Method and system for detecting industrial fa lenses based on visual detection

Abstract

This invention discloses a vision-based inspection method and system for industrial FA lenses. The invention relates to the field of vision inspection technology. It determines the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, and extracts multiple early defect features of the industrial FA lens in different dimensions along the frequency domain. Simultaneously, it determines multiple nonlinear distortion information based on the inverse decoupling of each current image, and constructs a corresponding distortion gradient field along these multiple nonlinear distortion information, improving the accuracy of the distortion gradient field of the industrial FA lens. Furthermore, it maps this multi-source heterogeneous data to a unified feature manifold space, thereby determining the optical attenuation trend of the industrial FA lens across the entire field of view and marking the corresponding lens drift data. Further, it combines a memory network and previous inspection data of the industrial FA lens to determine a dynamic inspection system for the industrial FA lens, improving the accuracy of the dynamic inspection system.

Description

Technical Field

[0001] This invention relates to the technical field of visual inspection, and more particularly to a method and system for inspecting industrial FA lenses based on visual inspection. Background Technology

[0002] As the core imaging component of machine vision systems, the industrial FA (Factory Automation Lens) directly determines the detection accuracy and operational stability of automated production lines. With the development of Industry 4.0 and precision manufacturing technologies, the market has placed extremely high demands on the surface quality, internal medium uniformity, and coating spectral characteristics of industrial FA lenses. Therefore, high-precision visual inspection and quality control are crucial during the lens manufacturing and usage processes.

[0003] Existing technologies for visual inspection of industrial FA lenses typically employ a single light source or a fixed illumination mode, often limited to single low-angle or high-angle illumination. This makes it impossible to integrate multi-dimensional information from low-angle dark fields, high-angle bright fields, and narrow-band light sources within the same inspection process. This single optical excitation method makes it difficult to effectively detect early minute defects located on or inside the lens surface, and it is impossible to further control multiple early defect features. This affects the accuracy of the distortion gradient field of industrial FA lenses, resulting in low accuracy of the dynamic inspection system for industrial FA lenses. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and system for detecting industrial FA lenses based on visual inspection.

[0005] This invention provides a vision-based inspection method for industrial FA lenses, comprising:

[0006] The industrial FA lens is transported to the vision inspection station, and the LED multi-zone matrix light source of the vision inspection station is adjusted to low-angle dark field, high-angle bright field and narrow-band light source, so as to collect multiple current images of the industrial FA lens in the multi-dimensional optical scene during the vision inspection of the industrial FA lens.

[0007] The frequency domain of the industrial FA lens is determined based on the dynamic transformation of multiple current images, and multiple early defect features of the industrial FA lens in different dimensions are extracted along the frequency domain. At the same time, multiple nonlinear distortion information is determined based on the inverse decoupling of each current image, and the corresponding distortion gradient field is constructed along the multiple nonlinear distortion information.

[0008] The early defect features and corresponding distortion gradient fields of the industrial FA lens are cross-fused to output digital twin data of the industrial FA lens in the current application scenario. The digital twin data is input into a preset optical performance mapping network and multiple topological nodes of the industrial FA lens are dynamically divided. The corresponding multi-source heterogeneous data is determined based on the tracing of each topological node.

[0009] The multi-source heterogeneous data is mapped to a unified feature manifold space to determine the optical attenuation trend of industrial FA lenses across the entire field of view and to mark the corresponding lens drift data. Furthermore, by combining the memory network and previous test data of industrial FA lenses, a dynamic testing system for industrial FA lenses is determined. This dynamic testing system can be adapted to the dynamic testing of different batches of industrial FA lenses.

[0010] This invention provides a vision-based inspection system for industrial FA lenses, which is applied to the aforementioned vision-based inspection method for industrial FA lenses. The vision-based inspection system for industrial FA lenses includes:

[0011] The vision inspection module is used to transport the industrial FA lens to the vision inspection station and adjust the LED multi-zone matrix light source of the vision inspection station to low-angle dark field, high-angle bright field and narrow-band light source, so as to acquire multiple current images of the industrial FA lens in a multi-dimensional optical scene during the vision inspection of the industrial FA lens.

[0012] The early defect feature module is used to determine the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, and extract multiple early defect features of the industrial FA lens in different dimensions along the frequency domain. At the same time, it determines multiple nonlinear distortion information based on the inverse decoupling of each current image, and constructs the corresponding distortion gradient field along the multiple nonlinear distortion information.

[0013] The multi-source heterogeneous data module is used to cross-fuse multiple early defect features and corresponding distortion gradient fields of industrial FA lenses, thereby outputting digital twin data of industrial FA lenses in the current application scenario. This digital twin data is input into a preset optical performance mapping network, and multiple topological nodes of industrial FA lenses are dynamically divided. The corresponding multi-source heterogeneous data is determined based on the tracing of each topological node.

[0014] The dynamic detection system module is used to map the multi-source heterogeneous data into a unified feature manifold space, thereby determining the optical attenuation trend of the industrial FA lens in the full field of view and marking the corresponding lens drift data. Furthermore, by combining the memory network and the previous detection data of the industrial FA lens, the dynamic detection system of the industrial FA lens is determined. This dynamic detection system can be adapted to the dynamic detection of different batches of industrial FA lenses.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] (1) The industrial FA lens is transported to the vision inspection station, and the LED multi-zone matrix light source of the vision inspection station is adjusted to low-angle dark field, high-angle bright field and narrow-band light source, so as to collect multiple current images of the industrial FA lens in the multi-dimensional optical scene during the vision inspection of the industrial FA lens; the frequency domain of the industrial FA lens is determined according to the dynamic transformation of multiple current images, and multiple early defect features of the industrial FA lens in different dimensions are extracted along the frequency domain. At the same time, multiple nonlinear distortion information is determined based on the inverse decoupling of each current image, and the corresponding distortion gradient field is constructed along the multiple nonlinear distortion information. The vision inspection of the industrial FA lens is introduced, and multiple early defect features are further controlled, which improves the accuracy of the distortion gradient field of the industrial FA lens.

[0017] (2) Cross-fusion of multiple early defect features and corresponding distortion gradient fields of industrial FA lens to output digital twin data of industrial FA lens in the current application scenario. Input the digital twin data into the preset optical performance mapping network and dynamically divide multiple topological nodes of industrial FA lens. Based on the tracing of each topological node, determine the corresponding multi-source heterogeneous data. Map the multi-source heterogeneous data to a unified feature manifold space to determine the optical attenuation trend of industrial FA lens in the full field of view and mark the corresponding lens drift data. Further combine the memory network and the previous test data of industrial FA lens to determine the dynamic detection system of industrial FA lens. Further control of multiple topological nodes of industrial FA lens, fully consider lens drift data, memory network and previous test data of industrial FA lens, and improve the accuracy of dynamic detection system of industrial FA lens. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the detection method for industrial FA lenses based on vision inspection in an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating step S11 of the detection method for industrial FA lenses based on vision detection in this embodiment of the invention.

[0020] Figure 3 This is a flowchart illustrating step S12 in the vision-based detection method for industrial FA lenses in this embodiment of the invention.

[0021] Figure 4 This is a flowchart illustrating step S13 in the vision-based detection method for industrial FA lenses in this embodiment of the invention.

[0022] Figure 5 This is a flowchart illustrating step S14 of the vision-based detection method for industrial FA lenses in this embodiment of the invention.

