Safety belt surface defect detection method and system based on industrial vision

Abstract

The present application relates to the technical field of industrial vision detection and intelligent manufacturing, in particular to a safety belt surface defect detection method and system based on industrial vision; comprising: acquiring the torque current signal of a winding servo motor and the longitudinal position of a safety belt; extracting the tension fluctuation component, and calculating the lateral displacement and lateral speed of the detection area according to the pre-acquired delay parameter and the pre-acquired conversion coefficient; judging the exposure condition according to the lateral speed, the longitudinal position and the speed threshold value, and triggering the industrial camera; recording the longitudinal position corresponding to the exposure, and performing longitudinal resampling on the original image to generate a reconstructed image; identifying the surface defects on the reconstructed image and extracting the image edge track, and correcting the delay parameter and the conversion coefficient according to the deviation between the actual lateral displacement and the calculated lateral displacement; the method is beneficial to reduce the edge blur caused by lateral swing and improve the detection clarity of small defects.

Description

Technical Field

[0001] This invention relates to the field of industrial vision inspection and intelligent manufacturing technology, specifically to a method and system for detecting surface defects in seat belts based on industrial vision. Background Technology

[0002] With the increasing demand for continuous production and online quality inspection of seat belts, real-time detection of surface defects of seat belts based on industrial vision has become an important technical means in the production process. Seat belts are usually in a continuous operation state during roll-to-roll conveying, and their surfaces, especially the edges, are prone to defects such as burrs, broken yarns, and snags. How to balance detection efficiency and imaging clarity during operation has become an important problem that needs to be solved in the field of online inspection of seat belts. Traditional visual inspection of seat belt surfaces currently relies mainly on the following methods: triggering industrial cameras to acquire images according to a fixed rhythm, simply shortening the exposure time to reduce motion blur, increasing the lighting intensity or using strobe to enhance the imaging effect, and limiting the swing of the seat belt through pressure rollers or guide structures. However, existing methods such as fixed-cycle acquisition, shortened exposure time, enhanced illumination, and mechanical limiting have the following drawbacks: fixed-cycle triggering is difficult to adapt to lateral swing changes during seat belt operation, and edge blurring is likely to occur when the lateral movement exceeds the preset speed threshold; simply shortening the exposure time will reduce image brightness and signal-to-noise ratio, especially for dark seat belts, reducing the image contrast of small defects; increasing illumination intensity or using higher power strobe will lead to increased system power consumption and reduced system stability; mechanical clamping or limiting methods may increase structural complexity and make it difficult to balance image clarity and longitudinal detection coverage under continuous operation. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for detecting surface defects in seat belts based on industrial vision, addressing the following technical problems: Existing visual inspection technologies for seat belt surfaces have limitations in adapting to lateral seat belt movement and ensuring image clarity and longitudinal coverage during continuous operation. This invention provides a detection method and system capable of dynamically adjusting exposure triggering to guarantee image quality and accurately identify defects. To achieve the above objective, this invention adopts the following technical solution: A method for detecting surface defects in seat belts based on industrial vision includes the following steps: acquiring the torque current signal of the seat belt winding servo motor and the longitudinal position of the seat belt; extracting the tension fluctuation component from the torque current signal, and calculating the lateral displacement and lateral velocity of the seat belt in the detection area based on pre-acquired delay parameters and pre-acquired conversion coefficients, wherein the delay parameters are used to characterize the lateral displacement response time difference of the seat belt corresponding to the tension fluctuation component, and the conversion coefficients are used to characterize the conversion relationship between the tension fluctuation component and the lateral displacement; Based on the lateral speed, the longitudinal position, and the preset speed threshold, it is determined whether the industrial camera exposure conditions are met. If they are met, an exposure trigger signal is generated. The longitudinal position of the seat belt corresponding to the exposure trigger signal is obtained, and the industrial camera is controlled to acquire the original image of the seat belt in response to the exposure trigger signal. The original image of the seat belt is longitudinally resampled to generate a reconstructed image with equal spacing. On the reconstructed image, surface defects of the seat belt are identified, the image edge trajectory of the seat belt is extracted, and the delay parameter and the conversion coefficient are corrected based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement calculated in the previous step.

[0004] Optionally, the process of calculating the lateral displacement and lateral velocity of the seat belt in the detection area includes: preprocessing the torque current signal to obtain the tension fluctuation component; calculating the lateral displacement based on the tension fluctuation component, the delay parameter, and the conversion factor, wherein the lateral displacement is the displacement obtained by aligning the tension fluctuation component according to the delay parameter and then converting it according to the conversion factor; and calculating the lateral velocity based on the change in lateral displacement at adjacent times and the corresponding time interval.

[0005] Optionally, the process of determining whether the exposure conditions of the industrial camera are met includes: generating a pre-trigger time when the longitudinal position reaches the set longitudinal sampling interval; and determining whether the lateral velocity corresponding to the pre-trigger time is less than or equal to the velocity threshold. If the speed is less than or equal to the speed threshold, the exposure conditions of the industrial camera are determined to be met, and the exposure trigger signal is issued at the pre-trigger time. If the speed threshold is greater than the speed threshold, it is determined whether the longitudinal distance between two adjacent industrial camera exposures exceeds the preset maximum skip distance. If the speed threshold is less than or equal to the maximum skip distance, the current pre-trigger time is canceled. If the speed threshold is greater than the maximum skip distance, the exposure trigger signal is issued at the moment with the smallest lateral speed among multiple pre-trigger times within a preset range starting from the current pre-trigger time.

[0006] Optionally, the process of selecting the moment with the smallest lateral velocity among multiple pre-trigger moments within a preset range to issue the exposure trigger signal includes: selecting the moment with the smallest lateral velocity among multiple pre-trigger moments within the current pre-trigger moment after reaching the maximum skip interval and within the subsequent preset time period to issue the exposure trigger signal.

[0007] Optionally, the process of longitudinally resampling the original seat belt image includes: obtaining the longitudinal position corresponding to each row of pixels in the original seat belt image according to the correspondence between the acquisition time sequence of the industrial camera and the longitudinal position of the seat belt; generating a longitudinal grid according to the set resampling interval; and interpolating and rearranging the original seat belt image according to the longitudinal grid to generate the reconstructed image.

