multiplier method to perform iterative decomposition on the prior weighted observation matrix. Its physical meaning is to separate the low-rank background matrix that represents normal smooth texture and the sparse anomaly matrix that represents local geometric abrupt changes.

[0111] In this process, the system first receives the single-row dynamic sparsity penalty parameter from the ratio construction module 40. It also calculates the arithmetic mean of the penalty parameters for all scan rows covered by the current two-dimensional observation matrix, and uses this as the global dynamic sparse penalty parameter. Subsequently, the convex optimization iterative unit constructs the augmented Lagrangian objective function, the specific formula of which is:

[0112]

[0113] In the formula, To augment the scalar output of the Lagrange function, The separated low-rank background matrix. The separated sparse outlier matrix, For Lagrange multiplier matrices, This is the iteration step size parameter. Representation matrix nuclear norm number, Representation matrix The L1 norm, Denotes the squared Frobenius norm of the residual matrix. This represents the inner product of matrices. The global dynamic sparsity penalty parameter calculated above is used to adaptively control the sparsity of the sparse matrix during iteration.

[0114] The convex optimization iterative unit continuously updates the matrix using the alternating direction multiplier method. ,matrix and Lagrange multiplier matrix For matrices The update is performed using a soft threshold shrinkage operator for the matrix. The update is performed using the singular value threshold shrinkage operator. To ensure strict convergence of the Lagrange function, the iteration step size parameter... After each iteration, follow the formula Perform monotonically increasing updates. As a preferred approach, the initial iteration step size... Set to 10 -3 Amplification factor It is set to a constant between 1.1 and 1.5. To prevent system failures caused by complex noise interference leading to non-convergence and an infinite loop, the internal logic presets a maximum number of iterations. and residual convergence threshold .

[0115] In this embodiment, the maximum number of iterations Set the residual convergence threshold to 50 to 200 iterations. Set to 10 -7 After each iteration, the system calculates the current reconstruction error. When the reconstruction error is less than the residual convergence threshold Or the current iteration round has reached the maximum number of iterations. When the iteration is terminated, the system forcibly terminates the iteration and outputs the finally converged sparse anomaly matrix. The final output sparse anomaly matrix is ​​a high-purity defect feature map that eliminates the background texture and background noise of the workpiece. For the parallel acceleration calculation logic of the singular value decomposition at the bottom layer of the alternating direction multiplier method, those skilled in the art can use readily available matrix operation algebra libraries for deployment. The underlying matrix basic algebra calculation process is a well-known technology in this field and will not be elaborated here.

[0116] In this embodiment, the defect output module 60 is mainly responsible for converting the pure mathematical matrix output from the matrix decomposition stage into engineering location information with practical physical guidance significance, and transmitting it to external devices. To achieve this conversion process, the defect output module 60 is logically divided into a threshold binarization unit, a coordinate mapping unit, and an output execution unit.

[0117] The sparse anomaly matrix output after the previous convex optimization iteration ideally contains non-zero elements only at defect locations. However, in actual floating-point operations, due to limitations imposed by computer truncation errors and the minimal retention of residual convergence thresholds, slight numerical fluctuations remain in the non-defect regions of the matrix. To eliminate this interference from numerical computation, a threshold binarization unit applies a global hard threshold to the extracted sparse anomaly matrix for binarization, thereby generating a binarized defect mask. Specifically, the system iterates through each element in the sparse anomaly matrix, setting an element to 1 if its absolute value is greater than the set threshold, and 0 otherwise. The corresponding hard threshold determination formula is:

[0118] ;

[0119] In the formula, The pixel values ​​in the output binarized defect mask. This represents the element value at the corresponding coordinate position in the sparse anomaly matrix input from the front end. This is a global hard threshold. As a preferred approach, this global hard threshold... Instead of being a fixed constant, it is adaptively set based on three times the statistical variance of the decomposition residuals of the standard defect-free sample during the equipment preheating stage. Its typical engineering value range is between 0.05 and 0.20. Through the above hard threshold truncation operation, the system generates a binary defect mask containing only 0 and 1, which allows the topological morphology of micro-defects to be accurately stripped away.