[0023] Figure 6 This is a schematic diagram of the structural composition of the vision-based industrial FA lens detection system in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] Please see Figures 1 to 6 A vision-based inspection method for industrial FA lenses is proposed, applied to vision inspection scenarios. The vision-based inspection method for industrial FA lenses includes:

[0026] Step S11: Transport the industrial FA lens to the vision inspection station, and adjust the LED multi-zone matrix light source of the vision inspection station to low-angle dark field, high-angle bright field and narrow-band light source, so as to acquire multiple current images of the industrial FA lens in the multi-dimensional optical scene during the vision inspection of the industrial FA lens.

[0027] Step S12: Determine the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, and extract multiple early defect features of the industrial FA lens in different dimensions along the frequency domain. At the same time, determine multiple nonlinear distortion information based on the inverse decoupling of each current image, and construct the corresponding distortion gradient field along the multiple nonlinear distortion information.

[0028] Step S13: Cross-fusion of multiple early defect features and corresponding distortion gradient fields of industrial FA lens to output digital twin data of industrial FA lens in the current use scenario. Input the digital twin data into the preset optical performance mapping network and dynamically divide multiple topology nodes of industrial FA lens. Determine the corresponding multi-source heterogeneous data based on the tracing of each topology node.

[0029] Step S14: Map the multi-source heterogeneous data to a unified feature manifold space to determine the optical attenuation trend of the industrial FA lens in the full field of view, and mark the corresponding lens drift data. Further combine the memory network and the previous test data of the industrial FA lens to determine the dynamic detection system of the industrial FA lens. This dynamic detection system can be adapted to the dynamic detection of different batches of industrial FA lenses.

[0030] refer to Figure 2 In step S11, the specific steps are as follows:

[0031] S111: When the industrial FA lens is transported to the vision inspection station, the LED multi-zone matrix light source at the vision inspection station responds to the arrival signal of the industrial FA lens. The LED multi-zone matrix light source adaptively flickers along the time sequence according to the signal of the industrial FA lens and the current environment of the vision inspection station, switching between low-angle dark field, high-angle bright field and narrowband light source. At this time, the low-angle dark field excites the scattered light on the surface of the FA lens with a grazing angle; the high-angle bright field penetrates the lens body and captures the transmission pattern of impurities; and the narrowband light source excites the interference fringe response of the multilayer coating under a specific spectrum through a bandpass filter with a preset center wavelength.

[0032] S112: It integrates low-angle dark field, high-angle bright field and narrowband light source into a multi-dimensional optical scene, and performs visual inspection on industrial FA lens. During the visual inspection process, it combines the multi-dimensional optical scene to determine multiple current images of the lens. Simultaneously, it completes the pre-filtering of the surface morphology, internal medium and coating spectral characteristics of industrial FA lens in the visual inspection end of industrial FA lens.

[0033] In the embodiments of this application, when the industrial FA lens is transported to the vision inspection station, the LED multi-zone matrix light source at the vision inspection station responds to the arrival signal of the industrial FA lens. The LED multi-zone matrix light source adaptively flickers along the time sequence according to the signal of the industrial FA lens and the current environment of the vision inspection station, switching between low-angle dark field, high-angle bright field and narrowband light source. At this time, the low-angle dark field excites the scattered light on the surface of the FA lens with a grazing angle; the high-angle bright field penetrates the lens body and captures the transmission pattern of impurities; and the narrowband light source, through a preset center wavelength bandpass filter, excites the interference fringe response of the multilayer coating under a specific spectrum, thus introducing the interference fringe response under a specific spectrum.

[0034] At this moment, the instant the industrial FA lens is precisely positioned to the vision inspection station by the conveying mechanism, the photoelectric position sensor located on the side of the station immediately captures the lens's arrival signal and sends a trigger pulse to the system. At this time, the system does not blindly turn on the light source, but simultaneously reads the data from the ambient light intensity sensor integrated inside the station to obtain the background illuminance in the current inspection dark box. Based on the preset inspection protocol of the lens model (such as a type A industrial FA lens) and the current ambient light data, the system calculates and generates the initial driving current and timing control command of the LED multi-zone matrix light source, thereby establishing a closed-loop response link of "signal-environment-light source" to ensure the stability and anti-interference capability of subsequent flicker switching.

[0035] Based on the generated timing instructions, the system controls the driver of the LED multi-zone matrix light source to switch at high frequency on a millisecond-level time axis. This switching is not a simple on / off switch, but an adaptive adjustment based on the optical characteristics of the A-type industrial FA lens: the system activates the peripheral low-angle light-emitting unit, the top high-angle light-emitting unit, and the central narrowband spectral unit in sequence according to the preset detection priority. During the switching process, the system monitors the actual output power of the light source in real time, eliminates brightness fluctuations caused by the thermal effect of the driving circuit, and ensures the uniformity and consistency of light energy output in the three modes of low-angle dark field, high-angle bright field, and narrowband light source, thereby constructing three independent and complementary optical imaging windows in a very short time.

[0036] When the sequence enters the low-angle dark field stage, the outer ring area of ​​the LED multi-zone matrix light source is lit, and the light is incident on the surface of the A-type industrial FA lens at a pre-set grazing angle (usually less than 30 degrees). Since this grazing angle is less than the critical angle of total internal reflection of the smooth surface of the lens, most of the light is specularly reflected by the lens surface and overflows the field of view, unable to enter the camera lens, so the background presents an extremely dark field effect. However, if there are micron-level scratches or chipped edges on the surface of the A-type industrial FA lens, these micro-defects will break the specular reflection conditions and produce strong diffuse reflection (scattered light). These scattered lights form high-contrast bright spot signals against the dark background, thus successfully highlighting the tiny physical damage characteristics of the lens surface.

[0037] When the light source mode is switched to high-angle bright field, the vertical illumination unit located at the top of the light source matrix is ​​turned on, and the light penetrates the lens body of the A-type industrial FA lens at a large angle (nearly vertical incidence). Under this illumination condition, the light mainly travels along the optical axis and enters the imaging system, and the lens body presents a bright and uniform bright field background. If there are abnormal media such as bubbles, cement layer impurities, or fingerprints inside the A-type industrial FA lens, these foreign objects will absorb, refract, or block the irradiated light, forming obvious shadows or light spots in the bright background. In this way, the system accurately captures the transmission morphology defects of the deep media hidden inside the lens.

[0038] In the final stage of the timing sequence, a narrowband light source is activated. This light source undergoes specific bandpass filtering to output monochromatic light with only the center wavelength (such as 525nm green light or a specific blue light). The surface of an A-type industrial FA lens is typically coated with multiple layers of antireflective coating. When narrowband monochromatic light shines on the coated surface, the light wave undergoes multi-beam interference on the upper and lower surfaces of the thin film. If the coating layer is of uniform thickness and of acceptable quality, the interference fringes will exhibit a specific uniform distribution or uniform color reflection. If the coating has uneven thickness, delamination, or oxidation defects, the interference optical path difference will change, causing the interference fringes to become distorted, broken, or exhibit abnormal mottled colors. By exciting this specific interference fringe response, the system can sensitively capture coating spectral characteristic defects that are difficult to detect with the naked eye.

[0039] Specifically, for the Type A industrial FA lens (specifications: focal length 12mm, F1.4 aperture, multi-layer coating), when the Type A industrial FA lens enters the inspection station with the conveyor belt, the photoelectric sensor senses its position, and after the system confirms that the ambient light intensity of the dark box meets the standard, the LED multi-zone matrix light source is activated; the system illuminates the low-angle ring light source, illuminating the front lens at a 25-degree grazing angle; at this time, the background of the image captured by the camera is completely black, but a faint bright spot is found on the edge of the lens, which is determined to be a tiny surface scratch that is not visible under normal lighting.