[0008] Optionally, the process of identifying seat belt surface defects on the reconstructed image includes: calculating the gradient magnitude and grayscale difference of the seat belt edge region in the reconstructed image; multiplying the gradient magnitude and the grayscale difference by preset edge weights respectively and then summing them by weight to obtain edge feature values; performing connected component analysis on pixel regions whose edge feature values ​​are greater than or equal to preset edge feature thresholds to obtain defect connected regions, and determining pixel regions whose edge feature values ​​are less than the edge feature thresholds as normal background regions; when the size range of the defect connected region in the vertical and horizontal directions meets preset defect determination size conditions, the region is determined to be a seat belt surface defect region.

[0009] Optionally, the process of correcting the delay parameter and the conversion coefficient based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement includes: calculating the horizontal coordinate of the center of the seat belt edge in a continuous image row to form the image edge trajectory, and calculating the fluctuation amount of the image edge trajectory; The delay parameter is updated based on the temporal offset that has the greatest correlation with the lateral displacement of the image edge trajectory; the conversion coefficient is updated based on the proportional relationship between the lateral relative displacement of the image edge trajectory relative to the reference image row and the lateral displacement, wherein the reference image row is a pre-selected row of images in the reconstructed image.

[0010] An industrial vision-based seat belt surface defect detection system for implementing the above method includes an acquisition module, a calculation module, a triggering module, a resampling module, an identification module, a correction module, and an output module. The acquisition module includes a current detection unit for acquiring torque current signals, a position detection unit for acquiring longitudinal position, and an industrial camera for acquiring original images of the seat belt. It is used to acquire the torque current signal of the seat belt winding servo motor and the longitudinal position of the seat belt, and to acquire original images of the seat belt. The calculation module is used to extract the tension fluctuation component in the torque current signal, and calculate the lateral displacement and lateral velocity of the seat belt in the detection area according to the pre-acquired delay parameters and pre-acquired conversion coefficients; the triggering module is used to determine whether the industrial camera exposure conditions are met according to the lateral velocity, the longitudinal position and the velocity threshold. If they are met, an exposure trigger signal is generated. The resampling module is used to obtain the longitudinal position of the seat belt corresponding to the exposure trigger signal, and respond to the exposure trigger signal to control the industrial camera to acquire the original image of the seat belt, perform longitudinal resampling on the original image of the seat belt, and generate a reconstructed image with equal spacing. The identification module is used to identify surface defects of the seat belt on the reconstructed image and extract the image edge trajectory of the seat belt; the correction module is used to correct the delay parameter and the conversion coefficient according to the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement; the output module is used to output the judgment result of the surface defect of the seat belt and the reconstructed image.

[0011] Compared with the prior art, the present invention has the following beneficial effects: This invention calculates the lateral speed of the seat belt by extracting the tension fluctuation component from the torque current signal, and triggers the exposure of the industrial camera when the lateral speed is less than or equal to the speed threshold. This method avoids image blurring caused by lateral swing without adding mechanical limiters or simply shortening the exposure time. In addition, by combining longitudinal resampling to generate equally spaced reconstructed images, it effectively balances image clarity and longitudinal detection coverage under continuous operation. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] Figure 1 This is a schematic flowchart of a seat belt surface defect detection method based on industrial vision provided in an embodiment of this application.

[0014] Figure 2 This is a schematic diagram of a seat belt surface defect detection system based on industrial vision provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0016] Reference Figure 1 A method for detecting surface defects in seat belts based on industrial vision includes the following steps: Step 1: Obtain the torque current signal of the seat belt winding servo motor and the longitudinal position of the seat belt; Step 2: Extract the tension fluctuation component from the torque current signal, and calculate the lateral displacement and lateral velocity of the seat belt in the detection area based on the pre-acquired delay parameters and pre-acquired conversion coefficients; Step 3: Based on the lateral velocity, longitudinal position, and preset velocity threshold, determine whether the industrial camera exposure conditions are met; if so, generate an exposure trigger signal. Step 4: Obtain the longitudinal position of the seat belt corresponding to the exposure trigger signal, and control the industrial camera to acquire the original image of the seat belt in response to the exposure trigger signal. Perform longitudinal resampling on the original image of the seat belt to generate a reconstructed image with equal spacing. Step 5: Identify surface defects of the seat belt on the reconstructed image, extract the image edge trajectory of the seat belt, and correct the delay parameter and conversion coefficient based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement calculated in Step 2.

[0017] Step 2 calculates the lateral displacement and lateral velocity of the seat belt in the detection area, including: The torque current signal is preprocessed to obtain the tension fluctuation component; The lateral displacement is calculated based on the tension fluctuation components, delay parameters, and conversion factors. The lateral displacement is the displacement obtained by aligning the tension fluctuation components according to the delay parameters and then converting them according to the conversion factors. The lateral velocity is calculated based on the change in lateral displacement at adjacent moments and the corresponding time interval.

[0018] Step 3 determines whether the industrial camera exposure conditions are met, including: When the longitudinal position reaches the set longitudinal sampling interval, a pre-trigger time is generated; Determine whether the lateral velocity at the pre-trigger moment is less than or equal to the velocity threshold; If it is less than or equal to, it is determined that the exposure conditions of the industrial camera are met, and an exposure trigger signal is issued at the pre-trigger time; If the speed threshold is exceeded, determine whether the longitudinal distance between two adjacent industrial camera exposures exceeds the set maximum skip distance. If the value is less than or equal to the maximum skip interval, then the current pre-triggering time is cancelled; If the distance is greater than the maximum skip interval, then among multiple pre-trigger times within a preset range starting from the current pre-trigger time, the time with the smallest lateral velocity will be selected to issue the exposure trigger signal.

[0019] Among multiple pre-trigger times within a preset range, the moment with the minimum lateral velocity is selected to issue the exposure trigger signal, including: Among the current pre-trigger time after reaching the maximum skip interval and multiple pre-trigger times within the subsequent preset time period, the moment with the smallest lateral velocity is selected to issue the exposure trigger signal.

[0020] Step 4 involves longitudinal resampling of the original seatbelt image, including: Based on the correspondence between the acquisition time sequence of the industrial camera and the longitudinal position of the seat belt, the longitudinal position corresponding to each row of pixels in the original image of the seat belt is obtained; A vertical grid is generated based on the set resampling interval; The original image of the seat belt is interpolated and rearranged according to the vertical grid to generate a reconstructed image.

