Vacuum coating surface defect online detection system and method based on industrial vision

Abstract

The invention provides a vacuum coating surface defect online detection system and method based on industrial vision. The system comprises a visual acquisition device and a control processing device. And the control processing equipment acquires a transmission step length parameter of the coated workpiece, and controls the visual acquisition equipment to perform high-frequency alternate gating so as to acquire multiple frames of surface associated images under different irradiation angles. The system extracts a highlight abnormal area in the image, calculates an actual movement track of the highlight abnormal area, and compares the actual movement track with a theoretical translation track calculated based on a transmission step length. When the track deviation degree exceeds a set threshold value, optical pseudo defects are judged to be eliminated, and when the track deviation degree does not exceed the set threshold value, real physical defects are judged, and process compensation parameters are generated and fed back to a machine table. According to the scheme, the surface reflection phenomenon and physical damage can be effectively distinguished, and the accuracy of detection operation is improved.

Description

Technical Field

[0001] This invention relates to the field of machine vision and defect detection technology, and in particular to an online detection system and method for defects on vacuum-coated surfaces based on industrial vision. Background Technology

[0002] Vacuum coating is commonly used to create a thin metal film with high reflectivity on the surface of various hardware and plastic parts. After the forming process is completed, the workpiece usually needs to be connected to a continuous conveyor area for appearance quality screening. In the physical environment of the implementation site, the formed coated workpiece is laid flat on the surface of the conveyor belt and undergoes physical spatial displacement with the direction of operation. External stray light refraction or illumination light reflection from inside the inspection station will produce local bright spots on the high-brightness curved surface of the coated workpiece, and these bright spots will undergo pixel shift with dynamic changes in the spatial observation angle.

[0003] Conventional static visual inspection logic typically performs single-frame feature extraction. A single two-dimensional image is insufficient to decouple the physical displacement and optical slippage of the light spot, easily mislabeling the aforementioned wandering localized reflective bright spots as physical defects such as coating peeling, delamination, or incomplete coating. This confusion between optical pseudo-defects and real physical defects interferes with the accurate sorting of defective products, thereby affecting the construction of the digital closed-loop system for correcting the underlying core manufacturing parameters. Summary of the Invention

[0004] The purpose of this invention is to provide an online detection system and method for surface defects in vacuum coating based on industrial vision, so as to solve the problems pointed out in the background art.

[0005] In a first aspect, the present invention provides an online detection system for surface defects in vacuum coating based on industrial vision, including a vision acquisition device and a control and processing device; the vision acquisition device is used to perform image acquisition operations on the coated workpiece moving on the conveyor belt and output the image sequence to be detected;

[0006] The control and processing equipment is equipped with a dynamic control module for removing false defects.

[0007] The pseudo-defect stripping dynamic control module is used to obtain the conveying step length parameters of the coated workpiece on the conveyor belt.

[0008] The pseudo-defect stripping dynamic control module sends a high-frequency alternating gating command to the visual acquisition device based on the transmission step size parameter, and controls the visual acquisition device to acquire multiple frames of surface association images under different illumination angles.

[0009] The pseudo-defect stripping dynamic control module extracts the bright abnormal regions contained within the multi-frame surface association images and calculates the actual movement trajectory of the bright abnormal regions in the multi-frame surface association images.

[0010] The pseudo-defect stripping dynamic control module compares the actual movement trajectory with the theoretical translation trajectory calculated based on the transmission step size parameter.

[0011] When the deviation between the actual movement trajectory and the theoretical translation trajectory exceeds a set threshold, the pseudo-defect stripping dynamic control module determines that the high-brightness abnormal area is an optical pseudo-defect caused by surface reflection and removes it. When the deviation does not exceed the set threshold, the pseudo-defect stripping dynamic control module determines that the high-brightness abnormal area is a real physical defect and feeds back the real physical defect generation process compensation parameters to the vacuum coating machine.

[0012] Optionally, the visual acquisition device includes an area array industrial camera and a multi-channel ring light source;

[0013] The pseudo-defect stripping dynamic control module divides the multi-channel ring light source into multiple sub-light sources;

[0014] The high-frequency alternating gating instruction is used to control the various sub-light sources inside the multi-channel ring light source to take turns lighting up.

[0015] The pseudo-defect stripping dynamic control module controls the area array industrial camera to synchronously perform exposure operations when each of the sub-light sources is lit, so as to form the multi-frame surface-related images.

[0016] Optionally, the control and processing device is also equipped with a basic image preprocessing module;

[0017] The basic image preprocessing module uses an edge detection operator to extract the outer contour boundary of the coated workpiece in the multi-frame surface association image, and performs a background segmentation operation based on the outer contour boundary to locate the high-brightness abnormal area within the coated workpiece.

[0018] Optionally, the conveyor belt is equipped with a servo drive shaft;

[0019] The online detection system for surface defects in vacuum coating based on industrial vision also includes an encoder unit;

[0020] The encoder unit is mounted on the end of the servo drive shaft;

[0021] The pseudo-defect stripping dynamic control module is communicatively connected to the encoder unit to read the physical displacement pulse signal of the conveyor belt and convert the physical displacement pulse signal into the conveying step length parameter.

[0022] Optionally, the control processing device is connected to a sorting execution unit;

[0023] After determining the actual physical defect, the control and processing equipment classifies the coated workpiece as defective based on the morphological attributes of the actual physical defect and sends a sorting action signal to the sorting execution unit.

[0024] The sorting execution unit receives the sorting action signal and moves the corresponding coated workpiece out of the main production line.

[0025] Optionally, when the pseudo-defect stripping dynamic control module calculates the theoretical translation trajectory, it extracts the center pixel coordinates of the bright abnormal region in the initial frame image of the multi-frame surface-related images, and combines the transmission step size parameter and the camera calibration matrix to obtain the expected set of theoretical pixel coordinates of the bright abnormal region in subsequent frames, and fits the set of theoretical pixel coordinates to the theoretical translation trajectory.

[0026] Optionally, when calculating the degree of deviation between the actual movement trajectory and the theoretical translation trajectory, the pseudo-defect stripping dynamic control module uses a dynamic time warping algorithm to quantify the cumulative Euclidean distance between the coordinate point pairs of the actual movement trajectory and the theoretical translation trajectory.

[0027] When the cumulative Euclidean distance is greater than the distance limit, it is determined that the degree of deviation exceeds the set threshold.

[0028] Optionally, the control processing device is further configured with a region division module;

[0029] The region division module calculates the surface reflectance gradient distribution matrix based on the surface grayscale image under natural light, and divides the surface of the coated workpiece into diffuse reflection region, specular reflection region and interference edge region according to the surface reflectance gradient distribution matrix.

[0030] The control and processing device activates high-frequency alternating gating commands at different angles for the specular reflection area, and adopts full-brightness constant-on illumination operation for the diffuse reflection area.

[0031] Optionally, the control processing device extracts the local color saturation variance for the interference edge region;

[0032] When the local color saturation variance is within the coating interference color difference threshold range and there are no abrupt pixel edges, the control processing device determines that the corresponding visual difference is a color interference artifact and suppresses the alarm action.