[0120] After acquiring the binarized defect mask in the image coordinate system, the coordinate mapping unit performs connected component extraction and inverse mapping calculation of physical coordinates. The system uses an eight-neighbor connectivity algorithm to search for clusters of non-zero pixels in the binarized defect mask, merging spatially adjacent non-zero elements into an independent non-zero element connected component. To avoid malfunctions of external actuators caused by random burst noise, the system performs pixel area statistics on each extracted connected component. When the total number of non-zero pixels contained in a connected component is less than a preset area filtering threshold (e.g., typically set to 3 to 5 consecutive pixels), the system determines it as a false defect and removes it; only connected components greater than or equal to this area filtering threshold are retained to represent real physical defect patches. For the retained valid non-zero element connected components, the system sums and averages the column coordinates and row coordinates of all pixels within it, using this arithmetic mean as the position coordinates of its geometric centroid representative point. To give the image coordinates physical meaning for guiding the operation of industrial robots or removal mechanisms, the coordinate mapping unit maps them inversely to workpiece physical coordinates. The specific physical coordinate mapping formula is as follows:

[0121] ;

[0122] ;

[0123] In the formula, and These are the horizontal and vertical physical coordinates generated after the reverse mapping, respectively, in millimeters. and The coordinates of the physical origin of the workpiece on the conveyor belt or platform; The spatial coordinates of the photosensitive pixel are the geometric centroids of the connected domain. The scan row coordinates are the geometric centroids of the connected domain; For the system's lateral spatial resolution, The vertical spatial resolution of the system. The horizontal spatial resolution. Vertical spatial resolution is statically determined by the camera pixel size and the magnification of the optical lens. It is then calculated by dynamically relating the encoder pulse equivalent of the transmission mechanism with the line scan trigger frequency.

[0124] After completing the physical coordinate system transformation, the output execution unit is responsible for encapsulating and outputting the final pure infrared defect location data on the workpiece surface. The system will then assign the lateral physical coordinates of each defect to the output. With vertical physical coordinates The area size of the defect connectivity region and the corresponding timestamp, among other parameters, are framed and packaged according to a preset industrial bus data structure. The final pure infrared defect location data of the industrial surface, after framing, is sent to an external actuator via an industrial Ethernet interface. Upon receiving the location data, the external actuator performs corresponding defect marking, defective product rejection, or sorting operations according to the on-site process requirements. For the data packetization and parsing process of the industrial Ethernet communication protocol involved in the output actuator, those skilled in the art can deploy it using existing standard protocol stack chips and driver libraries. The underlying network handshake, data link layer verification, and real-time guarantee mechanisms are well-known technologies in the field and will not be elaborated upon here.

[0125] To better understand the technical solution of this invention, the following description is based on a specific application scenario.

[0126] In this embodiment, the target under test is a continuously moving workpiece on a new energy lithium battery positive electrode coating production line. Due to the highly reflective metal foil substrate and rough, dark active material coating on the electrode surface, its surface emissivity fluctuates drastically, and tiny defects such as foil leaks or pinholes are often obscured by the complex background texture. The detection hardware system is deployed at the exit of the coating machine oven, and the camera has a total horizontal pixel count of [missing information]. Configured to 1024, sliding buffer depth The synchronization value is set to 1024, thus constructing a 1024×1024 square observation matrix. The system's lateral spatial resolution... With vertical spatial resolution All are calibrated to 0.02 mm / pixel.

[0127] After the system is powered on and continuous material feeding begins, the detector is inevitably affected by the rise in ambient temperature due to thermal convection at the oven outlet. To accurately capture this thermal drift, the thermal drift assessment unit adjusts the ring buffer depth. The system is set to 200 lines. By continuously extracting these 200 lines of historical dark field noise signals, the system calculates the temporal thermal drift variance of the current scan cycle in real time. Based on the previous calibration results on this type of electrode, the fundamental penalty constant... The regularization scaling factor is set to 0.03. Set to 1.2. Through the dynamic penalty mapping formula. The system generates dynamic sparse penalty parameters that adapt to the current thermal noise level. For example... Figure 3As shown, with the extended operating time of the device and the gradual increase in internal temperature from 25°C to 45°C, the dark current variance of the detector increases, while the penalty parameters calculated and output by this invention... It exhibits accurate linear following characteristics, thereby automatically tightening the sparsity constraints of subsequent matrix decomposition and effectively avoiding misjudging thermal noise exacerbated by temperature rise as a defect.