[0021] This implementation method is applicable to a continuous roll-to-roll safety belt inspection production line; the safety belt is continuously conveyed along the longitudinal direction, and the winding servo motor drives the winding shaft to complete the winding at the downstream end; the inspection area is equipped with an industrial camera, lens and lighting unit; the controller is connected to the servo driver, position detection unit and industrial camera respectively, and is used to receive motor feedback signals, length position signals and output industrial camera trigger signals; In this embodiment, the torque current signal refers to the quadrature-axis current feedback signal of the winding servo motor; this signal is proportional to the motor output torque, and therefore can reflect changes in the winding load; the longitudinal position refers to the cumulative length coordinate of the seat belt along the running direction, which is calculated by encoder pulses; Lateral displacement refers to the amount of swing of the seat belt relative to the reference lateral position within the detection area. Lateral velocity refers to how fast this amount of swing changes within adjacent sampling periods. The delay parameter is used to represent the propagation lag that is only reflected in the image after the tension fluctuation at the winding end is transmitted to the detection area. The conversion factor is used to convert the effective fluctuation amount in the current fluctuation into the amount of lateral displacement within the detection area. In step 1, the servo driver continuously outputs motor current feedback, and the position detection unit continuously outputs encoder cumulative pulses; the controller reads the current current value and the current position value in each control cycle, and stores them in pairs according to the same control cycle; A fixed-length circular queue is pre-allocated in the controller's memory as a data link; each element of the circular queue is defined as a data structure, which contains at least the following four core fields: a physical timestamp that accurately records the data sampling time with microsecond precision, the cross-axis current feedback value, the cumulative longitudinal position coordinates, and a trigger status flag used to mark whether exposure has occurred at that moment; The controller uses an incrementing write pointer to push the structure of each cycle into the circular queue in sequence. When the write pointer reaches the end of the queue, it automatically wraps around and overwrites the oldest data. The current fluctuation used to calculate the lateral displacement and the longitudinal position used to trigger the exposure are extracted by querying the corresponding timestamp within the same data link. In step 2, the controller first preprocesses the torque current signal; the preprocessing of the torque current signal extracts the AC component corresponding to the seat belt tension fluctuation; specifically, the quadrature-axis current is first converted into a torque signal based on the motor torque constant, then a local average value is calculated within a sliding window, and the current torque is subtracted from this local average value to remove the slow-changing baseline during the winding process. The time length of the sliding window is set to be greater than or equal to... ,in, This is the reference rotational frequency for the winding servo motor; The obtained torque fluctuation signal is then subjected to a bandpass filter. Specifically, the controller reads the current motor speed in real time to calculate the reference frequency, sets the lower cutoff frequency of the bandpass filter to 0.8 times the reference frequency, and sets the upper cutoff frequency to 1.2 times the highest relevant harmonic to be analyzed, thereby retaining the frequency band that is consistent with the winding frequency and its relevant harmonics, and filtering out low-frequency drift and high-frequency interference that deviate significantly from this frequency band; the signal obtained after bandpass filtering is used as the tension fluctuation component. The delay parameter characterizes the hysteresis of the seat belt's lateral sway after a change in tension; tension fluctuations propagate upwards along the belt in the form of elastic waves. The initial value of the delay parameter is composed of the wave propagation hysteresis time and the belt's dynamic response hysteresis time, and is dynamically identified during trial operation. Specifically, tension waves are tensile stress waves that propagate along the longitudinal direction of the belt. Therefore, the wave propagation lag time is equal to the belt length from the detection point to the winding end divided by the propagation speed of the longitudinal stress wave. This propagation speed is calculated by dividing the equivalent Young's modulus of the belt material by the square root of its density. The hysteresis time of the dynamic response is equal to the first-order system response time constant obtained by dividing the equivalent mass of the belt by the lateral damping coefficient; initial value of the delay parameter. The sum of these two lag times is calculated using the following formula: in, The distance from the detection point to the take-up end is a long distance. This is the equivalent Young's modulus of the belt material. For density, For the equivalent mass of the belt, The lateral damping coefficient is used for dynamic identification during trial operation. Specifically, considering that the tension fluctuation component belongs to the equal time interval sequence of the control period, while the historical lateral displacement sequence is not generated by equal frequency triggering, its sampling points are uneven on the time axis. The system extracts the physical timestamp recorded during each exposure acquisition and uses spline interpolation to perform one-dimensional time domain resampling of the historical lateral displacement sequence to align its time axis with the tension fluctuation component. The correlation coefficient between the aligned tension fluctuation component and the actual lateral displacement at different time shifts is calculated using the cross-correlation algorithm. The time shift corresponding to the maximum correlation coefficient is selected to update the delay parameter. The initial value of the conversion coefficient is set as follows: the maximum displacement amplitude from the peak to the trough of the image edge trajectory within a predetermined time period is extracted and divided by the maximum fluctuation difference of the tension fluctuation component within the same time period. The quotient is used as the initial value of the conversion coefficient. If a single proportional relationship is used, the lateral displacement is obtained by multiplying the tension fluctuation component after delay alignment by the conversion factor; if there is a significant phase deviation in the actual working condition, the in-phase component and the quadrature component can be used together for conversion. The calculation logic is to calculate the lateral displacement based on the tension fluctuation component, the delay parameter and the conversion factor. In terms of timing processing, the controller not only calculates the physical state at the current moment, but also performs phase extrapolation by extracting the tension fluctuation component with periodic characteristics that are strongly correlated with the winding frequency and related harmonics, combined with the delay parameter; so that the fluctuation component is aligned with the actual lateral oscillation in the detection area in the historical and current calculation sequences to complete the propagation timing; and then extrapolates to the future short time window using the periodic change law to generate the predicted fluctuation sequence. The specific extrapolation method is as follows: extract the sequence data of the aligned tension fluctuation component in the most recent one or more winding frequency cycles as the basic waveform, and cyclically copy the basic waveform according to the set prediction time window length according to the frequency cycle; The controller applies a weighted average-based cross-smoothing transition within the splicing interval. The smoothing mapping rule is as follows: assuming the splicing interval contains... The sampling point, for the th sampling point within the interval Each point is taken as the value of the amplitude of the corresponding point at the end of the previous cycle multiplied by itself. Multiply by the amplitude of the corresponding point at the beginning of the next cycle The sum of these two values ​​achieves linear cross-decay and increase, which is then spliced ​​to the current time to generate a smooth predictive fluctuation sequence covering the current and short future time windows. Alternatively, the Fast Fourier Transform can be used to extract the dominant frequency of the fundamental waveform and the amplitude and phase of its first N harmonic components (usually taken as 3 to balance fitting accuracy and real-time computation), constructing a model based on the reference bias. The Fourier series fitting function, composed of the superposition of sinusoidal functions of each harmonic order, has a time variable... The fitting formula is as follows: Wherein, the reference angular frequency