[0033] Secondly, the present invention provides an online detection method for defects on vacuum-coated surfaces based on industrial vision, comprising the following steps:

[0034] Obtain the conveying step length parameters of the coated workpiece on the conveyor belt;

[0035] Based on the transmission step size parameter, a high-frequency alternating gating command is issued to the visual acquisition device;

[0036] Control the visual acquisition device to acquire multiple frames of surface-related images under different illumination angles;

[0037] Extract the highlighted abnormal regions contained within the multi-frame surface association images;

[0038] Calculate the actual movement trajectory of the highlighted abnormal region in the multi-frame surface association image;

[0039] The actual movement trajectory is compared with the theoretical translation trajectory calculated based on the transmission step size parameter;

[0040] When the deviation between the actual movement trajectory and the theoretical translation trajectory exceeds a set threshold, the high-brightness abnormal area is determined to be an optical pseudo-defect caused by surface reflection and is removed.

[0041] When the deviation does not exceed the set threshold, the high-brightness abnormal area is determined to be a real physical defect, and the process compensation parameters are fed back to the vacuum coating machine based on the real physical defect.

[0042] The present invention has achieved the following beneficial effects:

[0043] This invention acquires the transfer step length parameters of the coated workpiece and controls a vision acquisition device to obtain multiple frames of surface-related images under different illumination angles. It extracts the actual movement trajectory of the high-brightness abnormal area in the multiple frames of surface-related images, and then compares it with the theoretical translation trajectory calculated based on mechanical means using Euclidean distance quantization. Based on the degree of deviation, it determines the removal of optical pseudo-defects caused by surface reflection. After determining that it is a real physical defect, it generates process compensation parameters and feeds them back to the vacuum coating machine. Based on its morphological attributes, it sends action signals to the sorting execution unit. This effectively reduces the misjudgment caused by optical interference and reflection, accurately distinguishes between optical phenomena and physical contamination, and realizes the physical separation of manufacturing defects and the feedback adjustment of the underlying equipment.

[0044] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0045] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is a schematic diagram of the architecture of the visual acquisition device and the control processing device of the detection system in an embodiment of the present invention;

[0048] Figure 2 This is a flowchart of steps S1 to S4 of the detection method in an embodiment of the present invention;

[0049] Figure 3 This is a flowchart of steps S5 to S11 of the detection method in this embodiment of the invention;

[0050] Figure 4 This is a flowchart of steps S12 to S13 of the detection method in an embodiment of the present invention. Detailed Implementation

[0051] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0052] An online detection system for surface defects in vacuum coating based on industrial vision is deployed in the continuous conveying area downstream of the vacuum coating machine. In the actual implementation environment, a conveyor belt is physically connected to the discharge port of the vacuum coating machine. The formed coated workpiece is laid flat on the conveyor belt surface and moves in the direction of the conveyor belt's rotation. Because the vacuum coating process generates a highly reflective metal thin film layer on the surface of metal and plastic parts, when the coated workpiece undergoes physical spatial displacement, external stray light refraction or internal illumination light reflection from the detection station will produce localized reflective bright spots on the high-brightness curved surface of the coated workpiece, which dynamically change with the spatial observation angle and exhibit pixel shift. If only conventional static visual inspection logic is used to perform single-frame feature extraction, these wandering localized reflective bright spots will be marked as physical defects such as coating peeling, delamination, or incomplete coating. To eliminate the aforementioned optical parallax misjudgment, the system establishes a data mapping channel between the mechanical conveying execution end and the visual acquisition top layer in the underlying data bus architecture.

[0053] See attached document Figure 1 As shown, the physical topology of the detection system includes a vision acquisition device 100 and a control processing device 200, with a bidirectional digital communication connection established between the vision acquisition device 100 and the control processing device 200. The control processing device 200 is configured as a combination of multiple hardware and software components, such as an industrial control motherboard, an edge computing node, or a programmable logic controller. The control processing device 200 integrates a basic image preprocessing module, a region segmentation module, and a dynamic control module for pseudo-defect removal. (See attached diagram.) Figure 2As shown, the system executes the following technical steps in a specific order:

[0054] Step S1: Obtain the conveying step length parameters of the coated workpiece on the conveyor belt.

[0055] The pseudo-defect removal dynamic control module is used to acquire the conveyor step length parameters of the coated workpiece on the conveyor belt. A servo motor is equipped at the physical drive shaft end of the conveyor belt, responsible for providing linear velocity mechanical traction. The system hardware also includes an encoder unit. The encoder unit is coaxially fixedly connected to the tail bearing of the servo motor. The pseudo-defect removal dynamic control module establishes a hardware pulse communication link with the encoder unit to read the physical displacement pulse signal of the conveyor belt and convert it into conveyor step length parameters.

[0056] Specifically, the encoder unit employs incremental rotary photoelectric encoder hardware. When the servo motor drives the conveyor belt, the transparent grating code disk inside the encoder unit rotates synchronously, continuously cutting the light beam projected by the internal light-emitting diodes. The photoelectric receiving element converts the received light signal into two square wave pulse signals, Phase A and Phase B, which contain 90-degree orthogonal physical phases. The control processing device 200's internal motherboard integrates a field-programmable gate array (FPGA). The FPGA is used to handle high-frequency hardware interrupt requests.

[0057] After receiving the Phase A and Phase B square wave pulse signals, the Field Programmable Gate Array (FPGA) invokes its internally configured digital filtering algorithm module. This module employs shift register majority voting filtering logic, sets a specific clock sampling window, and statistically analyzes the logic level states within the sampling window to filter out high-frequency glitches caused by electromagnetic interference from the industrial inverter. After filtering, the FPGA uses its internal phase detection and counting circuit to perform a fourth-harmonic frequency direction determination on the rising and falling edges of the Phase A and Phase B square wave pulse signals. If the phase transition of the Phase A square wave pulse signal leads the Phase B square wave pulse signal, the conveyor belt is determined to be in a forward conveying state, and a hardware register count accumulation operation is performed; otherwise, a hardware register count decrement operation is performed.

[0058] Understandably, the pseudo-defect removal dynamic control module has a preset pulse equivalent conversion coefficient value in its internal static random access memory. This pulse equivalent conversion coefficient value represents the actual linear displacement of the conveyor belt in millimeters corresponding to a single hardware pulse signal. During parameter calculation and initialization, the pseudo-defect removal dynamic control module reads the reciprocal parameter of the reducer's mechanical transmission ratio and the circumference parameter of the driving wheel segment pre-stored in the non-volatile register. The module performs a floating-point multiplication of the reciprocal parameter of the reducer's mechanical transmission ratio and the circumference parameter of the driving wheel segment to generate an intermediate product scalar. Subsequently, it divides this intermediate product scalar by the product of the encoder unit's inherent physical line count resolution per revolution and its fourth harmonic constant to derive the pulse equivalent conversion coefficient value. Within the cyclic detection cycle, the pseudo-defect removal dynamic control module records the number of pulse differences accumulated by the system between two adjacent visual hardware trigger signals. The module performs a double-precision floating-point multiplication of the pulse difference number and the pulse equivalent conversion coefficient value to obtain the transmission step size parameter within the corresponding time window. By collecting hardware pulse data from the mechanical actuator end, the step length parameter is transmitted, effectively compensating for the theoretical displacement prediction error caused by the elastic tensile deformation of the transmission belt and the backlash of the reducer.