[0128] After acquiring the spatial reflection differential ratio data that cancels out the absolute emissivity of the material, the system generates a non-overlapping two-dimensional observation matrix with a block size of 1024 rows. For this matrix, a weight amplification factor is introduced into the matrix weighting unit. Two-dimensional spatial weight matrix set to 1.5 Perform the Hadamard product operation to generate the prior weighted observation matrix. This step numerically amplifies the scattering gradient at the defect edge under asymmetric illumination. Subsequently, the convex optimization iterative unit uses the aforementioned averaged global dynamic sparsity penalty parameter. To constrain this, an alternating direction multiplier method iteration is initiated. As a preferred approach, the initial iteration step size is... Set to 10 -3 Amplification factor Set to 1.2, maximum number of iterations The number of iterations is limited to 100. After approximately 40 to 60 iterations, the reconstruction error typically decreases to the residual convergence threshold of 10. -7 The following also triggers the early termination condition. Combined with... Figure 2 The original observation matrix on the left contains a large amount of undulating normal coating texture background and random noise; after decomposition by the Lagrange multiplier method, the sparse anomaly matrix output on the right is greatly suppressed to near zero in most smooth areas, and only sharp numerical pulses are retained at the actual foil defect coordinates.

[0129] To convert this pure mathematical impulse into engineering execution instructions, the system utilizes a global hard threshold. (Set to 0.12 in this embodiment) The sparse anomaly matrix is ​​binarized and truncated, and tiny pseudo-defect connected components with an area less than 4 pixels are filtered out. The final extracted effective defect geometric centroids are then combined with the currently accumulated total offset of the global scan rows. Accurately mapped to absolute physical coordinates at the millimeter level It is then sent to the downstream laser marking and rejection station via the Profinet protocol.

[0130] A controlled comparative experimental environment was established using a constant temperature and humidity test chamber. The specific control parameters of the experimental environment were strictly set as follows: the ambient temperature inside the chamber was uniformly increased from 20℃ to 50℃ within 2 hours to simulate extreme thermal drift; the relative humidity was maintained at a constant 45%; and the circulating air speed inside the chamber was controlled below 0.5 m / s to eliminate interference from mechanical vibration of the workpieces. A conveyor belt ran at a linear speed of 60 m / min, continuously feeding in test sample rolls containing 500 manually calibrated real electrode sheets with minute defects.

[0131] Under the same test samples and experimental environment, the algorithm of this invention and the traditional standard robust principal component analysis algorithm were run respectively. Experimental statistical results show that, under the combined effects of complex emissivity interference and thermal drift, the average signal-to-noise ratio of the traditional algorithm is only 12.4 dB, while the average signal-to-noise ratio of the sparse anomaly matrix output by the algorithm of this invention in the defect region reaches 28.6 dB, demonstrating a substantial improvement in background suppression capability. Furthermore, combined with… Figure 4 The receiver operating characteristic (ROC) curves show that, under a baseline allowing for an extremely low false alarm rate of 0.5%, the traditional algorithm can only achieve a detection rate of 81.2% for minute defects in electrode sheets. This is because the fixed-parameter model generates a large number of spurious noise points at high temperatures, forcing the system to increase the subsequent decision threshold, thus missing real minute defects. In contrast, the algorithm of this invention, due to the prior enhancement of spatial weights at the front end and the integration of thermal drift adaptive feedback in the iterative layer, has an ROC curve that is closer to the ideal state in the upper left corner, and its real defect detection rate climbs to 98.7% under the same 0.5% false alarm rate condition.

Claims

1. A pure infrared light scanning detection system for industrial surface defects, characterized in that, include: The optical acquisition module is used to acquire the reflection signals of the first light source, the reflection signals of the second light source, and the dark field noise signal in a time-division multiplexed sequence. The frequency domain equalization module is used to construct a spatially variable equalization kernel based on the divergence angle distribution, and to use the spatially variable equalization kernel to filter the denoised second light source reflection signal to generate a frequency-aligned second light source reflection signal. An energy modulation module is used to extract a geometric gradient using the reflection signal of the first light source and the frequency-aligned reflection signal of the second light source to adjust the luminous energy. The ratio construction module is used to calculate the thermal drift variance based on the dark field noise signal to generate dynamic sparse penalty parameters, and to construct differential ratio data splicing using the first light source reflection signal and the frequency-aligned second light source reflection signal to generate a two-dimensional observation matrix; The matrix decomposition module is used to decompose the two-dimensional observation matrix using the spatial weight matrix of the divergence angle difference and the dynamic sparsity penalty parameter to output a sparse anomaly matrix. The defect output module is used to binarize and coordinate map the sparse anomaly matrix to output the surface defect detection results.