corresponding to the winding frequency is denoted as , No. The amplitude and phase of the first harmonic component are denoted as follows: and Substitute future candidate times into the fitting function. The predicted fluctuation sequence is calculated from the corresponding time variables; the aligned and extrapolated sequence is converted into a predicted lateral displacement sequence that includes the current and future using conversion factors. The controller calculates the continuous displacement difference between adjacent predicted sampling points and divides it by the corresponding time interval to obtain a predicted lateral velocity sequence that includes the current and future candidate moments; the subsequent industrial camera triggering judgment is based on this predicted lateral velocity sequence. In step 3, the controller considers both longitudinal sampling requirements and lateral motion requirements. The longitudinal sampling interval refers to the interval that should be met between two adjacent ideal samples along the length of the seat belt, which is used to ensure that the entire seat belt is covered. Whenever the cumulative longitudinal position reaches the sampling interval, the controller generates a pre-trigger moment. At this time, the controller does not expose immediately, but continues to read the lateral velocity corresponding to the pre-trigger moment and compare it with the velocity threshold. The speed threshold is used to limit the extent of lateral movement of the safety belt during the exposure of an industrial camera. It is set based on the fact that the maximum amount of lateral blur that can be generated during the exposure does not exceed a certain proportion of the object resolution. Specifically, if the lateral object resolution of the industrial camera is the actual size corresponding to one pixel, and the exposure time is determined by the imaging signal-to-noise ratio requirement through pre-calibration, then when the lateral speed multiplied by the exposure time is not greater than the allowable blur amount, it can be considered that small edge defects will not be significantly widened due to lateral swing. The specific formula for calculating the speed threshold is as follows: For the detection of fine filamentous yarn breaks, the allowable ambiguity coefficient is set to 0.1 to 0.2; for larger burrs, the allowable ambiguity coefficient is set to 0.3 to 0.5. When the lateral velocity corresponding to the pre-trigger moment is less than or equal to the velocity threshold, the controller determines that the current moment meets the exposure conditions and immediately outputs an exposure trigger signal to the industrial camera; if the lateral velocity is greater than the velocity threshold, it means that the speed of the belt in the lateral direction exceeds the velocity threshold, and the edge defect trailing caused by direct exposure will occur. At this time, the controller will not trigger immediately, but will first determine whether the longitudinal distance between two adjacent industrial camera exposures exceeds the set maximum skip distance. The maximum skip gap is used to avoid longitudinal undersampling after skipping too many pre-triggering moments consecutively. The maximum skip gap is set under the condition that its value is no greater than half the length of the minimum target defect, so that any defect still has enough chance to be covered by at least two samplings in the longitudinal direction. If the longitudinal distance between two adjacent industrial camera exposures does not exceed the maximum skip gap, the current pre-triggering moment is canceled, and the judgment is made again after a new pre-triggering moment. If the longitudinal distance between two adjacent industrial camera exposures has exceeded the maximum skip gap, the compensation trigger selection process is entered. The compensation trigger selection process corresponds to the implementation details; that is, within the current pre-triggering moment after reaching the maximum skip interval and within the subsequent preset time period, the system does not actually physically transmit to all candidate moments, but the controller uses the aforementioned generated look-ahead predictive lateral velocity sequence to extract the predicted lateral velocity corresponding to multiple candidate pre-triggering moments within the future preset time period in the digital space; By comparing the predicted lateral velocity magnitudes of these candidate moments one by one, the moment with the smallest absolute value is selected, and an exposure trigger signal is issued at this optimal moment, or a corresponding delay is set to wait for triggering; the preset time period is set within the allowable range according to the requirement of no missed sampling, thereby taking into account both image clarity and vertical coverage; In step 4, the industrial camera acquires the original image of the seat belt after receiving the exposure trigger signal. Since the trigger is not a fixed uniform line frequency, but is determined by both the longitudinal position condition and the lateral speed condition, the actual longitudinal distance between adjacent image lines in the original image is usually not constant. In order to unify the longitudinal distance parameter, the controller records the longitudinal position of the seat belt corresponding to each exposure while outputting the exposure trigger signal, and establishes a one-to-one correspondence between the position and the original image line. During resampling, the controller first calculates the vertical position of each row of pixels in the original image based on the correspondence between the acquisition time sequence and the vertical position of the industrial camera; then it generates a uniform vertical grid according to the set resampling interval. The resampling interval is calculated and determined based on the vertical object resolution of the industrial camera or the ratio of the set minimum target defect vertical size. When a target grid position lies between the vertical positions of two adjacent rows of original data, the grayscale values ​​are interpolated and rearranged according to the vertical distance between these two rows. The specific mapping relationship is as follows: for the target grid position... Based on the position of the adjacent upper original image row and its corresponding grayscale and the position of the original image row below. and its corresponding grayscale Calculate the grayscale of the target grid pixel as This results in a reconstructed image with equal spacing; the row spacing of the image is converted into uniform spatial coordinates through interpolation rearrangement. In step 5, the controller performs surface defect identification and extracts image edge trajectories on the reconstructed image; the image edge trajectory refers to the trajectory formed by the change of the horizontal coordinate of the seat belt edge center with the vertical position in a continuous image row; since the image edge center will shift laterally when the seat belt swings laterally in the detection area, the edge trajectory itself can characterize the actual lateral displacement change in the detection area. The controller compares the predicted lateral displacement sequence obtained in step 2 with the actual lateral displacement sequence represented by the image edge trajectory; if the two are not completely consistent in time, it means that the delay parameter needs to be adjusted; if there is a stable proportional difference in the amplitude, it means that the conversion coefficient needs to be adjusted. In specific corrections, the controller uses the correlation between the image edge trajectory and the predicted lateral displacement as the basis for judgment, changes the delay parameter within the allowable range, and finds the delay value that makes the correlation larger; then, based on the ratio between the lateral relative displacement of the image edge trajectory relative to the reference image line and the predicted lateral displacement, the conversion coefficient is updated; the corrected parameters are reused in subsequent control cycles to calculate the lateral displacement and lateral velocity, so that the subsequent industrial camera triggering judgment is closer to the real motion state in the detection area. This embodiment completes the exposure when the lateral velocity is less than or equal to the velocity threshold, provided that the imaging signal-to-noise ratio is met. In one embodiment, step 5, identifying surface defects of the seatbelt on the reconstructed image, includes: Calculate the gradient magnitude and grayscale difference of the seatbelt edge region in the reconstructed image; The gradient magnitude and the gray-level difference are multiplied by preset edge weights respectively, and then the weighted sum is obtained to obtain the edge feature value. Connectivity analysis is performed on pixel regions with edge feature values ​​greater than or equal to a preset edge feature threshold to obtain defective connected regions, and pixel regions with edge feature values ​​less than the edge feature threshold are determined as normal background regions. When the dimensions of the connected defect area in both the longitudinal and transverse directions meet the preset defect determination size conditions, the area is determined to be a surface defect area of ​​the seat belt.