[0059] Step S2: Calculate the surface reflectance gradient distribution matrix based on the surface grayscale image under natural light, and divide the surface of the coated workpiece into specific optical reflection regions according to the surface reflectance gradient distribution matrix.

[0060] Before triggering the multi-angle alternating illumination action, in order to avoid the invalid occupation of system computing resources in flat, non-reflective areas, the region division module configured inside the control processing device 200 calculates the surface reflectivity gradient distribution matrix based on the surface grayscale image under natural light, and divides the surface of the coated workpiece into diffuse reflection region, specular reflection region and interference edge region according to the surface reflectivity gradient distribution matrix.

[0061] Specifically, the area-array industrial camera built into the vision acquisition device 100 performs static scanning acquisition on the coated workpiece entering the inspection field of view under backlight illumination, outputting a grayscale image of the surface under natural light. This grayscale image constitutes a two-dimensional digital pixel grayscale matrix. The region segmentation module reads the grayscale image under natural light and constructs a square pixel traversal matrix frame of fixed pixel size within the image's two-dimensional coordinate system. This square pixel traversal matrix frame then slides and scans within the grayscale image under natural light, stepping through the pixels according to a set pixel coordinate increment.

[0062] At the sliding and dwelling coordinate point, the region segmentation module extracts the grayscale value of the central reference pixel corresponding to the center coordinate of the square pixel traversal matrix frame, and extracts the array pixel grayscale values ​​of the surrounding associated adjacent pixel groups. The region segmentation module uses the arithmetic logic unit to calculate the sum of squares of the differences between the grayscale value of the central reference pixel and the grayscale values ​​of all associated adjacent pixel groups, generating a gradient scalar representing the intensity of local reflectivity spatial fluctuations. The calculation process includes: subtracting the grayscale value of the central reference pixel from the grayscale value of each associated adjacent pixel, and obtaining the difference; performing a self-multiplication operation on the obtained difference to obtain the pixel squared deviation value; and summing the pixel squared deviation values ​​corresponding to all associated adjacent pixels. The gradient scalar calculated at the sliding and dwelling coordinate point is re-arrayed according to the physical order of the planar spatial coordinates to generate a surface reflectivity gradient distribution matrix covering the effective field of view of the coated workpiece surface.

[0063] At a deeper level, the region division module internally stores two calibration threshold parameters: a primary threshold for diffuse reflection and a secondary threshold for interference. The primary threshold for diffuse reflection is set to be less than the secondary threshold for interference. The region division module reads the gradient scalar of each specified grid coordinate node in the surface reflectivity gradient distribution matrix. When the gradient scalar of a specified grid coordinate node is less than the primary threshold for diffuse reflection, it indicates that the surface geometry of the corresponding coordinate region is gentle and the reflectivity distribution is uniform; the region division module marks this region as a diffuse reflection region. When the gradient scalar of a specified grid coordinate node is greater than or equal to the primary threshold for diffuse reflection and less than the secondary threshold for interference, it indicates that the corresponding coordinate region has a surface morphology change or directional specular reflection characteristics; the region division module marks this region as a specular reflection region. When the gradient scalar of a specified grid coordinate node is greater than or equal to the secondary threshold for interference, it indicates that the corresponding coordinate region belongs to the acute edge of the workpiece or the boundary of the coating thickness gradient, possessing the physical structural conditions for generating optical thin-film interference phenomena; the region division module marks this region as an interference edge region. The control and processing device 200 adopts a differentiated scheduling strategy based on the output results. For the specular reflection area, it allocates processing threads and enables high-frequency alternating gating instructions; for the diffuse reflection area, it employs a fully illuminated, always-on lighting operation.

[0064] Step S3: Extract the local color saturation variance for the interference edge region and perform color interference artifact identification and motion masking.

[0065] At the chamfered edge of the coated workpiece, the film thickness decreases gradually due to the sputtering angle of the target and the geometric shadowing effect. This spatially varying thickness of the film causes phase-shift interference with the incident light wave, forming optical thin-film interference colored fringes. Static detection logic can easily confuse these optical interference fringes with surface oxidation discoloration defects, leading to the output of rejection commands.

[0066] Specifically, the area array industrial camera outputs a basic color pixel matrix containing red, green, and blue channels. The control processing device 200 extracts the local color saturation variance for the interference edge region. The control processing device 200 reads the color pixel matrix data marked within the interference edge region, performs a color space coordinate transformation matrix operation, and converts the aforementioned basic color channel data into color space model data containing hue, saturation, and lightness components.

[0067] The control processing device 200 extracts saturation component data separately to construct an independent matrix and sets a local pixel detection window. The statistical discrete variance of the saturation components within the local pixel detection window is calculated to obtain the local color saturation variance. The variance mathematical algorithm logic is as follows: calculate the arithmetic mean of all saturation values ​​within the window; calculate the square of the deviation of each saturation value from the arithmetic mean; sum all squared deviations and divide by the total number of pixels involved in the calculation. Surface chemical oxidation discoloration defects manifest as continuous dark color patches with a flat saturation spatial distribution; while physical thin-film interference phenomena manifest as alternating color spaces with spatial periodic oscillations in saturation.

[0068] The system's static memory stores the coating interference color difference threshold range, calibrated based on the good product database. When the extracted local color saturation variance falls within the coating interference color difference threshold range, the control processing device 200 calls the Laplace second-order differential operator to perform spatial domain convolution operations on the lightness component data of the corresponding region. If the second-order derivative response matrix calculated by the convolution operation does not contain any transition edges crossing zero-point judgments, it indicates that the color exhibits a smooth optical phase spatial gradient. The control processing device 200 determines that the visual difference phenomenon in the corresponding region is caused by physical thin-film interference, belonging to color interference artifacts. The control processing device 200 writes an alarm mask logic word for the interference edge region into the system alarm control register. The alarm mask is responsible for blocking the outward execution of the defective product rejection instruction, distinguishing between optical phenomena and physical contamination in the underlying logic judgment.

[0069] Step S4: Issue a high-frequency alternating gating command to obtain multi-frame surface association images under different illumination angles.

[0070] Acquiring multi-angle reflective layers with independent incident directions is a fundamental data prerequisite for trajectory deformation extraction and decoupling of optical false defects. The vision acquisition device 100 includes an area array industrial camera and a multi-channel ring light source. The multi-channel ring light source is positioned around the receiving surface of the area array industrial camera lens, converging the emitted light onto the detection field of view. The false defect removal dynamic control module issues high-frequency alternating gating commands to the vision acquisition device 100 based on the transmission step size parameters.