2. The industrial surface defect pure infrared light scanning detection system according to claim 1, characterized in that, The optical acquisition module includes an imaging receiving unit, a trigger control unit, and an asymmetric optical illumination unit: The imaging receiving unit is vertically positioned above the workpiece being measured. The trigger control unit is coaxially connected to the mechanical transmission mechanism and outputs a trigger cycle signal and a rigid, constant exposure window time to the imaging receiving unit according to the equidistant physical movement distance. The asymmetric optical illumination unit is arranged in parallel on both sides of the field of view of the imaging receiving unit, including a first light source and a second light source with a spatially gradient optical lens mounted on the surface. The first light source and the second light source generate a spatially asymmetric physical divergence angle mapping function on the pixel spatial coordinate axis. The trigger control unit outputs a three-phase time-division multiplexing trigger timing control signal within the exposure window time, respectively driving the first light source to light up individually, the second light source to light up individually, and the two light sources to turn off synchronously, so as to control the imaging receiving unit to generate the first light source reflection signal, the second light source reflection signal, and the dark field noise signal in sequence without cross-boundary blurring.

3. The industrial surface defect pure infrared light scanning detection system according to claim 1, characterized in that, The frequency domain equalization module internally achieves frequency consistency reconstruction through the coordinated operation of a background denoising unit, an equalization kernel construction unit, and a convolutional filtering unit. The background noise reduction unit uses the dark field noise signal at the same spatial coordinate position to directly cancel it out through a subtraction operation and outputs the first light source reflection signal and the second light source reflection signal after removing the dark field noise signal. The equalization kernel construction unit extracts the absolute difference distribution between the first light source and the second light source on the physical divergence angle mapping function, and establishes a linear mapping relationship of local standard deviation in combination with the optical broadening mapping constant of the system, thereby dynamically constructing a one-dimensional spatially variable Gaussian equalization kernel for each pixel coordinate position. The convolutional filtering unit applies the one-dimensional spatially variable Gaussian equalization kernel to the second light source reflection signal after removing the dark field noise signal, and performs a one-dimensional sliding window multiplication and addition operation pixel by pixel using a boundary extension mechanism in the direction of the pixel space coordinate axis, thereby eliminating the edge differences caused by the divergence angle asymmetry and driving the generation of the frequency-aligned second light source reflection signal.

4. The industrial surface defect pure infrared light scanning detection system according to claim 1, characterized in that, The ratio construction module is internally configured with a thermal drift evaluation unit to achieve adaptive adjustment of the penalty threshold: The thermal drift evaluation unit uses a ring cache with a fixed row depth to continuously store the dark field noise signal of the historical scanning cycle, and extracts all dark field image data in the entire cache spatiotemporal domain to calculate the temporal thermal drift variance that deviates from the global statistical mean. The thermal drift evaluation unit further performs a linear combination mapping between the calculated temporal thermal drift variance and the pre-calibrated basic penalty constant and regularization scaling factor, driving the generation of the dynamic sparse penalty parameters used to offset the dark flow artifacts of the high-temperature detector.

5. The industrial surface defect pure infrared light scanning detection system according to claim 1, characterized in that, The ratio construction module completes cross-dimensional reorganization of data through its internal ratio calculation unit and cache splicing unit: The ratio calculation unit uses the difference between the first light source reflection signal after removing the dark field noise signal at the same pixel position and the second light source reflection signal aligned with the frequency as the molecular data. The ratio calculation unit uses the sum of the two sets of signals plus a normal number as the normalized denominator data, and constructs a single row of spatial reflection differential ratio data through division operation, so as to completely cancel the interference of the absolute emissivity constant of the material on the workpiece surface at the mathematical level. The buffer stitching unit opens a sliding buffer inside and pushes the spatial reflection differential ratio data output line by line into the stack in sequence according to the scanning time. When the number of data lines reaches the preset horizontal cell depth, the two-dimensional observation matrix containing the planar topology is stitched together and the clear and reset operation is performed to maintain the data source replacement required for manifold optimization with non-overlapping block step size.