[0040] The light source was switched to a high-angle bright field mode, with light penetrating the lens group perpendicularly. Against a bright imaging background, the camera clearly captured a tiny bubble with a diameter of approximately 0.05 mm in the lens adhesive layer. This bubble appeared as a black shadow dot, indicating a packaging defect in the internal medium. Further, the system illuminated a narrowband blue light source with a center wavelength of 450 nm to shine on the antireflective coating layer on the surface of the Type A industrial FA lens. Regular interference rings were observed in the image captured by the camera. However, in the lower right corner of the lens, the interference fringes showed obvious local breaks and distortions, directly reflecting a deviation in the coating thickness in this area (uneven coating), which is a functional optical defect.

[0041] Furthermore, low-angle dark field, high-angle bright field, and narrowband light source are fused together to construct a corresponding multi-dimensional optical scene. At the same time, the industrial FA lens is visually inspected, and multiple current images of the lens are determined in combination with this multi-dimensional optical scene during the visual inspection process. Simultaneously, the surface morphology, internal medium, and coating spectral characteristics of the industrial FA lens are pre-filtered at the visual inspection end of the industrial FA lens. This is compatible with the overall consideration of combining this multi-dimensional optical scene during the visual inspection process, ensuring the accuracy of multiple current images of the lens.

[0042] At this point, after acquiring the time-series image in step S111, the system extracts optical physical quantities under each light source mode, including the grazing angle azimuth vector of the low-angle dark field, the light intensity transmittance distribution of the high-angle bright field, and the center wavelength and bandwidth characteristics of the narrowband light source. The system uses coordinate system transformation to fuse the above heterogeneous optical parameters in a unified spatial reference system to construct a multi-dimensional optical scene model containing a three-dimensional vector of "spatial angle-light intensity distribution-spectral response". This model completely reproduces the theoretical response state of the A-type industrial FA lens under different physical illumination excitations in the digital domain, providing a standardized optical reference surface for subsequent feature extraction.

[0043] Based on the constructed multidimensional optical scene, the visual inspection system maps the discrete sequence images acquired by the camera into the scene model. The system eliminates the micron-level mechanical displacement deviation of the lens during multi-frequency switching through feature registration, and strictly aligns the scattering image of the dark field, the transmission image of the bright field, and the interference image of the narrow band at the pixel level. The "multiple current images" determined in this way are no longer isolated gray-level matrices, but composite datasets carrying multidimensional optical attribute labels, which accurately characterize the actual physical state of the A-type industrial FA lens in the full field of view, ensuring the consistency of defect coordinates in subsequent analysis.

[0044] For dark-field image data, the visual inspection end performs a pre-filtering method based on morphological gradients. Since low-angle grazing light is very likely to form non-defective structural reflections at the lens edges or chamfers, the system uses geometric masks in the multi-dimensional optical scene to remove the bright interference areas of the lens mechanical shell and edge chamfers. At the same time, the system calculates the scattering intensity entropy of local areas of the image, sets a dynamic threshold to filter out low signal-to-noise ratio speckle caused by environmental dust, and retains only the scattering signals with high gradient changes that are located within the effective field of view, thereby accurately separating the feature areas that represent micro-morphological defects such as surface scratches and edge chipping.

[0045] For bright field image data, the system implements a pre-filtering strategy based on transmittance uniformity analysis. Given that the edges of the cemented layer inside the FA lens often produce non-defective dark rings or halos due to light refraction, the system uses the optical path propagation model in the optical scene to identify and remove these inherent optical artifacts. The system further analyzes the grayscale histogram distribution of the image, filters out large areas of uniformly transmitted background (i.e., qualified areas), and focuses on locking heterogeneous areas where transmittance changes abruptly. This effectively distinguishes and retains the projected shadows formed by internal medium defects such as bubbles and impurities, and eliminates artifact interference caused by stress concentration.

[0046] For narrowband light source images, the visual inspection end executes a pre-filtering mechanism based on frequency domain texture analysis. The system uses Gabor filter banks to perform multi-directional scanning and frequency response analysis on the interference fringe texture of the lens surface. Since a qualified coating layer will produce continuous and uniform interference fringes, while surface contaminants such as fingerprints and oil stains affect the appearance but do not change the physical thickness of the coating layer, they will not destroy the interference structure. Based on this principle, the system filters out non-coating layer contaminant signals that do not have spectral interference characteristics, and focuses on marking the areas where the interference fringes are broken, twisted, or have wavelength drift, thereby accurately locating deep defects that affect optical performance, such as coating peeling and uneven thickness.

[0047] Specifically, at the inspection station, the system fuses the three original images of the Type A industrial FA lens acquired in step S111 to construct a dedicated multi-dimensional optical scene for the 12mm focal length lens. This scene accurately covers the optical path characteristics under the F1.4 large aperture structure of the Type A industrial FA lens and the spectral response benchmark of the multi-layer coating. In the surface morphology pre-filtering stage, the system automatically identifies and blocks the high-brightness reflective interference of the metal outer ring of the Type A industrial FA lens. At the same time, it determines that a small bright spot in the upper left corner of the image is floating dust (low scattering entropy) on the surface and filters it out. However, a high-contrast scratch signal located in the center of the lens is retained and confirmed as a surface defect.

[0048] In the internal media pre-filtering stage, for bright field images, the system detected a faint dark ring at the edge of the lens cement layer. The system, combined with the optical scene model, determined that it was an inherent cement edge diffraction effect (artifact) of this lens model and filtered it out. Meanwhile, a well-defined circular dark spot located on the lower right side of the lens was identified as an internal air bubble and retained.

[0049] In the pre-filtering stage of coating spectral characteristics, for narrowband blue light images, the system analysis found a blurry oil stain on the lens surface. However, since the interference fringes in this area were still continuous and regular, the system determined that it was only surface contamination and would not affect the coating spectral characteristics, so it was removed through pre-filtering. Conversely, obvious breaks and abnormal wavelength reflections were detected in the interference fringes at the edge of the lens. The system confirmed that the coating thickness was uneven at this point and marked it as a functional optical defect. Through this series of pre-filtering steps, the signal-to-noise ratio of the test data of the A-type industrial FA lens was greatly improved, providing clean source data for the frequency domain analysis in the subsequent S12 step.

[0050] refer to Figure 3 In step S12, the specific steps are as follows:

[0051] S121: Sort multiple current images along the time domain to form a corresponding current image sequence. Perform a three-dimensional Fourier transform on the current image sequence and combine it with a wavelet packet decomposition mechanism for dynamic transformation, thereby mapping the current image sequence to the frequency domain space. In this frequency domain space, use radial frequency slices to extract multiple early defect features corresponding to different defect scales.

[0052] S122: Synchronously mark each current image, reverse decouple each current image, and dynamically calibrate the image offset content of the target object. Thus, determine multiple nonlinear distortion information by dynamically tracking the image offset content. The multiple nonlinear distortion information covers lens eccentricity information, tilt deformation information, and surface error information. At the same time, using the pixel neighborhood as the topological unit, construct a distortion gradient field with continuous vector distribution characteristics along the multiple nonlinear distortion information. This gradient field can map the deflection state and aberration evolution path of the industrial FA lens in an intuitive flow field shape.