[0071] At the underlying driver allocation level, the pseudo-defect stripping dynamic control module divides the multi-channel ring light source's emission array into multiple relatively independently powered sub-light sources. Each sub-light source occupies a different spatial illumination azimuth sector. The multi-channel ring light source is divided into multiple sub-light sources occupying different spatial azimuth angles (e.g., front, rear, left, and right illumination sub-light sources). High-frequency alternating gating instructions are used to control the sequential lighting operation of each sub-light source; the pseudo-defect stripping dynamic control module controls the area array industrial camera to synchronously perform exposure operations when each sub-light source is lit, thereby forming multiple frames of surface-related images.

[0072] Understandably, the underlying hardware timing synchronization logic is configured as follows: the pseudo-defect stripping dynamic control module has a built-in timing synchronization microcontroller chip. When the timing synchronization microcontroller chip detects that the cumulative absolute displacement scalar of the transmission step size parameter reaches the preset spatial trigger step threshold, it triggers the multi-threaded parallel synchronization state machine. The timing synchronization microcontroller chip sends an enable high-level pulse to the constant current drive control board of the multi-channel ring light source to activate the front illumination sub-light source, and simultaneously sends the first exposure control pulse to the opto-isolated trigger pin of the area array industrial camera through a physical wire. The photosensitive array of the area array industrial camera starts the charge integration light-receiving operation. After the set exposure time ends, the initial frame image acquisition operation is completed. Subsequently, the timing synchronization microcontroller chip pulls down to cancel the enable pulse of the front illumination sub-light source to cut off the illumination, outputs a secondary enable pulse to activate the left illumination sub-light source, and sends an exposure pulse to the area array industrial camera to acquire subsequent associated image frames. Following the aforementioned switching matrix logic loop, this continues until all sub-light sources in the set orientations have completed their turn-on lighting and collaborative acquisition.

[0073] Because the single illumination cycle of each sub-light source at different angles is within the set exposure time level, the spatial offset of the physical displacement of the coated workpiece on the conveyor belt during the alternating selection polling cycle is objectively less than the set pixel equivalent threshold. The discrete digital image matrix continuously output by the area array industrial camera forms multi-frame surface-correlated images. These multi-frame surface-correlated images are coherent in the temporal dimension, encompass the coated workpiece entity in a relatively stationary displacement state in spatial geometry, and optically map and record the spatial migration distribution of reflected light spots formed on the vacuum-coated surface by light rays at different spatial incident angles. The hardware-level synchronization mechanism shields against natural light interference, giving the sequential images a clear spatial orientation attribute of the light source.

[0074] See attached document Figure 3 As shown, it also includes:

[0075] Step S5: Execute system memory read / write management control to complete the underlying transfer of visual data stream.

[0076] To support the physical storage and exchange of high-speed acquired visual image data streams, the control and processing device 200 has a hard real-time extension module residing in the operating system kernel layer to improve responsiveness under multi-threaded concurrent data processing. In terms of software architecture, the acquisition and monitoring process responsible for receiving low-level network data packets is given high hardware interrupt handling priority by the system kernel scheduler.

[0077] The acquisition and monitoring process utilizes the peripheral component interconnection data bus master mechanism and built-in data transfer channel to bypass the central processing unit's control core register instruction loop, directly writing the received valid image data blocks into the pre-allocated memory buffer pool's contiguous physical address space pre-mounted in the motherboard's physical memory space. The pre-allocated memory buffer pool's data structure employs a lock-free circular queue. The write data physical pointer is driven forward by the underlying network card hardware receive interrupt service routine; the read data physical pointer is driven forward by the image recognition underlying data parsing thread. Through low-level calls to system-level atomic operations for comparison and exchange of execution instructions, the synchronization and alignment of the write and read pointer positions are maintained under concurrent read / write environments. The lock-free concurrent circular queue data operation structure reduces the additional overhead of context switching caused by operating system mutex locking operations in multi-process resource contention, avoiding address out-of-bounds or deadlock blocking during memory physical address allocation, addressing, and write / read operations under conveyor belt operation and camera image capture loads.

[0078] Step S6: Use the edge detection low-level operator to perform outer contour tracking and background digital region pixel masking operations.

[0079] Within the original field of view of the acquired multi-frame surface association images, the conveyor belt bearing surface is characterized by dark textured grooves of anti-slip rubber or seams between the front and rear sections of the conveyor belt. Incorporating environmental background clutter into the dynamic motion trajectory deduction and comparison analysis model would lead to excessive computational resources for coordinate optimization and trigger coordinate point matching actions for surrounding background surface noise lattice regions. Therefore, regional constraint shielding measures must be implemented.

[0080] The control and processing device 200 is internally equipped with a basic image preprocessing module. This module retrieves multi-frame surface-associated image matrices from a lock-free circular queue cache in memory. For each retrieved independent image frame matrix array, the module calls an edge detection operator. During algorithm execution, a Gaussian two-dimensional smoothed normal distribution filter spatial convolution filtering operation is applied to the original pixel grayscale matrix to filter out discrete high-frequency noise. Then, Sobel matrix convolution operators in the horizontal and vertical directions are used to perform discrete partial differential operations in the corresponding spatial dimensions to obtain the magnitude parameter of the overall synthesized grayscale gradient vector and the scalar angle between the normal direction and the normal.

[0081] Next, mathematical processing to suppress non-maximum interfering pixels is performed according to the calculated gradient direction angle scalar. A comparison is made along the gradient vector pointing axis; if the grayscale amplitude variable parameter of the pixel at the processing location center is not a local maxima, a non-maximum grayscale suppression and zeroing action is performed to eliminate excessively widened edge ghost pixels. Following this, a hysteresis dual-threshold logic framework with high and low threshold parameters is configured to divide pixel features into initial connected extension search seed nodes of different strength types. Along the search boundary, continuous spatial paths with drastic grayscale step changes in the image plane distribution are found and extended. Boundary tracing connection algorithms are used to reconstruct a complete set of closed and connected outer contour boundary coordinate line segments.

[0082] Subsequently, the basic image preprocessing module generates a black-and-white two-dimensional logical matrix mask with the same horizontal and vertical dimensions as the original image, based on the calculated outer contour boundary. The true values ​​of each node within the mask are set according to their positional relationships. For pixel positions located within the effective spatial coordinate range of the outer contour boundary, corresponding to the surface of the workpiece being coated, the corresponding coordinate logical storage pixel scalar of the binary judgment matrix is ​​set to a high-level logical value representing the right to read. For pixel positions located outside the outer contour boundary, corresponding to the background interference space of the conveyor belt, the mask matrix matching position data scalar is set to a low-level logical value representing the right to read. In the final stage, the basic image preprocessing module performs a bitwise AND-DOF operation on the generated binary mask matrix and the original multi-frame surface-related image to be processed, performing a data zeroing operation on the original pixel grayscale data of the background area containing the low-level blocking data. This effectively removes and masks optical background clutter at the underlying data processing level.

[0083] Step S7: Extract the connected block region of physical pixels with abnormal surface brightness.