[0022] In step 5, the delay parameter and conversion factor are corrected based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement itself. This includes: Calculate the x-coordinate of the center of the seat belt edge in a continuous row of images to form the image edge trajectory, and calculate the fluctuation of the image edge trajectory; The delay parameters are updated based on the temporal offset that has the strongest correlation with the image edge trajectory and the lateral displacement. The conversion coefficients are updated based on the ratio between the lateral relative displacement of the image edge trajectory and the lateral displacement of the reference image row, where the reference image row is a pre-selected row of images in the reconstructed image.

[0023] This embodiment is used to illustrate the defect identification process and parameter correction process on the reconstructed image; the execution subject is still the controller, and the input object is the reconstructed image that has been resampled longitudinally; Before defect identification, the controller first locates the detection zone area where the left and right edges of the seat belt are located in the reconstructed image; the detection zone width is set according to the belt width calibration result and the preset expansion range of edge defects; Within the edge region, the controller calculates the gradient magnitude and grayscale difference for each pixel position. Specifically, the Sobel operator is used to obtain the grayscale gradient values ​​of the current pixel in the horizontal and vertical directions, and the square root of the sum of the squares of the two is taken as the gradient magnitude of the pixel. When calculating the grayscale difference, the absolute difference between the grayscale value of the current pixel and the local average grayscale value is taken as the grayscale difference. The gradient magnitude reflects the steepness of the grayscale change at that location, used to represent geometric changes such as edge breaks and burr protrusions; the grayscale difference reflects the difference between the grayscale of this pixel and the average grayscale of its neighborhood, used to represent the brightness deviation caused by changes in the weave structure at points of yarn breakage or filament pulling; the local average grayscale is centered on the current pixel. The value is obtained within a rectangular window, where the window side length is... The setting is based on the pixel span value obtained by dividing the actual physical size of a single weave texture cycle on the surface of the seat belt by the object resolution of the industrial camera, in order to extract the difference features between defects and the background and filter out global brightness fluctuation interference. After obtaining the gradient magnitude and grayscale difference, the controller performs normalization processing on the data. Considering that the local average grayscale value of the dark seat belt background may tend to zero, direct division will cause data overflow and calculation abnormalities. The normalization processing is as follows: the gradient magnitude is divided by the maximum gradient value allowed by the image, such as 255, to obtain the normalized gradient. When calculating the normalized grayscale difference, the grayscale difference is divided by the sum of the local average grayscale value of the seat belt background and the set non-zero smoothing constant. This non-zero smoothing constant is calibrated according to the dark current noise level of the industrial camera substrate, thereby avoiding division by zero abnormalities and numerical distortion in extremely dark backgrounds. The controller is multiplied by its corresponding weight and then summed using a weighted average. The calculation method involves multiplying the gradient weights. The product of the normalized gradient and the grayscale difference weight. The product of the edge feature value and the normalized gray level difference yields the edge feature value, i.e.: And satisfy ; The settings for gradient weights and grayscale difference weights depend on the type of target defect: when edge burrs are prominent, the value range of W1 is empirically determined to be 0.65 to 0.85; when yarn breakage or fraying is more obvious in grayscale, the value range of W2 is set to 0.65 to 0.85; the principle for selecting weights is to ensure that the feature values ​​corresponding to normal edge textures are mostly lower than the preset edge feature threshold, while the feature values ​​corresponding to real defect areas are more likely to exceed the preset edge feature threshold. The controller compares the edge feature values ​​with the preset edge feature threshold and filters out pixel regions that are greater than or equal to the threshold. The preset edge feature threshold is obtained by: pre-statistically calculating the edge feature values ​​of all pixels in the defect-free normal sample image region, calculating their mean and standard deviation, and setting the preset edge feature threshold as the mean plus 3 to 5 times the standard deviation, so that normal texture fluctuations are not triggered as much as possible, while the target defects are preserved as much as possible. For the selected pixel regions, the controller further performs connected component analysis to merge adjacent abnormal pixels into several defective connected regions. Specifically, at the algorithm processing level, the region growing method based on eight connected neighborhoods is used for analysis: the system scans the image sequentially, and when it encounters a pixel whose edge feature value is greater than or equal to the preset edge feature threshold and has not yet been marked, it is used as a seed point to set a new defect label and pushed into the coordinate stack. Enter the conditional judgment loop. As long as the coordinate stack is not empty, pop the top pixel and check its eight adjacent pixels in the surrounding eight directions. If the adjacent pixels also exceed the threshold and are not marked, set the same defect label and push it onto the coordinate stack. Repeat this process until the coordinate stack is empty, thereby completing the merging of the current connected region. This completes the distinction between scattered noise and continuous morphological defects. After obtaining the connected regions of each defect, the controller continues to determine whether the size range of each connected region in the longitudinal and transverse directions meets the set size conditions. The size conditions include longitudinal size limits and transverse size limits, which are specifically obtained by dividing the minimum acceptable physical size of the target defect and the maximum abnormal size by the object resolution of the industrial camera, respectively, to calculate the corresponding minimum number of connected pixels and maximum number of connected pixels. These are used as specific numerical limits for filtering effective areas in the longitudinal and transverse directions to exclude isolated noise points that are significantly smaller than the minimum acceptable defect size, and also to exclude abnormal lighting areas that are much larger than the normal size of small edge defects. When a connected region simultaneously satisfies both the vertical and horizontal dimension conditions, i.e., the vertical pixel row count of the bounding rectangle of the region is greater than or equal to the lowest connected pixel count and not higher than the highest connected pixel count, and the horizontal pixel column count is also within the corresponding reasonable range, the controller determines the region as an edge defect region and outputs the corresponding defect location and defect image; the defect location is specifically represented as a data packet containing the upper left corner coordinates, rectangle