6. The industrial surface defect pure infrared light scanning detection system according to claim 1, characterized in that, The matrix factorization module combines prior weight construction units with matrix weighting units to inject physical prior knowledge into subsequent optimization: The prior weight construction unit extracts the absolute difference between the physical divergence angle mapping functions of the first light source and the second light source and performs maximum value normalization, and then combines the weight amplification coefficient to generate a one-dimensional weight feature. The prior weight construction unit expands and replicates the one-dimensional weight feature by equal value in the scanning row direction, thereby constructing a two-dimensional spatial weight matrix that perfectly matches the dimension of the two-dimensional observation matrix. The matrix weighting unit receives the two-dimensional spatial weight matrix and the two-dimensional observation matrix generated by the front end, and generates a priori weighted observation matrix by performing element-wise Hadamard product operation between the two to numerically amplify and display the geometric scattering gradient of the real defect edge under asymmetric illumination.

7. The industrial surface defect pure infrared light scanning detection system according to claim 6, characterized in that, The matrix factorization module utilizes convex optimization iterative units to remove defects from complex backgrounds: The convex optimization iterative unit calculates the arithmetic mean of the dynamic sparsity penalty parameters within the current matrix block as the global dynamic sparsity penalty parameters, and uses them as sparsity constraints. Combined with the prior weighted observation matrix, it constructs an augmented Lagrange objective function. The convex optimization iterative unit uses the alternating direction multiplier method to iteratively update the low-rank background matrix and the sparse anomaly matrix in the augmented Lagrange objective function, and uses the singular value threshold shrinkage operator to solve the low-rank background matrix, and uses the soft threshold shrinkage operator combined with the global dynamic sparse penalty parameter to solve the sparse anomaly matrix. The convex optimization iterative unit updates the iteration step size parameter monotonically according to the set amplification factor after each iteration. When the residual matrix reaches the preset residual convergence threshold or the rounds are exhausted, the iteration process is blocked, and the sparse anomaly matrix representing pure geometric mutation is separated from the aliased signal and output.

8. The industrial surface defect pure infrared light scanning detection system according to claim 1, characterized in that, The defect output module uses a threshold binarization unit to perform mathematical truncation determination of the morphology: The threshold binarization unit extracts the statistical variance of the decomposition residual of the defect-free sample to adaptively generate a global hard threshold. The threshold binarization unit iterates through the absolute values ​​of each element in the sparse anomaly matrix one by one, sets the coordinates that exceed the global hard threshold to a valid high level state, and otherwise sets them to a logic zero state, thereby converting the floating-point matrix into a binary defect mask that only contains morphological contour features.

9. The industrial surface defect pure infrared light scanning detection system according to claim 8, characterized in that, The defect output module utilizes the coordinate mapping unit to perform a spatial reconstruction process from relative pixels to global physical coordinates: The coordinate mapping unit uses the eight-neighborhood connectivity algorithm to aggregate non-zero element connected regions in the binarized defect mask, removes pseudo-noise patches with pixel areas smaller than a preset area filtering threshold, and calculates the arithmetic mean of the row and column indices of each pixel for the effective non-zero element connected regions to generate the relative geometric centroid position coordinates. The coordinate mapping unit obtains the total offset parameter of the global scan row recorded synchronously, and adds it to the relative row index in the relative geometric centroid position coordinates to transform it into an absolute row coordinate with global uniqueness. The coordinate mapping unit integrates the horizontal spatial resolution, vertical pulse resolution, and the coordinate parameters of the initially calibrated physical origin on the workpiece of the system, and projects the absolute pixel index system in reverse to the physical coordinate system of the workpiece through linear transformation, thereby generating continuous horizontal and vertical physical coordinates without any breaks.

10. The industrial surface defect pure infrared light scanning detection system according to claim 9, characterized in that, The defect output module is equipped with an output execution unit at its end to enable industrial interconnection execution. The output execution unit will convert the generated horizontal physical coordinates, the vertical physical coordinates, the calculated physical area of ​​the connected domain with effective non-zero elements, and the corresponding timestamp parameters, and perform frame packing operations according to the preset industrial bus data structure. The packaged industrial location guidance data is verified by the underlying data link layer and then pushed to the downstream external actuators through the industrial Ethernet interface to directly drive the external actuators to perform defective product rejection or laser marking processes with real physical location significance.

Images