[0053] In the embodiments of this application, multiple current images are sorted along the time domain to form a corresponding current image sequence. The current image sequence is subjected to a three-dimensional Fourier transform and dynamically transformed by combining a wavelet packet decomposition mechanism, thereby mapping the current image sequence to the frequency domain space. In this frequency domain space, multiple early defect features corresponding to different defect scales are extracted using radial frequency slicing. This introduces the extraction of multiple early defect features corresponding to different defect scales in the frequency domain space using radial frequency slicing.

[0054] At this point, after completing the pre-filtering in step S112, the system logically sorts the three current images—dark field, bright field, and narrowband—along a strict time axis according to the triggering sequence of the LED multi-zone matrix light source in step S111, constructing an ordered sequence of current images. This sequence is not a simple set, but rather establishes a correspondence between "time-illumination mode-image features," ensuring that subsequent processing can track the image evolution under specific illumination conditions, providing structured input data for three-dimensional frequency domain analysis.

[0055] The system treats the constructed current image sequence as a three-dimensional data volume (xy spatial dimension, t time / sequence dimension) and performs a three-dimensional fast Fourier transform (3D-FFT) on it. Through the transform, the spatial gray-level distribution information of the image is projected into the frequency domain space, where the low-frequency component represents the background contour of the lens and the overall illumination distribution, and the high-frequency component represents the edges, textures and noise. The system introduces a wavelet packet decomposition mechanism to dynamically transform the frequency domain signal. By utilizing the multi-resolution analysis characteristics of wavelet packets, the frequency domain signal is decomposed into subspaces of different frequency bands, thereby achieving fine decoupling of spectral energy and effectively distinguishing high-frequency random noise from defect signals with specific spatial structures.

[0056] Within the generated frequency domain space, the system performs radial frequency slicing operations along different radii with the center of the spectrum as the origin. Since different defect types (such as point, line, and area) correspond to different energy accumulation regions in the frequency domain, the system extracts feature components corresponding to different spatial frequencies by setting ring bandpass filters with different radius ranges. By calculating the energy spectral density and spectral texture features of each slice region, the system can accurately capture high-frequency protrusion signals representing early micro-defects, thereby extracting multiple early defect features corresponding to different defect scales and realizing frequency domain quantitative characterization of micro-hidden dangers.

[0057] Specifically, the system arranges the filtered dark field, bright field, and narrowband images according to the lighting trigger order to construct the current image sequence of the A-type industrial FA lens; the system performs a three-dimensional Fourier transform on the sequence to map the spatial image data of the A-type industrial FA lens to the frequency domain; in the spectrum diagram, the normal lens edge and the F1.4 large aperture structure form a low-frequency main energy ring, while the background noise appears as discrete stray high-frequency points.

[0058] To extract early defects, the system uses wavelet packet decomposition to refine the frequency domain signal and performs radial frequency slicing. In the mid-to-high frequency slice (corresponding to a spatial frequency of 0.5-1.0 cycles / mm), the system detects an abnormal peak in the energy spectral density, which corresponds to an early microcrack on the surface of the Type A industrial FA lens with a depth of only micrometers (this crack is extremely inconspicuous in the spatial domain, but causes high-frequency resonance in a specific direction in the frequency domain). In the high-frequency slice at a specific angle, the system identifies another set of directional high-frequency energy accumulations, which are determined to be minute micro-stripes defects in the lens coating layer. Through this radial frequency slicing technique, the system successfully separates and extracts the early microcrack and micro-stripe features of the Type A industrial FA lens in the frequency domain, achieving precise targeting of early defects that are difficult to detect with the naked eye.

[0059] Furthermore, each current image is simultaneously labeled, and the images are decoupled in reverse. The image offset content of the target object is dynamically calibrated, thereby determining multiple nonlinear distortion information by dynamically tracking the image offset content. These nonlinear distortion information include lens eccentricity information, tilt deformation information, and surface error information. Simultaneously, using pixel neighborhoods as topological units, a distortion gradient field with continuous vector distribution characteristics is constructed along the multiple nonlinear distortion information. This gradient field can map the deflection state and aberration evolution path of the industrial FA lens in an intuitive flow field shape, taking into account the overall consideration of multiple nonlinear distortion information and ensuring the accuracy of the distortion gradient field with continuous vector distribution characteristics. At the same time, visual inspection of the industrial FA lens is introduced to further control multiple early defect features, thereby improving the accuracy of the distortion gradient field of the industrial FA lens.

[0060] At this point, the system synchronously marks the multiple current images output in step S112 and performs differential comparison with the preset standard FA lens ideal imaging model. Using a reverse decoupling method, the pixels in the current image are traced back along the optical path to the theoretical imaging plane, removing parallax artifacts caused by changes in the light source angle. During this process, the system dynamically calibrates the image offset of the target object (i.e., the actual image of the FA lens) relative to the ideal model. This offset is not a simple translation, but a non-uniform displacement field that includes local geometric deformation, accurately recording the degree of deviation of each local area from the standard position.

[0061] Based on the calibrated image offset content, the system employs optical flow field tracking technology to dynamically track the offset vector along its direction and modulus, transforming discrete displacement data into continuous deformation parameters. This allows for the calculation of multiple nonlinear distortion information, which not only includes conventional radial distortion coefficients but also deeply encompasses specific process defect characteristics: lens eccentricity information manifests as an asymmetric offset vector of the image's geometric center relative to the optical axis; tilt deformation information is reflected as a trapezoidal distortion feature with progressive stretching or compression on one side of the image; and surface error information is mapped as irregular twisting in local areas, reflecting local deviations in the curvature of the lens surface.

[0062] After acquiring nonlinear distortion information, the system defines the surrounding area of ​​each pixel in the image as a pixel neighborhood topological unit. The system calculates the difference vector of distortion (including offset and direction) between the pixel and its neighboring pixels, and integrates and concatenates these difference vectors over the entire domain to construct a distortion gradient field with continuous vector distribution characteristics. This gradient field no longer describes the distortion with isolated values, but forms a visualized flow field shape, where the magnitude of the vector represents the severity of the distortion, and the direction of the vector points to the source or direction of the distortion.

[0063] The system utilizes the constructed distortion gradient field to physically map the optical state of industrial FA lenses. The convergence points (source points) or divergence points (sink points) of streamlines in the gradient field intuitively reflect the stress or assembly abnormality areas inside the lens. By analyzing the overall topology of the flow field, the system can clearly map the deflection trend of the FA lens (such as the overall tilt direction of the lens) and the aberration evolution path (such as how local aberrations spread to the edge), providing an intuitive physical basis for determining whether there is assembly stress concentration or optical axis misalignment in the lens.

[0064] Specifically, the system reverse-decouples the dark field, bright field, and narrowband images of the Type A industrial FA lens and finds that the central crosshair of the bright field image has a slight counterclockwise rotation offset relative to the standard position, and there is non-uniform stretching in the edge area. By dynamically tracking these offsets, the system calculates that the Type A industrial FA lens has a lens eccentricity of 0.05mm (causing the image center to deviate from the optical axis) and a slight tilt distortion (indicating the wedge angle caused by uneven shims during lens assembly).

[0065] The system constructs a distortion gradient field using pixel neighborhoods as units. In the gradient field streamline diagram, obvious vector aggregation and divergence are observed in the lower right corner of the lens, which in turn indicates that there is a significant surface error in this area. This is due to excessive locking force during the assembly of the F1.4 large aperture structure, which causes local stress compression deformation on the lens surface. This distortion gradient field not only quantifies the degree of distortion, but also directly reveals the evolution path of aberrations spreading from the lower right corner to the center.