width, and rectangle height, which is passed to the output module; In this embodiment, the extraction of image edge trajectories is performed simultaneously with the aforementioned defect identification; the controller searches for the horizontal coordinate of the center of the seat belt edge on a continuous image row; here, the horizontal coordinate of the center of the edge refers to the representative position of the seat belt boundary in the horizontal direction within an image row, which can be obtained from the gray-scale transition position of the left and right edges; The specific calculation rule is as follows: For each image row, scan horizontally from the outside to the inside of the band. Pixels whose grayscale value first jumps from the low grayscale of the background to exceeding a set edge grayscale threshold are recorded as the left edge point of the current row. The set edge grayscale threshold... The calculation formula is: in, The average low grayscale value of the background. This represents the average normal grayscale value within the band. The threshold coefficient is between 0.3 and 0.5. The corresponding right edge point is obtained by scanning in reverse from the other side. The average of the horizontal coordinates of the left edge point and the right edge point is taken as the horizontal coordinate of the edge center of the row. When the horizontal coordinates of the edge centers of consecutive image rows are arranged in vertical row order, the image edge trajectory is formed. In order to use image information to correct the front-end prediction parameters, the controller first calculates the fluctuation of the image edge trajectory. Specifically, the average value of the abscissa of the edge center is calculated within a set update window. The update window is set to include a continuous reconstructed image row interval containing a preset fixed number of rows. Then, the average value is subtracted from the abscissa of the edge center of each image row to obtain the fluctuation sequence relative to the average position of the window. Image edge trajectories are extracted from reconstructed images with equal spatial spacing. The predicted lateral displacement sequence is a time-domain coordinate with equal time periods. Before comparing sequences, the controller performs time-space transformation and alignment: using the timestamp-vertical position mapping relationship stored synchronously in each control cycle, the predicted lateral displacement sequence is mapped to a non-uniform vertical spatial position domain; a vertical grid is generated using the same resampling interval as in step 4, and the mapped sequence is rearranged by one-dimensional spatial interpolation to generate an interpolated predicted displacement sequence that is perfectly aligned point-to-point with the image edge trajectory in the vertical spatial coordinates. When correcting the delay parameters, since the delay adjustment occurs in the time domain rather than the spatial domain, the controller takes multiple candidate delay parameters within the allowed search range, performs delay adjustment in the time domain respectively, and re-executes the above time-space conversion alignment process to generate multiple sets of corresponding spatial interpolation prediction displacement sequences. The linear correlation between the Pearson correlation coefficient and the image edge trajectory fluctuation vector is calculated. The candidate delay parameter corresponding to the maximum value of the Pearson correlation coefficient indicates that the predicted displacement is closest to the actual image swing in terms of dynamic response. Therefore, this parameter is updated as the new delay parameter. When correcting the conversion coefficients, the controller first selects one row in the reconstructed image as the reference image row, and takes the lateral displacement of other image rows relative to the reference image row as the relative displacement measured by the image; and compares this relative displacement sequence with the interpolated predicted lateral displacement sequence proportionally. The specific update logic is as follows: Extract the relative displacement sequence and interpolated predicted lateral displacement sequence of the image within the update window, and calculate the standard deviation of both; divide the standard deviation of the relative displacement of the image by the standard deviation of the interpolated predicted lateral displacement to obtain the amplitude scaling factor; multiply the current conversion factor by the amplitude scaling factor to obtain the target coefficient, and calculate the updated conversion coefficient by performing a weighted moving average with the current conversion coefficient using a set smoothing coefficient. The specific calculation formula is as follows: Among them, the updated conversion factors From the conversion factor before the update Combined with the relative displacement standard deviation of the image and interpolation prediction of lateral displacement standard deviation The calculation was performed using a smoothing coefficient ranging from 0.1 to 0.3. A weighted moving average is performed. If the lateral displacement is calculated using both in-phase and quadrature components, the weighted moving average update logic based on the standard deviation ratio is applied to the coefficients of the in-phase component and the coefficients of the quadrature component respectively, so that the lateral swing amplitude observed in the image is consistent with the predicted swing amplitude. The images are used for final defect determination and to correct the parameters of lateral sway calculated from the torque current; Reference Figure 2 In one embodiment, a seatbelt surface defect detection system based on industrial vision for any of the methods described in the embodiments includes: The acquisition module includes a current detection unit for acquiring torque current signals, a position detection unit for acquiring longitudinal position, and an industrial camera for acquiring raw images of the seat belt. It is used to acquire the torque current signals of the seat belt winding servo motor and the longitudinal position of the seat belt, and to acquire raw images of the seat belt. The calculation module is used to extract the tension fluctuation component in the torque current signal and calculate the lateral displacement and lateral velocity of the seat belt in the detection area based on the pre-acquired delay parameter and the pre-acquired conversion coefficient. The delay parameter is used to characterize the lateral displacement response time difference of the seat belt corresponding to the tension fluctuation component, and the conversion coefficient is used to characterize the conversion relationship between the tension fluctuation component and the lateral displacement. The trigger module is used to determine whether the exposure conditions of the industrial camera are met based on the horizontal speed, vertical position and preset speed threshold; if they are met, an exposure trigger signal is generated. The resampling module is used to obtain the longitudinal position of the seat belt corresponding to the exposure trigger signal, and to control the industrial camera to acquire the original image of the seat belt in response to the exposure trigger signal. The original image of the seat belt is then resampled longitudinally to generate a reconstructed image with equal spacing. The recognition module is used to identify surface defects of the seat belt on the reconstructed image and extract the image edge trajectory of the seat belt; The correction module is used to correct the delay parameters and conversion coefficients based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement. The output module is used to output the judgment results and reconstructed images of surface defects in seat belts.