[0084] After background masking, within the effective non-zero pixel set only containing the physical working area of ​​the coated workpiece, the pseudo-defect stripping dynamic control module performs extraction of the physical pixel positions of the high-brightness abrupt optical response characteristics. Statistical algorithms are used to calculate the overall mean of the grayscale arithmetic mean and the standard deviation of the grayscale values ​​contained in the current non-background effective working pixel set. The algorithm logic performs numerical multiplication based on the set grayscale compensation multiplier ratio, adding a dynamically adjusted deviation amplitude summation to the average base value to generate an adaptive floating high-brightness recognition truncation threshold. By traversing all effective working pixels, target pixel units with grayscale values ​​higher than the adaptive floating high-brightness recognition truncation threshold are extracted.

[0085] For the selected target pixel units, the system executes a connected region topology analysis algorithm. Adjacent target pixel units are assigned the same cluster label number, extracting physically isolated bright pixel connected clusters. To eliminate black spots and voids caused by microscopic particles within reflective patches, the pseudo-defect stripping dynamic control module performs mathematical morphological compound closure operations on each bright pixel connected cluster. First, morphological dilation is performed on boundary pixels using a structuring element matrix, followed by morphological erosion using an equivalent structuring element matrix. Subsequently, the system uses the polygonal spatial geometric image moment operation rules to calculate the zero-order moment of the planar area and the first-order position moment of the spatial coordinates of all bright pixel connected clusters. By dividing the first-order position moment by the zero-order moment, the two-dimensional pixel coordinates of the geometric centroid of all bright pixel connected clusters are calculated. The extracted bright pixel connected clusters are marked as bright abnormal regions and include associated initial spatial centroid coordinate parameters.

[0086] Step S8: Calculate the actual movement trajectory of the highlighted abnormal region in the multi-frame surface association image.

[0087] The actual movement trajectory represents the continuous tracking path of the true pixel offset of the bright patch within the projection field of view of the 2D image from the area array industrial camera. The pseudo-defect stripping dynamic control module extracts the center geometric centroid coordinates of the patch region in the initial frame input image as the starting judgment anchor point. Within the secondary frame associated images arranged according to the input time sequence, the pseudo-defect stripping dynamic control module opens a 2D square local search neighborhood with a set pixel width around the spatial coordinate position mapped by the starting judgment anchor point.

[0088] The system calculates the comprehensive cost of matching features between each candidate bright patch and the initial decision anchor point within a two-dimensional square local search neighborhood. Specifically, this includes: using the analytic geometric Euclidean distance formula to calculate the spatial distance cost between the centroid coordinates of the candidate bright patch and the coordinates of the initial decision anchor point; and calculating the absolute area difference cost between the total area of ​​the pixels contained in the candidate bright patch and the total area of ​​the reference patch. The system then uses specified weight parameters to perform a weighted sum of the spatial distance cost and the absolute area difference cost to obtain a comprehensive cost score for each candidate bright patch.

[0089] The pseudo-defect stripping dynamic control module calls the minimum value filtering operator to select the specific candidate bright patch with the lowest comprehensive cost score as the cross-frame matching target object, and extracts the geometric centroid coordinates of the matched object as trajectory extension nodes. The module iterates through all frame data contained in the multi-frame surface association image. The centroid coordinates of the successfully matched cross-frame points are concatenated according to the triggering axis to form a discrete coordinate data structure. The module then calls the cubic spline mathematical interpolation fitting analysis model algorithm to perform a nonlinear geometric coherence smoothing operation on the discrete coordinate sequence, ultimately outputting the actual movement trajectory representing the dynamic position observed visually.

[0090] Step S9: Calculate the theoretical translation trajectory based on the transmission step size parameter.

[0091] To establish an alignment comparison benchmark for rigid body motion, the pseudo-defect stripping dynamic control module performs spatial numerical quantization comparison between the actual movement trajectory and the theoretical translation trajectory calculated based on the conveyor step length parameters. In the calculation phase, the pseudo-defect stripping dynamic control module extracts the center pixel coordinates of the highlighted abnormal region in the initial frame image and, combined with the conveyor step length parameters and camera calibration parameter matrix, calculates the expected pixel coordinate set, fitting this set to the theoretical translation trajectory. The theoretical translation trajectory characterizes the ideal expected pixel offset path appearing in the two-dimensional image field of the area array sensor when physical defects fixed to the physical surface of the coated workpiece are mechanically transported by the conveyor belt.

[0092] Specifically, the control processing device 200 stores a camera calibration parameter matrix in its memory, which includes an internal optical parameter matrix and an external pose parameter transformation matrix. The internal optical parameter matrix includes the camera's equivalent focal length scalar and the optical principal point offset coordinate scalar; the external pose parameter transformation matrix includes orthogonal rotation matrix components and spatial translation vector components.

[0093] When the simulation begins, the pseudo-defect stripping dynamic control module uses the coefficients of the lens radial distortion multivariate correction parameter polynomial and the tangential distortion position correction parameter polynomial to perform distortion correction geometric processing on the center pixel coordinates, obtaining the corresponding absolute coordinate parameters of the ideal projection plane. The absolute coordinate parameters of the ideal projection plane are then multiplied by the inverse matrix of the internal optical parameter matrix to generate a three-dimensional vector of the viewing direction in the normalized camera coordinate system. Subsequently, the rotation orthogonal component matrix and translation component three-dimensional vector of the external pose parameter transformation matrix are called to perform coordinate space system rotation and offset transformation operations, physically mapping the viewing direction three-dimensional vector to the global world reference coordinate system. Combining the height vertical section constraint equation contained in the coating workpiece bearing surface in the global world reference coordinate system, the three-dimensional analytical geometric equations are used to calculate the specific numerical absolute parameter values ​​of the three-dimensional physical intersection point of the viewing direction three-dimensional vector and the height vertical section. The obtained numerical absolute parameter values ​​of the three-dimensional physical intersection point represent the initial physical entity reference three-dimensional spatial coordinates of the center pixel coordinates in the physical world.

[0094] After establishing the initial physical entity reference 3D spatial coordinates, the pseudo-defect stripping dynamic control module reads the transfer step size parameters generated by the system between the trigger times of each single-frame image. Along the positive direction vector axis of the conveyor belt mechanical transport set within the world reference coordinate system, the transfer step size parameters read each time are used as the absolute increment of the rigid body's linear physical displacement and sequentially superimposed onto the components of the initial physical entity reference 3D spatial coordinates. Through spatial vector addition calculations, the expected 3D physical spatial coordinate set of the assumed entity defect, as it moves with the conveyor belt, will arrive at subsequent exposure times is derived.

[0095] After completing the three-dimensional spatial coordinate calculations, the pseudo-defect stripping dynamic control module uses an external pose parameter transformation matrix to perform matrix multiplication on the expected three-dimensional physical spatial coordinate set, transforming it back to the camera's internal spatial coordinate system. Then, using the internal optical parameter matrix, perspective division and forward perspective projection matrix operations are performed to map and reduce the three-dimensional spatial coordinates to the two-dimensional image plane pixel coordinate system, extracting the expected theoretical pixel coordinate set of the highlighted abnormal region in the subsequent image acquisition stage. Polynomial spline smoothing mathematical interpolation and connection operations are then performed on each discrete coordinate data within the theoretical pixel coordinate set to construct a theoretical translation trajectory exhibiting continuous, uninterrupted tangent lines.