[0066] refer to Figure 4 In step S13, the specific steps are as follows:

[0067] S131: Collect multiple early defect features of industrial FA lens, input multiple early defect features and corresponding distortion gradient fields into cross-fusion module, and perform cross-matching of the physical location of early defect features and the mechanical distribution of distortion gradient field, thereby dynamically outputting digital twin data of industrial FA lens in the current use scenario during the matching process. This digital twin data contains full-dimensional attributes of "morphology-distortion-optics".

[0068] S132: Obtain a preset optical performance mapping network. After loading the twin data into the preset optical performance mapping network, the optical performance mapping network dynamically divides multiple topological nodes representing different optical performance ranges of industrial FA lenses according to the density of aberration distribution and field of view attenuation boundary. At this time, the topological nodes are used as anchor points and traced backward to extract the node features of the current frame. It can also cross-modal associate the context information corresponding to the topological node. In the process of aggregating context information, multi-source heterogeneous data with causal relationship is formed. This context information covers the material information of industrial FA lenses, the corresponding lens processing route, and the historical environment.

[0069] In the embodiments of this application, multiple early defect features of an industrial FA lens are collected, and the multiple early defect features and the corresponding distortion gradient field are input into a cross-fusion module. The physical location of the early defect features and the mechanical distribution of the distortion gradient field are cross-matched, thereby dynamically outputting digital twin data of the industrial FA lens in the current usage scenario during the matching process. The digital twin data contains full-dimensional attributes of "morphology-distortion-optics".

[0070] At this point, the system extracts discrete early defect feature data from the frequency domain analysis results of step S12, and extracts continuous distortion gradient field data from the distortion analysis results. These two types of data are heterogeneous: early defect features are manifested as energy abrupt changes or texture anomalies at specific coordinate points, while distortion gradient fields are manifested as a continuous vector distribution across the entire domain. The system synchronously inputs these two types of data into the preset cross-fusion module and performs data format normalization at the input interface to ensure that the point features of defects and the field features of distortion are computable in the spatial coordinate system.

[0071] Within the fusion module, the system executes a core cross-matching method. This method is not a simple superposition, but a logical association based on physical causes: the system maps the physical location coordinates of early defect features to the distortion gradient field, queries the vector magnitude and direction (i.e., mechanical distribution) at that location, and calculates the geometric relationship between the defect feature point and the local distortion gradient to determine whether the defect is an isolated surface damage or a derivative defect induced by internal stress distortion. During the matching process, the system dynamically corrects the defect edges based on the continuity of the gradient field, eliminates false defect signals caused by distortion, and strengthens the real defect features located in the high gradient region, thereby achieving a deep association between "defect location and distortion attribution".

[0072] Based on the results of cross-matching, the system reconstructs a digital twin model of an industrial FA lens in virtual space. This model is no longer a static three-dimensional mesh, but a dynamic data volume carrying rich physical properties. The system fuses and encodes surface topography information (from early defect features), geometric morphology information (from distortion gradient field), and implicit optical performance information (derived from the former two). The final output digital twin data contains full-dimensional attributes of "topography-distortion-optics", which can reflect the real physical state of the lens in the current usage scenario in real time and provide accurate input parameters for subsequent optical performance mapping.

[0073] Specifically, the system collects the early microcrack features extracted in step S121 (located in the lower right quadrant of the lens) and the distortion gradient field constructed in step S122 (showing a significant stress compression vector field in the lower right region). In the cross-fusion module, the system maps the coordinates of the microcrack onto the distortion gradient field and finds that the crack is located in the high stress concentration area (the region of maximum vector magnitude) of the gradient field.

[0074] Through cross-matching, the system determined that the microcrack was not an accidental surface scratch, but a stress crack (physical location defect) caused by mechanical stress compression (distortion mechanical distribution) during lens assembly. Based on this correlation analysis, the system dynamically output digital twin data of the A-type industrial FA lens: in the digital twin model, this area was marked as a red high-risk area. Its data attributes not only included the geometric size of the crack (morphological attributes), but also the distortion attribute of "compression stress-induced" distortion, and deduced that this area would produce obvious astigmatism and coma (optical attributes). This full-dimensional attribute data provided a decisive basis for subsequent determination of whether the lens was qualified, distinguishing between simple surface defects and structural damage.

[0075] Furthermore, a preset optical performance mapping network is obtained. After loading the twin data into the preset optical performance mapping network, the optical performance mapping network dynamically divides multiple topological nodes representing different optical performance ranges of the industrial FA lens according to the density of aberration distribution and the field of view attenuation boundary. At this time, the topological nodes are used as anchor points and traced backward to extract the node features of the current frame. It is also possible to cross-modal associate the context information corresponding to the topological node. In the process of aggregating the context information, multi-source heterogeneous data with causal relationship is formed. This context information covers the material information of the industrial FA lens, the corresponding lens processing route, and the historical environment.

[0076] At this point, the system retrieves the pre-trained deep neural network model—the optical performance mapping network. This network has learned the nonlinear mapping relationship between "appearance defects / distortion features" and "optical transfer function (MTF) / aberration distribution" through a large number of historical samples. The digital twin data containing the full-dimensional attributes of "morphology-distortion-optics" generated in step S131 is loaded into the input layer of this network. The network transforms the geometric distortion gradient and surface defect features in the digital twin data into specific aberration distribution maps, thereby reconstructing the actual optical performance of the industrial FA lens in the feature space.

[0077] In the hidden space of the optical performance mapping network, the system performs adaptive clustering analysis based on the density of aberration distribution and the field attenuation boundary. For regions with dense aberrations and a sharp decline in optical performance (such as high-stress areas of the distortion gradient field), the network identifies them as independent functionally abnormal regions. For regions with significant energy attenuation at the edge of the field of view, the system delineates the boundary based on the attenuation gradient. Based on this, the network dynamically divides multiple topological nodes that characterize different optical performance ranges. These nodes are not fixed physical grids, but logical units constructed based on the similarity of optical performance. Each node represents a specific optical state cluster (such as "central high-resolution region", "edge coma region", "stress-induced defect region").

[0078] Using the identified topological nodes as spatial anchors, the system initiates a backward tracing mechanism to retrieve the original feature set upon which the node depends. The system not only extracts the texture and frequency domain features (node ​​features) of the node region in the current frame image, but also triggers a cross-modal association engine to retrieve unstructured data matching the node's location and features from the database. By comparing the defect features with the time stamps of the production logs, the system aggregates relevant contextual information and establishes a logical chain tracing back from "optical performance" to "production causes," thereby transforming simple image features into engineering data with causal explanatory capabilities.

[0079] The system deeply integrates visual inspection data with heterogeneous data from the Industrial Internet to generate a multi-source heterogeneous dataset with causal relationships. This dataset not only describes "what defects exist" but also explains "why the defects occur." The integrated contextual information specifically covers: material batch information of industrial FA lenses (such as the batch uniformity of refractive index of optical glass), corresponding lens processing routes (such as fine grinding process parameters and coating deposition rate), and historical environmental data (such as temperature and humidity fluctuations in the processing workshop). Each topology node carries a complete physical identity and process genes, realizing a leap from appearance inspection to root cause analysis.

[0080] Specifically, the system loads the digital twin data of the A-type industrial FA lens into the optical performance mapping network. Network analysis reveals a significant aberration clustering phenomenon in the lower right quadrant of the lens, and an abnormal energy attenuation boundary appears at the edge of the field of view in this region. Therefore, this region is dynamically divided into a topological node "T-Node_03", which represents the optical performance range of "local coma exceeding the standard and resolution dropping sharply".