[0024] This embodiment provides a seat belt surface defect detection system based on industrial vision for implementing the aforementioned method. The system is set up at the detection station where the seat belt is running continuously, and consists of a data acquisition module, a calculation module, a triggering module, a resampling module, an identification module, a correction module, and an output module. Each module can be centrally deployed in the same controller, or it can be implemented collaboratively by an industrial computer and an industrial camera control unit, as long as the modules maintain the corresponding data transmission and reception relationship. The current detection unit in the acquisition module is connected to the take-up servo driver to read the quadrature-axis current feedback value of the servo motor; the position detection unit is connected to the encoder to output accumulated pulses and calculate the longitudinal position of the seat belt; the industrial camera is installed above the detection area to acquire the original image of the seat belt when the trigger signal arrives; the data output by the acquisition module includes: the torque current sequence corresponding to the control cycle, the longitudinal position sequence, and the original image row data corresponding to each exposure; After receiving the torque-current sequence from the acquisition module, the calculation module first performs baseline removal and bandpass filtering to extract the tension fluctuation component. Then, based on the frequency rotation periodicity characteristics of this fluctuation component and the currently saved delay parameters, it performs phase extrapolation to generate a predicted lateral displacement sequence covering the current and future short time windows. The predicted lateral velocity sequence is then calculated based on the displacement difference between adjacent points and the time interval. The predicted velocity sequence and the current position output by the calculation module are not directly used as the display result, but are sent to the trigger module to participate in the forward exposure determination. At the same time, the lateral displacement is sent to the correction module as the prediction quantity for subsequent parameter comparison. The trigger module receives the predicted lateral velocity sequence, longitudinal position, and velocity threshold. Its working sequence is as follows: first, it determines whether the set longitudinal sampling interval has been reached based on the longitudinal position. If it has, the current pre-trigger moment is formed. Then, it reads the velocity at this moment and compares it with the velocity threshold. If the conditions are met, it directly outputs the exposure trigger signal. If the conditions are not met, the longitudinal distance between two adjacent industrial camera exposures is checked to see if it exceeds the set maximum skip distance. If it exceeds the maximum skip distance, the trigger module proactively retrieves multiple future candidate pre-trigger times from the predicted lateral velocity sequence, selects the best time with the minimum predicted lateral motion, and issues an exposure trigger signal or issues a timing command for waiting compensation trigger. In this way, the trigger module actually undertakes the task of merging the position coverage requirement and the low ambiguity requirement based on forward prediction into the same trigger determination process. The resampling module receives the exposure trigger signal output by the trigger module to control the industrial camera to acquire images. On the other hand, it reads the vertical position corresponding to the exposure trigger signal and associates the position with the original image row acquired by the industrial camera. Since the actual exposure cycle may not be uniform, the resampling module establishes an equally spaced vertical grid based on these image row-vertical position correspondences and interpolates and rearranges the original image to output an equally spaced reconstructed image. After receiving the reconstructed image, the recognition module calculates the gradient magnitude and grayscale difference in the seat belt edge region, and obtains the edge feature value by weighted summation. Then, it performs connected component analysis on the pixel region that reaches the threshold to filter out the edge defect region that meets the vertical and horizontal size conditions. The output of the recognition module includes at least the defect judgment result and the sequence of the horizontal coordinates of the edge center in the continuous image row, which is the image edge trajectory. The correction module receives the image edge trajectory output by the recognition module and the predicted lateral displacement sequence output by the calculation module. Since the image edge trajectory has spatial domain isometry and the predicted sequence has temporal domain isoperiodicity, and the reconstructed image row does not have a single real timestamp, the correction module uses the synchronously paired stored timestamp-vertical position mapping relationship to map the predicted lateral displacement sequence from the temporal domain and interpolate and resample it to the vertical spatial grid with the same spacing, so as to achieve consistent alignment with the image edge trajectory on the spatial reference. The fluctuation of the image edge trajectory relative to the average position is calculated. Then, within the allowed search range, different candidate temporal delay parameters are applied to generate multiple sets of spatially aligned interpolated predicted lateral displacement sequences. The set of sequences with the strongest linear correlation to the fluctuation is found, and the delay parameters are updated with their corresponding delay values. At the same time, the conversion coefficients are updated according to the standard deviation ratio between the lateral relative displacement of the image edge trajectory relative to the reference image row and the final aligned interpolated predicted lateral displacement. The updated parameters are written back to the calculation module for continued use in subsequent control cycles. The output module receives the defect judgment results from the recognition module and the reconstructed image from the resampling module, and outputs them to the upper monitoring interface or storage unit. The output content may include the vertical location range of the defect, the edge of the defect, the corresponding reconstructed image, and the judgment results of the current batch. The output module synchronously associates and outputs the image and the corresponding defect detection results in the regular spatial coordinates. In this embodiment, each module processes physical quantities related to the actual operating state of the seat belt; the current detection unit collects the electrical quantity of motor current, the position detection unit collects the displacement quantity of cumulative length of the seat belt, the industrial camera collects the image grayscale information formed by the reflection of the seat belt surface, and the calculation module and correction module process the lateral displacement and lateral velocity corresponding to the belt swaying situation in the detection area. The system controls the exposure of the industrial camera by processing the collected physical quantities and extrapolating the features, and restores the image to a reconstructed image for defect identification.