[0096] Step S10: Use the dynamic time warping algorithm to measure the cumulative Euclidean distance after aligning the time axes of the two sets of trajectories.

[0097] Within the working environment, due to response jitter between light source triggering and camera exposure, the projected coordinate points in continuous images do not possess a strictly equidistant, discrete distribution pattern over time. Directly using a sequence number alignment and absolute difference algorithm to calculate the distance will result in time axis misalignment errors. The pseudo-defect stripping dynamic control module employs a dynamic time warping algorithm to calculate the cumulative Euclidean distance between two sets of trajectory coordinate point pairs, achieving optimal matching through elastic scaling in the time dimension.

[0098] The algorithm logic is as follows: It extracts all actual pixel coordinates arranged in chronological order from the actual movement trajectory to form the measured sequence; it also extracts all expected pixel coordinates from the theoretical translation trajectory to form the baseline sequence. The control processing device 200 allocates a two-dimensional planar cost search comparison matrix in the system's dynamic random access memory space. The row dimension corresponds to the coordinate node index of the measured sequence, and the column dimension corresponds to the coordinate node index of the baseline sequence. The system sets the element at the top-left origin position to zero and fills the remaining node elements in the first row and first column (excluding the origin) with a set infinite constant parameter to construct a boundary blocking penalty.

[0099] Initiate a nested loop to traverse the inner non-boundary operation nodes. For specific x-axis and y-axis index nodes, use the Euclidean distance formula to calculate the square of the difference between the actual pixel coordinate x-coordinate value and the expected pixel coordinate x-coordinate value, add the square of the difference between the two points' y-coordinate values, and perform a square root operation to obtain the absolute value cost variable of the spatial span of the corresponding node, which is temporarily stored in the arithmetic register as the local execution distance cost.

[0100] Next, the historical global cumulative path cost values ​​already accumulated and stored within the adjacent left horizontal node, directly above vertical node, and upper left diagonal node in the two-dimensional plane cost search comparison matrix are read. A comparator is invoked to select the priority cumulative path cost component with the smallest cost value among the three adjacent nodes. This priority cumulative path cost component is then numerically added to the temporarily stored local execution distance cost. The summation result is written to the memory address corresponding to the current coordinate index node, serving as the updated global cumulative distance cost parameter.

[0101] After the double traversal, the final numerical parameter stored inside the index node of the diagonal vertex at the end of the matrix represents the absolute sum of the cumulative Euclidean distances after the completion of the time-dimension elastic alignment operation. The system divides the absolute sum of the cumulative Euclidean distances by the actual number of nodes traversed in the regular matching optimization backtracking path, performs a numerical standardization and smooth convergence operation, and generates the final cumulative Euclidean distance scalar for judgment.

[0102] Step S11: Determine whether to eliminate optical false defects based on the calculated degree of deviation.

[0103] The control processing device 200 has pre-written tolerance distance parameters in its memory. When the deviation between the actual movement trajectory and the theoretical translation trajectory exceeds the set threshold, that is, when the cumulative Euclidean distance scalar after standardization is greater than the tolerance distance parameter, the pseudo-defect removal dynamic control module determines that the bright abnormal area is an optical pseudo-defect caused by surface reflection and removes it.

[0104] At the physical judgment level, the cumulative Euclidean distance scalar exceeding the parameter confirms that the bright reflective patches obtained by visual extraction did not follow the linear geometric rigid body displacement law of the conveyor belt's mechanical conveying traction direction. Instead, they exhibited a nonlinear optical slippage phenomenon on the specular reflective surface of the coated workpiece due to the alternation of multiple light source orientations. The system identified the corresponding bright abnormal area in spatial location as an optical false defect. For the alarm marker object within the system, the false defect removal dynamic control module triggered a memory release and address pointer cleanup interrupt program. The instruction operation required the central processing unit to overwrite and zero out the corresponding data structure in the current detection sequence linked list, performing a digital zeroing and clearing operation to remove false alarm source information in the data logic judgment link and prevent external hardware facilities from issuing erroneous physical sorting actions.

[0105] See attached document Figure 4 As shown, it also includes:

[0106] Step S12: Confirm the actual defect attributes and issue a closed-loop digital feedback control adjustment communication message command.

[0107] When the deviation does not exceed the set threshold, i.e., when the cumulative Euclidean distance scalar calculated by the dynamic time warping algorithm is less than or equal to the tolerance distance parameter, the pseudo-defect stripping dynamic control module determines that the deviation remains within the judgment limit parameter range and determines that the highlighted abnormal area is a real physical defect. The logical judgment process verifies that the actual offset trajectory data of the highlighted patch in the time-series image meets the spatial trajectory alignment judgment with the underlying mechanical physical displacement data driven by the servo motor hardware. The system confirms that the corresponding target is a real physical entity fixed within the surface material of the coated workpiece.

[0108] Based on real physical defects, the system generates process compensation parameters and feeds them back to the vacuum coating machine, establishing a digital closed loop for correcting the underlying core manufacturing parameters.

[0109] The control and processing equipment 200 internally houses a subroutine module for in-depth analysis of the microscopic morphology of defects. This subroutine module extracts multi-dimensional physical geometric digital structure, scale, contour characteristics, and corresponding optical grayscale contrast quantification evaluation indicators from pixel regions confirmed as genuine physical defects. The extracted and calculated output parameter set includes: the pixel length of the perimeter of the physical defect's outer closed envelope, the pixel count of the physical defect's boundary envelope space, the roundness characteristic value of the defect region, and the grayscale contrast difference between the arithmetic mean absolute brightness value of the pixels within the defect region and the arithmetic mean absolute brightness value of the outer healthy background region.

[0110] During system processing, the calculation logic for the roundness feature value of the defect region is determined as follows: multiply the number of pixels of the physical defect boundary envelope area by a preset shape coefficient (such as a constant of 4) and pi. Then, divide the resulting product by a scalar parameter multiplied by the square of the pixel length of the physical defect's outer closed envelope perimeter.

[0111] The control and processing device 200 inputs the extracted and calculated multi-dimensional fused fingerprint feature array into a pre-loaded and fixed process mapping lookup tree architecture module in the motherboard's non-volatile read-only space. This architecture is generated based on the clustering and labeling of a large number of pre-configured physical entity samples with definite defect causes.

[0112] When the system searches and finds that the roundness feature value of the extracted defect area meets the constant parameter condition of the approximation calibration as the lower limit, and the grayscale step span contrast difference variable parameter has the physical property of attenuation towards lower grayscale, the process mapping lookup tree outputs a diagnostic judgment conclusion confirming that the physical source of the real physical defect is a macroscopic metal molten particle sputtering and falling contamination fault caused by partial discharge of the cathode of the metal sputtering target inside the vacuum production chamber. If the comprehensive judgment determines that the real physical defect exhibits a narrow, long, strip-shaped, large-area peeling physical appearance outline, and the photometric readings collected in the internal area show the original color reflectivity characteristics of the underlying plastic material resin, the process mapping lookup tree logically determines that the physical defect is a peeling phenomenon caused by weak adhesion of the metal film deposition surface due to an imbalance in the distribution of protective process gas input in the chamber or insufficient energy supply for pretreatment ion bombardment and polishing.