[0081] The system traces backward using "T-Node_03" as the anchor point, extracting node features in the frequency domain image where high-frequency energy is missing. Simultaneously, the cross-modal correlation engine quickly searches the production database and finds that during the processing of this type A industrial FA lens, the coating process record in the lower right quadrant shows that "the vacuum chamber pressure fluctuated slightly at time T-120" (historical environment), and the material report for this batch of lenses shows that the corresponding region has "stress birefringence slightly higher than the standard deviation" (material information).

[0082] The system aggregated multi-source heterogeneous data at the “T-Node_03” node, clearly indicating that the optical performance of this topological node was abnormal (dense aberration distribution). The root cause was the uneven coating thickness caused by material internal stress and environmental air pressure fluctuations (processing route deviation). This multi-source heterogeneous data with a causal relationship provided accurate data support for the subsequent correction of the processing technology of this batch of lenses.

[0083] refer to Figure 5 In step S14, the specific steps are as follows:

[0084] S141: Acquire the multi-source heterogeneous data, nonlinearly map the multi-source heterogeneous data to a unified feature manifold space, sequentially form the optical attenuation trend of the industrial FA lens in the full field of view along the evolution of multiple attenuation features in the feature manifold space, and mark the lens drift data caused by temperature fluctuations or mechanical vibrations in real time; simultaneously acquire the previous test data of the industrial FA lens, perform multi-factor fusion of the previous test data and lens drift data of the industrial FA lens, and construct the corresponding dynamic detection framework during the fusion process;

[0085] S142: Dynamically identify the dynamic detection framework and incorporate a memory network mechanism during the identification process. This allows for multi-level iteration along the memory network, thereby generating a corresponding dynamic detection system. This dynamic detection system can autonomously adjust the judgment threshold and key information of each topology node based on the initial manifold distribution of different batches of industrial FA lenses. This enables intelligent closed-loop detection of various industrial FA lenses without relying on manual parameter tuning.

[0086] In the embodiments of this application, the multi-source heterogeneous data is acquired, and the multi-source heterogeneous data is nonlinearly mapped to a unified feature manifold space. In the feature manifold space, the optical attenuation trend of the industrial FA lens in the full field of view is sequentially formed along the evolution of multiple attenuation features, and the lens drift data caused by temperature fluctuations or mechanical vibrations is marked in real time. Simultaneously, the previous test data of the industrial FA lens is acquired, and the previous test data and lens drift data of the industrial FA lens are fused by multiple factors. In the fusion process, a corresponding dynamic detection framework is constructed, and a corresponding dynamic detection framework is introduced.

[0087] At this point, the system acquires multi-source heterogeneous data containing the causal relationship of "material-process-environment" generated in step S132. Since these data have huge differences in dimension, scale and distribution characteristics (such as continuous temperature curves and discrete processing codes), the system adopts nonlinear manifold learning methods (such as kernel principal component analysis or autoencoder mapping) to project these high-dimensional heterogeneous data into a unified low-dimensional feature manifold space. In this manifold space, the data that originally had very different physical meanings are transformed into feature vectors with geometric measurability, so that process defects with similar potential attributes are spatially close to each other, thereby realizing the global fusion and normalized expression of multi-source data.

[0088] Within a unified feature manifold space, the system performs topological connection and trajectory tracking on multiple attenuation features across the entire field of view of industrial FA lenses. Based on the evolution path of feature points in the manifold space, the system simulates the loss trend of light energy during transmission from the center to the edge of the lens, thereby sequentially forming an optical attenuation trend map. This trend not only reflects the inherent optical characteristics of the lens design, but also reveals abnormal energy transmission paths caused by local defects or material inhomogeneities through changes in manifold curvature, achieving precise quantification of the decay law of optical performance across the entire field of view.

[0089] While analyzing the optical attenuation trend, the system introduces a high-frequency time-series analysis module to capture minute jitters at manifold feature points in real time. By comparing with a standard static benchmark, the system accurately identifies and marks lens drift data caused by external environmental interference. Among these, temperature fluctuations manifest as a systematic translation of feature vectors in the manifold space (reflecting focal plane drift caused by thermal expansion and contraction), while mechanical vibrations manifest as high-frequency oscillations of feature vectors (reflecting image plane jitter caused by external excitation sources). The system isolates and marks these dynamic deviations caused by non-inherent defects to prevent them from interfering with the determination of the lens's intrinsic quality.

[0090] The system synchronously calls up the previous test data of industrial FA lenses (i.e., the historical test records of the entire life cycle), and performs multi-factor fusion of the steady-state characteristics in the historical data with the currently marked drift data. The system dynamically adjusts the test logic according to the fusion results: if the historical data shows that the lens performance is stable and the current drift is minimal, then a standard and stringent test framework is adopted; if significant temperature drift or vibration interference is detected, an anti-interference factor and a threshold correction matrix are adaptively introduced to build a dynamic test framework for the current working conditions, ensuring that the test system can maintain high robustness under different environments and batches.

[0091] Specifically, the system maps the material batch data, fine grinding process parameters, and aberration features of the current test of the A-type industrial FA lens to a unified feature manifold space. In the manifold space, the system finds that the feature points representing the edge field of view of the lens exhibit an evolutionary trajectory that converges towards the low light energy region, forming an optical attenuation trend of "high brightness at the center - nonlinear attenuation at the edge" in sequence, which is consistent with the edge light energy reduction characteristics of the F1.4 large aperture lens.

[0092] During the real-time labeling process, the system detected minute periodic oscillations in the characteristic trajectories within the manifold space. Upon analysis, the oscillation frequency was found to be consistent with the vibration frequency of the factory workshop's foundation. The system determined that this was lens drift data caused by mechanical vibration and labeled it as "external interference" to avoid misjudging it as a resolution defect in the lens itself.

[0093] The system retrieved the test records of the Type A industrial FA lens over the past three months and found that its historical data performance was stable. Combined with the existing minor vibration drift, the system automatically triggered the "dynamic image stabilization mode" in the dynamic detection framework during the multi-factor fusion process. It appropriately relaxed the judgment tolerance of the image edge sharpness threshold and enabled the inter-frame averaging filtering method. This dynamic detection framework successfully eliminated vibration interference and output the true quality rating of the Type A industrial FA lens, ensuring the objectivity and accuracy of the test results.

[0094] Furthermore, the dynamic detection framework is dynamically identified, and a memory network mechanism is introduced during the identification process. This allows for multi-level iteration along the memory network, thereby generating a corresponding dynamic detection system. This dynamic detection system can autonomously adjust the judgment thresholds and key information of each topology node based on the initial manifold distribution of different batches of industrial FA lenses. This enables intelligent detection of various industrial FA lenses without relying on manual parameter tuning. Simultaneously, it further controls multiple topology nodes of the industrial FA lenses, fully considering lens drift data, combined memory networks, and previous detection data of the industrial FA lenses, thus improving the accuracy of the dynamic detection system for industrial FA lenses.

[0095] At this point, the system performs real-time status identification on the dynamic detection framework constructed in step S141, analyzes its current logical structure and parameter configuration, and determines whether it is suitable for the current detection task. The system introduces a memory network mechanism, which has long short-term memory units built inside, used to store classic benchmark detection models and short-term changing process fluctuation characteristics, respectively. The memory network is connected to the dynamic detection framework as an external knowledge base, enabling it to have intelligent cognitive capabilities of "memory-retrieval-comparison", providing historical experience support for subsequent iterative optimization.