[0025] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting surface defects in seat belts based on industrial vision, characterized in that, Includes the following steps: Step 1: Obtain the torque current signal of the seat belt winding servo motor and the longitudinal position of the seat belt; Step 2: Extract the tension fluctuation component from the torque current signal, and calculate the lateral displacement and lateral velocity of the seat belt in the detection area based on the pre-acquired delay parameters and pre-acquired conversion coefficients; Step 3: Based on the lateral velocity, the longitudinal position, and the preset velocity threshold, determine whether the industrial camera exposure conditions are met; if so, generate an exposure trigger signal. Step 4: Obtain the longitudinal position of the seat belt corresponding to the exposure trigger signal, and control the industrial camera to acquire the original image of the seat belt in response to the exposure trigger signal. Perform longitudinal resampling on the original image of the seat belt to generate a reconstructed image with equal spacing. Step 5: Identify surface defects of the seat belt on the reconstructed image, extract the image edge trajectory of the seat belt, and correct the delay parameter and the conversion coefficient based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement calculated in Step 2.

2. The method for detecting surface defects of seat belts based on industrial vision according to claim 1, characterized in that, Step 2, which calculates the lateral displacement and lateral velocity of the seat belt in the detection area, includes: The torque current signal is preprocessed to obtain the tension fluctuation component; The lateral displacement is calculated based on the tension fluctuation component, the delay parameter, and the conversion factor, wherein the lateral displacement is the displacement obtained by aligning the tension fluctuation component according to the delay parameter and then converting it according to the conversion factor. The lateral velocity is calculated based on the change in lateral displacement at adjacent moments and the corresponding time interval.

3. The method for detecting surface defects of seat belts based on industrial vision according to claim 1, characterized in that, Step 3, determining whether the industrial camera exposure conditions are met, includes: When the longitudinal position reaches the set longitudinal sampling interval, a pre-trigger time is generated; Determine whether the lateral velocity corresponding to the pre-trigger moment is less than or equal to the velocity threshold; If it is less than or equal to, it is determined that the industrial camera exposure conditions are met, and the exposure trigger signal is issued at the pre-trigger time; If the speed is greater than the speed threshold, determine whether the longitudinal distance between two adjacent industrial camera exposures exceeds the set maximum skip distance; If the value is less than or equal to the maximum skip interval, then the current pre-triggering time is cancelled; If the distance is greater than the maximum skip interval, then the exposure trigger signal is emitted at the moment with the smallest lateral velocity among multiple pre-trigger moments within a preset range starting from the current pre-trigger moment.

4. The method for detecting surface defects of seat belts based on industrial vision according to claim 3, characterized in that, Among multiple pre-trigger times within a preset range, the exposure trigger signal is emitted at the moment with the minimum lateral velocity, including: Among the current pre-trigger time after reaching the maximum skip gap and multiple pre-trigger times within the subsequent preset time period, the time with the minimum lateral velocity is selected to issue the exposure trigger signal.

5. The method for detecting surface defects of seat belts based on industrial vision according to claim 1, characterized in that, Step 4 involves longitudinal resampling of the original seatbelt image, including: Based on the correspondence between the acquisition time sequence of the industrial camera and the longitudinal position of the seat belt, the longitudinal position of each row of pixels in the original image of the seat belt is obtained; A vertical grid is generated based on the set resampling interval; The original image of the seat belt is interpolated and rearranged according to the vertical grid to generate the reconstructed image.

6. The method for detecting surface defects of seat belts based on industrial vision according to claim 1, characterized in that, Step 5, which involves identifying surface defects of the seatbelt on the reconstructed image, includes: Calculate the gradient magnitude and grayscale difference of the seatbelt edge region in the reconstructed image; The gradient magnitude and the grayscale difference are multiplied by preset edge weights respectively, and then weighted and summed to obtain the edge feature value; Connectivity analysis is performed on pixel regions whose edge feature values ​​are greater than or equal to a preset edge feature threshold to obtain defective connected regions, and pixel regions whose edge feature values ​​are less than the edge feature threshold are determined as normal background regions; When the dimensions of the defective connected area in the longitudinal and transverse directions meet the preset defect determination size conditions, the area is determined to be a seat belt surface defect area.

7. The method for detecting surface defects of seat belts based on industrial vision according to claim 1, characterized in that, Step 5 involves correcting the delay parameter and the conversion coefficient based on the deviation between the actual lateral displacement of the seatbelt represented by the image edge trajectory and the lateral displacement itself. This includes: Calculate the abscissa of the center of the seat belt edge in a continuous row of images to form the image edge trajectory, and calculate the fluctuation amount of the image edge trajectory; The delay parameter is updated based on the temporal offset that has the strongest correlation with the image edge trajectory and the lateral displacement; The conversion coefficients are updated based on the ratio between the lateral relative displacement of the image edge trajectory with respect to the reference image row and the lateral displacement, wherein the reference image row is a pre-selected row of images in the reconstructed image.

8. A seatbelt surface defect detection system based on industrial vision for implementing the method according to any one of claims 1-7, characterized in that, include: The acquisition module includes a current detection unit for acquiring torque current signals, a position detection unit for acquiring longitudinal position, and an industrial camera for acquiring raw images of the seat belt. It is used to acquire the torque current signals of the seat belt winding servo motor and the longitudinal position of the seat belt, and to acquire raw images of the seat belt. The calculation module is used to extract the tension fluctuation component in the torque current signal and calculate the lateral displacement and lateral velocity of the seat belt in the detection area according to the pre-acquired delay parameter and the pre-acquired conversion coefficient. The delay parameter is used to characterize the lateral displacement response time difference of the seat belt corresponding to the tension fluctuation component, and the conversion coefficient is used to characterize the conversion relationship between the tension fluctuation component and the lateral displacement. The trigger module is used to determine whether the exposure conditions of the industrial camera are met based on the lateral velocity, the longitudinal position, and a preset velocity threshold; if the conditions are met, an exposure trigger signal is generated. The resampling module is used to obtain the longitudinal position of the seat belt corresponding to the exposure trigger signal, and to control the industrial camera to acquire the original image of the seat belt in response to the exposure trigger signal, and to perform longitudinal resampling on the original image of the seat belt to generate a reconstructed image with equal spacing. The recognition module is used to identify surface defects of the seat belt on the reconstructed image and extract the image edge trajectory of the seat belt; The correction module is used to correct the delay parameter and the conversion coefficient based on the deviation between the actual lateral displacement of the seat belt represented by the image edge trajectory and the lateral displacement. The output module is used to output the judgment result of the surface defects of the seat belt and the reconstructed image.

Images