[0113] Based on the physical causes of underlying manufacturing defects identified through deconstruction analysis, the integrated digital computing control loop within the control processing equipment 200 generates process compensation parameters with targeted parameter correction and adjustment functions. For macroscopic particle sputtering failures, the control processing equipment 200 generates a combination of digital adjustment compensation instructions to lower the standard operating value of the cathode bias voltage output from the external sputtering power supply and to increase the input flow rate limit of the protective gas mass flow controller in the production chamber. For thin film peeling failures, the control processing equipment 200 generates a combination of digital adjustment compensation instructions to increase the upper limit threshold of the preheating temperature of the infrared heating tube array network on the chassis substrate of the vacuum production chamber and to extend the specified duration of plasma bombardment surface cleaning from the anode beam source.

[0114] After the instruction parameter value message is assembled, the control processing device 200 calls the industrial local area Ethernet communication hardware controller link module integrated in the computer motherboard architecture area. The system utilizes the transmission control data protocol specification system and the Internet routing network layer protocol stack mechanism to encapsulate the process compensation parameters, including the operation content of digital offset setpoint variables, into standard control bus communication messages. The standard control bus communication message data packets are sent to the Ethernet communication receiving port of the main control programmable logic controller built into the vacuum coating machine through an externally deployed twisted-pair shielded communication physical link. The main control programmable logic controller parses and extracts the message data packet content, and according to the included register offset addressing variable instructions, overwrites the control reference analog voltage parameter output by the corresponding digital-to-analog signal conversion hardware module, dynamically changing the operating power load output limit of the underlying electrical actuators of the vacuum coating machine in real time.

[0115] Step S13: The pulse shift queue is used to drive the sorting execution unit to complete the physical separation of waste products.

[0116] In the scrap removal process, the control processing equipment 200 is externally electrically controlled by a sorting execution unit. After identifying a real physical defect, the control processing equipment 200 classifies the coated workpiece as defective based on the morphological attributes of the real physical defect and sends a sorting action signal to the sorting execution unit. The sorting execution unit receives the sorting action signal and applies a physical offset thrust to the coated workpiece containing the physical defect, moving it out of the main production conveyor.

[0117] To ensure precise synchronization and matching between the visual capture and judgment execution time and the mechanical gear conveying delay in spatial positioning, the control processing device 200 establishes and maintains an asynchronous first-in-first-out buffer shift register data stack queue in the dynamic high-speed working memory area, based on the absolute count pulse triggering pointer action of the encoder unit hardware. When the system's operational logic identifies a real physical defect on the surface of a specific coated workpiece, the control processing device 200's interrupt response program captures and freezes the total number of digital values ​​returned by the encoder unit's counter accumulator in the current state, assigning and saving this value as the initial pulse total value variable.

[0118] The control processing device 200 reads the pre-written static spatial physical layout distance span value. This static spatial physical layout distance span value reflects the difference in linear displacement physical length between the vertical coordinates of the optical axis of the main viewing center of the area array industrial camera and the coordinates of the center of the airflow outlet at the pneumatic nozzle of the sorting execution unit. The control processing device 200 performs a floating-point division operation, dividing the acquired linear displacement physical length difference of the fixed assembly spacing by the pulse equivalent conversion coefficient value, and then performs a downward hardware truncation and rounding operation on the quotient. It then calculates the total number of inherent mechanical travel delay compensation pulses required for the workpiece to be rejected to move from the camera capture start point to the tail-end side rejection operation station interval.

[0119] The control processing device 200 uses accumulator logic to add the temporarily stored total value of the initial pulse to the calculated total value of the inherent mechanical travel delay compensation pulses. This calculates the expected threshold limit parameter for the specific target trigger pulse that is expected to trigger the hardware action. The calculated expected threshold limit parameter for the specific target trigger pulse is pushed onto the system main control stack and written to the end of the asynchronous first-in-first-out buffer shift register data stack queue, stored at the physical address where the allocation pointer is located, for later verification.

[0120] During the continuous operation of the conveyor belt, which carries and pushes materials forward, the background polling service program of the 200 core configuration of the control processing equipment independently occupies a dedicated channel to continuously and at high speed read and refresh the accumulated real-time pulse dynamic readings returned by the encoder unit. When the system verifies that the acquired absolute cumulative sum of pulses matches the expected threshold limit parameter of the specific target trigger pulse pointed to by the read pointer at the front pop port of the asynchronous first-in-first-out buffer shift register data stack queue, and determines that the defective workpiece being tracked has entered the pneumatic sorting execution area, the system concludes that the defective workpiece has entered the pneumatic sorting execution area.

[0121] The moment the alignment condition is met, the control processing device 200 uses its internal hardware registers to drive the general-purpose input / output logic pins of the digital peripheral device to flip levels, sending a sorting action signal to the sorting execution unit. The sorting execution unit's physical structure includes a direct-acting solenoid valve control body with a preset diameter and a high-velocity directional pneumatic jet high-pressure exhaust nozzle. Upon receiving the sorting action signal and triggering the flip edge, the direct-acting solenoid valve control body overcomes spring resistance within its inherent mechanical response hysteresis time to open the internal high-pressure compressed airflow exhaust channel. The high-velocity directional pneumatic jet high-pressure exhaust nozzle releases a controlled high-speed airflow jet, which acts on the side of the target object with physical defects through jet thrust. The airflow jet applies lateral thrust, driving the specific defective target object to change its original direction, causing lateral physical displacement and deviating from its original forward trajectory along the main conveyor belt. It then falls into a waste collection trough deployed beside the main line.

[0122] After the physical sorting and rejection process is completed, the control processing equipment 200 automatically operates the asynchronous first-in-first-out buffer shift register data stack queue head pointer to find the current relative position according to the memory stack queue maintenance and management rules. Logically, the system cache memory space and physical operation area previously occupied by the corresponding instruction task are cleared and released from the software memory structure. The entire spatial closed-loop process effectively avoids the time delay error and action deviation caused by purely relying on clock time for estimation and judgment due to the specially designed register queue space hysteresis pulse compensation synchronous matching calculation verification mechanism that relies on real-time tracking and comparison of the high-frequency pulse counting digital scalar of the underlying peripherals. The aforementioned error and deviation are caused by the decrease in traction motor speed, which is due to the resistance change caused by the irregular fluctuation of the conveyor belt's load weight.

[0123] Furthermore, based on the same inventive concept, embodiments of the present invention also provide an online detection method for surface defects in vacuum coating based on industrial vision. This method applies to and relies on the aforementioned online detection system for operation.