[0096] The system iterates at multiple levels along the topology of the memory network. In the feature-level iteration, the system compares the feature manifold of the current shot with historical successful cases in the memory network and corrects the weights of feature extraction. Secondly, in the decision-level iteration, the system uses the error backpropagation mechanism in the memory unit to dynamically adjust the decision logic within the detection framework. Through this multi-level iteration, the dynamic detection framework can continuously approach the optimal solution under the current working conditions, eliminate systematic biases caused by sudden environmental changes or sample differences, and thus gradually generate a highly adaptive dynamic detection system during the iteration process.

[0097] This dynamic detection system has autonomous decision-making capabilities. When different batches of industrial FA lenses enter the inspection station, the system quickly scans the initial manifold distribution of the lenses and identifies the common characteristics of the batch (such as background grayscale shifts caused by specific material batches or edge aberration distributions caused by specific processing techniques). Based on the identification results, the system autonomously adjusts the judgment thresholds and key information of each topology node. For example, for batches with high background noise, it automatically increases the filtering intensity and relaxes the absolute grayscale threshold; for batches sensitive to edge aberrations, it specifically strengthens the judgment weight of edge topology nodes. This process is completely free from manual intervention and realizes the automated configuration of the detection logic.

[0098] Through the aforementioned adaptive adjustments, the dynamic detection system ultimately achieves a closed-loop intelligent detection system for various industrial FA lenses. The system not only outputs the final pass / fail judgment result, but also transmits the data characteristics of this test back to the memory network in real time to update its weight parameters, achieving continuous evolution of "the more it is tested, the more accurate it becomes". This closed-loop mechanism ensures that the detection system can seamlessly adapt to fluctuations in incoming materials, equipment aging and environmental changes, and always maintain a high-precision detection state.

[0099] Specifically, when inspecting a batch of newly arrived Type A industrial FA lenses, the system identifies that the lenses use a new type of low-refractive-index glass material, which slightly improves the imaging brightness, causing the original background grayscale threshold to no longer be applicable; at this time, the dynamic detection framework constructed in step S141 is activated and connected to the memory network.

[0100] The system retrieved historical experience regarding "high-transmittance materials" from the memory network and discovered that such lenses are highly prone to overexposure artifacts at a large aperture of F1.4. The system then initiated an iterative process, adjusting the cutoff frequency of the frequency domain filter at the feature level and dynamically lowering the overexposure alarm threshold of the central field-of-view topology node at the decision level. Simultaneously, based on the initial manifold distribution of this batch of Type A industrial FA lenses, the dynamic detection system autonomously increased the judgment threshold of the central region by 5% and fine-tuned the distortion tolerance of the edge topology nodes. The system accurately identified a coating defect in this batch of Type A industrial FA lenses, without misjudging due to increased brightness. Furthermore, the detection data was written into the memory network as a correction basis for subsequent detection of similar Type A industrial FA lenses, successfully achieving an unattended intelligent detection closed loop.

[0101] Please see Figure 6 , Figure 6 This is a schematic diagram of the structural composition of the vision-based industrial FA lens detection system according to an embodiment of the present invention; the vision-based industrial FA lens detection system is applied to the above-mentioned vision-based industrial FA lens detection method; the vision-based industrial FA lens detection system includes:

[0102] The vision inspection module 21 is used to transport the industrial FA lens to the vision inspection station and adjust the LED multi-zone matrix light source of the vision inspection station to a low-angle dark field, a high-angle bright field and a narrow-band light source, so as to acquire multiple current images of the industrial FA lens in a multi-dimensional optical scene during the vision inspection of the industrial FA lens.

[0103] The early defect feature module 22 is used to determine the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, and extract multiple early defect features of the industrial FA lens in different dimensions along the frequency domain. At the same time, it determines multiple nonlinear distortion information based on the inverse decoupling of each current image, and constructs the corresponding distortion gradient field along the multiple nonlinear distortion information.

[0104] The multi-source heterogeneous data module 23 is used to cross-fuse multiple early defect features and corresponding distortion gradient fields of the industrial FA lens, thereby outputting digital twin data of the industrial FA lens in the current application scenario. The digital twin data is input into a preset optical performance mapping network and the multiple topological nodes of the industrial FA lens are dynamically divided. The corresponding multi-source heterogeneous data is determined based on the tracing of each topological node.

[0105] The dynamic detection system module 24 is used to map the multi-source heterogeneous data to a unified feature manifold space, thereby determining the optical attenuation trend of the industrial FA lens in the full field of view and marking the corresponding lens drift data. Furthermore, it combines the memory network and previous detection data of the industrial FA lens to determine the dynamic detection system of the industrial FA lens. This dynamic detection system can be adapted to the dynamic detection of different batches of industrial FA lenses.

[0106] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided and embodied by multiple modules.

[0107] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A vision-based inspection method for industrial FA lenses, characterized in that, include: The industrial FA lens is transported to the vision inspection station, and the LED multi-zone matrix light source of the vision inspection station is adjusted to low-angle dark field, high-angle bright field and narrow-band light source, so as to collect multiple current images of the industrial FA lens in the multi-dimensional optical scene during the vision inspection of the industrial FA lens. The frequency domain of the industrial FA lens is determined based on the dynamic transformation of multiple current images, and multiple early defect features of the industrial FA lens in different dimensions are extracted along the frequency domain. At the same time, multiple nonlinear distortion information is determined based on the inverse decoupling of each current image, and the corresponding distortion gradient field is constructed along the multiple nonlinear distortion information. The early defect features and corresponding distortion gradient fields of the industrial FA lens are cross-fused to output digital twin data of the industrial FA lens in the current application scenario. The digital twin data is input into a preset optical performance mapping network and multiple topological nodes of the industrial FA lens are dynamically divided. The corresponding multi-source heterogeneous data is determined based on the tracing of each topological node. The multi-source heterogeneous data is mapped to a unified feature manifold space to determine the optical attenuation trend of industrial FA lenses across the entire field of view and to mark the corresponding lens drift data. Furthermore, by combining the memory network and previous test data of industrial FA lenses, a dynamic testing system for industrial FA lenses is determined. This dynamic testing system can be adapted to the dynamic testing of different batches of industrial FA lenses.

2. The detection method for industrial FA lenses based on vision inspection according to claim 1, characterized in that, The process involves transporting the industrial FA lens to the vision inspection station, and adjusting the LED multi-zone matrix light source at the vision inspection station to a low-angle dark field, a high-angle bright field, and a narrowband light source. This allows for the acquisition of multiple current images of the industrial FA lens within a multi-dimensional optical scene during the vision inspection of the industrial FA lens, including: When the industrial FA lens is delivered to the vision inspection station, the LED multi-zone matrix light source at the vision inspection station responds to the arrival signal of the industrial FA lens. The LED multi-zone matrix light source adaptively flickers along the time sequence according to the signal of the industrial FA lens and the current environment of the vision inspection station, switching between low-angle dark field, high-angle bright field and narrowband light source. At this time, the low-angle dark field excites the scattered light on the surface of the FA lens with a grazing angle; the high-angle bright field penetrates the lens body and captures the transmission pattern of impurities; and the narrowband light source, through a preset center wavelength bandpass filter, excites the interference fringe response of the multilayer coating under a specific spectrum.

3. The detection method for industrial FA lenses based on vision inspection according to claim 2, characterized in that, The process of transporting the industrial FA lens to the vision inspection station and adjusting the LED multi-zone matrix light source of the vision inspection station to a low-angle dark field, a high-angle bright field, and a narrow-band light source, thereby acquiring multiple current images of the industrial FA lens in a multi-dimensional optical scene during the vision inspection of the industrial FA lens, also includes: The low-angle dark field, high-angle bright field and narrowband light source are fused to construct a corresponding multi-dimensional optical scene. At the same time, the industrial FA lens is visually inspected, and multiple current images of the lens are determined in combination with the multi-dimensional optical scene during the visual inspection process. Simultaneously, the surface morphology, internal medium and coating spectral characteristics of the industrial FA lens are pre-filtered at the visual inspection end of the industrial FA lens.

4. The detection method for industrial FA lenses based on vision inspection according to claim 1, characterized in that, The process involves determining the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, extracting multiple early defect features of the industrial FA lens in different dimensions along the frequency domain, and simultaneously determining multiple nonlinear distortion information based on the inverse decoupling of each current image. A corresponding distortion gradient field is then constructed along these multiple nonlinear distortion information, including: Multiple current images are sorted along the time domain to form a corresponding current image sequence. A three-dimensional Fourier transform is performed on the current image sequence, and dynamic transformation is performed by combining wavelet packet decomposition mechanism, thereby mapping the current image sequence to the frequency domain space. In this frequency domain space, multiple early defect features corresponding to different defect scales are extracted by radial frequency slicing.

5. The detection method for industrial FA lenses based on vision inspection according to claim 4, characterized in that, The process of determining the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, extracting multiple early defect features of the industrial FA lens in different dimensions along the frequency domain, and simultaneously determining multiple nonlinear distortion information based on the inverse decoupling of each current image, and constructing a corresponding distortion gradient field along the multiple nonlinear distortion information, further includes: Each current image is simultaneously labeled, and each current image is decoupled in reverse. The image offset content of the target object is dynamically calibrated, thereby determining multiple nonlinear distortion information by dynamically tracking the image offset content. These multiple nonlinear distortion information include lens eccentricity information, tilt deformation information, and surface error information. At the same time, using the pixel neighborhood as the topological unit, a distortion gradient field with continuous vector distribution characteristics is constructed along the multiple nonlinear distortion information. This gradient field can map the deflection state and aberration evolution path of the industrial FA lens in an intuitive flow field shape.

6. The detection method for industrial FA lenses based on vision inspection according to claim 1, characterized in that, The process involves cross-fusing multiple early defect features and corresponding distortion gradient fields of the industrial FA lens to output digital twin data of the industrial FA lens in the current application scenario. This digital twin data is then input into a preset optical performance mapping network, and multiple topological nodes of the industrial FA lens are dynamically divided. Based on the tracing of each topological node, the corresponding multi-source heterogeneous data is determined, including: Multiple early defect features of industrial FA lenses are collected, and these early defect features and their corresponding distortion gradient fields are input into a cross-fusion module. The physical location of the early defect features and the mechanical distribution of the distortion gradient field are cross-matched, thereby dynamically outputting digital twin data of the industrial FA lens in the current usage scenario during the matching process. This digital twin data contains full-dimensional attributes of "morphology-distortion-optics".

7. The detection method for industrial FA lenses based on vision inspection according to claim 6, characterized in that, The process involves cross-fusing multiple early defect features and corresponding distortion gradient fields of the industrial FA lens to output digital twin data of the industrial FA lens in the current application scenario. This digital twin data is then input into a preset optical performance mapping network, and multiple topological nodes of the industrial FA lens are dynamically divided. Based on the tracing of each topological node, the corresponding multi-source heterogeneous data is determined. The process also includes: A preset optical performance mapping network is obtained. After loading the twin data into the preset optical performance mapping network, the optical performance mapping network dynamically divides multiple topological nodes representing different optical performance ranges of industrial FA lenses according to the density of aberration distribution and field of view attenuation boundary. At this time, the topological nodes are used as anchor points and traced backward to extract the node features of the current frame. It is also possible to associate the context information corresponding to the topological node across modes. In the process of aggregating context information, multi-source heterogeneous data with causal relationship is formed. This context information covers the material information of industrial FA lenses, the corresponding lens processing route, and historical environment.

8. The detection method for industrial FA lenses based on vision inspection according to claim 1, characterized in that, The process involves mapping the multi-source heterogeneous data into a unified feature manifold space to determine the optical attenuation trend of the industrial FA lens across the entire field of view, and marking the corresponding lens drift data. Furthermore, by combining a memory network and previous test data of the industrial FA lens, a dynamic testing system for the industrial FA lens is determined. This dynamic testing system is adaptable to the dynamic testing of different batches of industrial FA lenses, including: The multi-source heterogeneous data is acquired and nonlinearly mapped to a unified feature manifold space. Within this feature manifold space, the optical attenuation trend of the industrial FA lens is sequentially formed along the evolution of multiple attenuation features, and lens drift data caused by temperature fluctuations or mechanical vibrations is marked in real time. Simultaneously, the previous test data of the industrial FA lens is acquired, and the previous test data and lens drift data of the industrial FA lens are fused by multiple factors. During the fusion process, a corresponding dynamic detection framework is constructed.

9. The detection method for industrial FA lenses based on vision inspection according to claim 8, characterized in that, The process involves mapping the multi-source heterogeneous data into a unified feature manifold space to determine the optical attenuation trend of the industrial FA lens across the entire field of view, and marking the corresponding lens drift data. Further, it combines a memory network and previous test data of the industrial FA lens to determine a dynamic testing system for the industrial FA lens. This dynamic testing system is adaptable to the dynamic testing of different batches of industrial FA lenses and also includes: The dynamic detection framework is dynamically identified, and a memory network mechanism is introduced during the identification process to perform multi-level iterations along the memory network, thereby generating a corresponding dynamic detection system. This dynamic detection system can autonomously adjust the judgment threshold and key information of each topology node according to the initial manifold distribution of different batches of industrial FA lenses, thereby achieving intelligent detection closed loop for various industrial FA lenses without relying on manual parameter adjustment.

10. A vision-based inspection system for industrial FA lenses, characterized in that, The vision-based inspection system for industrial FA lenses is applied to the vision-based inspection method for industrial FA lenses as described in any one of claims 1-9; the vision-based inspection system for industrial FA lenses comprises: The vision inspection module is used to transport the industrial FA lens to the vision inspection station and adjust the LED multi-zone matrix light source of the vision inspection station to low-angle dark field, high-angle bright field and narrow-band light source, so as to acquire multiple current images of the industrial FA lens in a multi-dimensional optical scene during the vision inspection of the industrial FA lens. The early defect feature module is used to determine the frequency domain of the industrial FA lens based on the dynamic transformation of multiple current images, and extract multiple early defect features of the industrial FA lens in different dimensions along the frequency domain. At the same time, it determines multiple nonlinear distortion information based on the inverse decoupling of each current image, and constructs the corresponding distortion gradient field along the multiple nonlinear distortion information. The multi-source heterogeneous data module is used to cross-fuse multiple early defect features and corresponding distortion gradient fields of industrial FA lenses, thereby outputting digital twin data of industrial FA lenses in the current application scenario. This digital twin data is input into a preset optical performance mapping network, and multiple topological nodes of industrial FA lenses are dynamically divided. The corresponding multi-source heterogeneous data is determined based on the tracing of each topological node. The dynamic detection system module is used to map the multi-source heterogeneous data into a unified feature manifold space, thereby determining the optical attenuation trend of the industrial FA lens in the full field of view and marking the corresponding lens drift data. Furthermore, by combining the memory network and the previous detection data of the industrial FA lens, the dynamic detection system of the industrial FA lens is determined. This dynamic detection system can be adapted to the dynamic detection of different batches of industrial FA lenses.

Images