[0124] The method mainly includes the following steps:

[0125] The process involves: acquiring the conveying step length parameters of the coated workpiece on the conveyor belt; issuing a high-frequency alternating gating command to the vision acquisition device based on the conveying step length parameters to acquire multiple frames of surface-related images under different illumination angles; extracting the high-brightness abnormal areas within the multiple frames of surface-related images and calculating their actual movement trajectories; estimating the theoretical translation trajectory based on the conveying step length parameters; using a dynamic time warping algorithm to measure the cumulative Euclidean distance between the actual movement trajectory and the theoretical translation trajectory after aligning their time axes; and determining whether to eliminate optical pseudo-defects or confirm real physical defects based on the degree of deviation of the cumulative Euclidean distance and issuing a closed-loop control command to perform waste removal.

[0126] It is understood that the specific data flow logic, algorithm derivation process, and corresponding technical effects in the above-described detection method embodiments correspond one-to-one with steps S1 to S13 in the aforementioned embodiment of the online detection system for vacuum coating surface defects based on industrial vision. To avoid redundancy in the specification, these details will not be elaborated upon here.

[0127] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An online detection system for surface defects in vacuum coating based on industrial vision, comprising a vision acquisition device and a control and processing device; the vision acquisition device is used to perform image acquisition operations on a coated workpiece moving on a conveyor belt and output a sequence of images to be detected; Its features are, The control and processing equipment is equipped with a dynamic control module for removing false defects. The pseudo-defect stripping dynamic control module is used to obtain the conveying step length parameters of the coated workpiece on the conveyor belt. The pseudo-defect stripping dynamic control module sends a high-frequency alternating gating command to the visual acquisition device based on the transmission step size parameter, and controls the visual acquisition device to acquire multiple frames of surface association images under different illumination angles. The pseudo-defect stripping dynamic control module extracts the bright abnormal regions contained within the multi-frame surface association images and calculates the actual movement trajectory of the bright abnormal regions in the multi-frame surface association images. The pseudo-defect stripping dynamic control module compares the actual movement trajectory with the theoretical translation trajectory calculated based on the transmission step size parameter. When the deviation between the actual movement trajectory and the theoretical translation trajectory exceeds a set threshold, the pseudo-defect stripping dynamic control module determines that the high-brightness abnormal area is an optical pseudo-defect caused by surface reflection and removes it. When the deviation does not exceed the set threshold, the pseudo-defect stripping dynamic control module determines that the high-brightness abnormal area is a real physical defect and feeds back the real physical defect generation process compensation parameters to the vacuum coating machine.

2. The online detection system for vacuum coating surface defects based on industrial vision according to claim 1, characterized in that, The visual acquisition device includes an area array industrial camera and a multi-channel ring light source; The pseudo-defect stripping dynamic control module divides the multi-channel ring light source into multiple sub-light sources; The high-frequency alternating gating instruction is used to control the various sub-light sources inside the multi-channel ring light source to take turns lighting up. The pseudo-defect stripping dynamic control module controls the area array industrial camera to synchronously perform exposure operations when each of the sub-light sources is lit, so as to form the multi-frame surface-related images.

3. The online detection system for vacuum coating surface defects based on industrial vision according to claim 1, characterized in that, The control and processing device is also equipped with a basic image preprocessing module. The basic image preprocessing module uses an edge detection operator to extract the outer contour boundary of the coated workpiece in the multi-frame surface association image, and performs a background segmentation operation based on the outer contour boundary to locate the high-brightness abnormal area within the coated workpiece.

4. The online detection system for vacuum coating surface defects based on industrial vision according to claim 1, characterized in that, The conveyor belt is equipped with a servo drive shaft; The online detection system for surface defects in vacuum coating based on industrial vision also includes an encoder unit; The encoder unit is mounted on the end of the servo drive shaft; The pseudo-defect stripping dynamic control module is communicatively connected to the encoder unit to read the physical displacement pulse signal of the conveyor belt and convert the physical displacement pulse signal into the conveying step length parameter.

5. The online detection system for vacuum coating surface defects based on industrial vision according to claim 1, characterized in that, The control and processing equipment is connected to a sorting execution unit; After determining the actual physical defect, the control and processing equipment classifies the coated workpiece as defective based on the morphological attributes of the actual physical defect and sends a sorting action signal to the sorting execution unit. The sorting execution unit receives the sorting action signal and moves the corresponding coated workpiece out of the main production line.

6. The online detection system for vacuum coating surface defects based on industrial vision according to claim 1, characterized in that, When calculating the theoretical translation trajectory, the pseudo-defect stripping dynamic control module extracts the center pixel coordinates of the bright abnormal region in the initial frame image of the multi-frame surface-related images, and combines the transmission step size parameter and the camera calibration matrix to obtain the expected set of theoretical pixel coordinates of the bright abnormal region in subsequent frames, and fits the set of theoretical pixel coordinates to the theoretical translation trajectory.

7. The online detection system for vacuum coating surface defects based on industrial vision according to claim 6, characterized in that, When calculating the degree of deviation between the actual movement trajectory and the theoretical translation trajectory, the pseudo-defect stripping dynamic control module uses a dynamic time warping algorithm to quantify the cumulative Euclidean distance between the coordinate point pairs of the actual movement trajectory and the theoretical translation trajectory. When the cumulative Euclidean distance is greater than the distance limit, it is determined that the degree of deviation exceeds the set threshold.

8. The online detection system for vacuum coating surface defects based on industrial vision according to claim 1, characterized in that, The control and processing device is also equipped with a region division module; The region division module calculates the surface reflectance gradient distribution matrix based on the surface grayscale image under natural light, and divides the surface of the coated workpiece into diffuse reflection region, specular reflection region and interference edge region according to the surface reflectance gradient distribution matrix. The control and processing device activates high-frequency alternating gating commands at different angles for the specular reflection area, and adopts full-brightness constant-on illumination operation for the diffuse reflection area.

9. The online detection system for vacuum coating surface defects based on industrial vision according to claim 8, characterized in that, The control processing device extracts the local color saturation variance for the interference edge region; When the local color saturation variance is within the coating interference color difference threshold range and there are no abrupt pixel edges, the control processing device determines that the corresponding visual difference is a color interference artifact and suppresses the alarm action.

10. An online detection method for surface defects in vacuum coating based on industrial vision, characterized in that, The execution steps include the following: Obtain the conveying step length parameters of the coated workpiece on the conveyor belt; Based on the transmission step size parameter, a high-frequency alternating gating command is issued to the visual acquisition device; Control the visual acquisition device to acquire multiple frames of surface-related images under different illumination angles; Extract the highlighted abnormal regions contained within the multi-frame surface association images; Calculate the actual movement trajectory of the highlighted abnormal region in the multi-frame surface association image; The actual movement trajectory is compared with the theoretical translation trajectory calculated based on the transmission step size parameter; When the deviation between the actual movement trajectory and the theoretical translation trajectory exceeds a set threshold, the high-brightness abnormal area is determined to be an optical pseudo-defect caused by surface reflection and is removed. When the deviation does not exceed the set threshold, the high-brightness abnormal area is determined to be a real physical defect, and the process compensation parameters are fed back to the vacuum coating machine based on the real physical defect.

Images