An intelligent detection method for shoe upper defects based on industrial vision

Abstract

The present application relates to the technical field of industrial vision detection, in particular to a kind of intelligent detection method of shoe surface defect based on industrial vision, comprising: obtaining the trigger signal of shoe surface entering imaging area and starting color industrial camera exposure;In exposure time, red stroboscopic light source and blue stroboscopic light source are triggered in turn, and single-frame color image is collected;Red channel image and blue channel image are separated;According to the conveying speed, the time difference of two light sources triggering and the pixel calibration coefficient, displacement compensation is carried out to blue channel image along the conveying direction;The difference value graph of red channel image and compensated blue channel image is calculated;Candidate defect area is extracted, and real defect, wrinkle false defect, noise or environmental block interference is judged according to area, boundary length and consistency coefficient;The present application converts the deformation difference in time dimension into color channel response difference, and cancels rigid body displacement through speed feedforward compensation.

Description

Technical Field

[0001] This invention relates to the field of industrial vision inspection technology, specifically to an intelligent detection method for shoe upper defects based on industrial vision. Background Technology

[0002] During the dynamic inspection process of continuous conveying of flexible shoe uppers such as flyknit, mesh, and synthetic leather, it is difficult to accurately distinguish between real defects such as holes, broken yarns, and scratches and false anomalies such as wrinkles, reflections, and noise.

[0003] Currently, traditional methods for detecting defects in shoe uppers mainly rely on the following approaches, but all of them have significant drawbacks. They are difficult to effectively eliminate interference caused by displacement and deformation during the shoe upper transportation process, and wrinkles are easily misjudged as real defects. They also have high requirements for the synchronization accuracy of cameras, light sources, and transportation, and are easily affected by timing jitter and motion blur on high-speed production lines. Furthermore, due to the lack of comprehensive constraints on the area of ​​candidate regions, boundary morphology, and differences in response at two different times, they are difficult to adapt to the detection needs of shoe uppers with different materials and reflective properties, leading to increased false detection and false negative rates. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent detection method for shoe upper defects based on industrial vision, avoiding the misjudgment of wrinkles as real defects due to displacement and deformation during dynamic transportation, and more accurately distinguishing between real defects and false anomalies in shoe uppers of different materials, thereby reducing the false detection rate and the missed detection rate. To achieve the above objectives, this invention adopts the following technical solution:

[0005] Acquire the trigger signal of the shoe upper entering the imaging area and start the color industrial camera exposure;

[0006] During the exposure time, a red strobe light source and a blue strobe light source are triggered sequentially to capture a single frame of color image;

[0007] Channel separation is performed on a single-frame color image to obtain a red channel image and a blue channel image;

[0008] The conveying speed of the shoe upper along the conveying direction is obtained. Based on the conveying speed, the time difference between the two light source triggers and the pixel calibration coefficient of the color industrial camera, the blue channel image is displacement compensated along the conveying direction to obtain the compensated blue channel image.

[0009] Calculate the difference map between the red channel image and the compensated blue channel image; extract candidate defect regions based on the difference map, and calculate the area, boundary length, and consistency coefficient of each candidate defect region;

[0010] The preset area threshold, boundary length threshold, and consistency threshold are retrieved as corresponding thresholds. Based on the comparison results of the area, boundary length, and consistency coefficient with the corresponding thresholds, the real defects, wrinkle pseudo-defects, noise, or environmental blocky interference are determined.

[0011] Preferably, the pulse width of both the red and blue strobe sources is no greater than [value missing]. The camera's exposure time is to .

[0012] Preferably, the displacement compensation includes: determining the moving distance based on the conveying speed of the shoe upper and the time difference between the two light source triggers, and shifting the blue channel image by the corresponding number of pixels in combination with the pixel calibration coefficient of the color industrial camera.

[0013] Preferably, extracting candidate defect regions based on the difference map includes: extracting pixels with pixel values ​​greater than a preset difference threshold from the difference map, and combining adjacent connected pixels into candidate defect regions.

[0014] Preferably, the comparison results include: when the area and boundary length of the candidate defect region meet the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, it is determined to be a real defect; when the area of ​​the candidate defect region meets the corresponding threshold and the consistency coefficient is greater than the corresponding threshold, it is determined to be a wrinkle pseudo-defect; when the area of ​​the candidate defect region does not meet the corresponding threshold, it is determined to be noise; when the area of ​​the candidate defect region meets the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, but the boundary length does not meet the corresponding threshold, it is determined to be an environmental blocky interference.

[0015] Preferably, the area is the total number of pixels contained in the candidate defect region, the boundary length is the contour length of the candidate defect region, and the consistency coefficient is the average pixel difference between the red channel image and the compensated blue channel image within the candidate defect region.

[0016] Preferably, after determining whether a defect is a real defect, a crease, or noise, the defect detection result is output, specifically including: when a defect is determined to be a real defect, a defect marker is generated in the image and output; when a defect is determined to be a crease, no marker is generated; when a defect is determined to be noise, the candidate defect area is directly discarded.

[0017] Preferably, the difference threshold, area threshold, boundary length threshold, and consistency threshold are determined based on the statistical results of good quality shoe upper samples, genuine defective shoe upper samples, and wrinkled shoe upper samples.

[0018] Preferably, the 95th percentile of the difference distribution of good quality shoe upper samples is used as the difference threshold, the 95th percentile of the area distribution of real defect samples is used as the area threshold, the 95th percentile of the boundary length distribution of real defect samples is used as the boundary length threshold, and the 95th percentile of the consistency coefficient distribution of wrinkled shoe upper samples is used as the consistency threshold.

[0019] The technical solution disclosed herein has the following beneficial effects:

[0020] 1. This invention achieves single-frame dual-color time-division sampling by sequentially triggering two different colored strobe light sources within a single exposure cycle. Combined with displacement compensation determined by the difference between the transport speed and the trigger time, it effectively counteracts the interference caused by the rigid body motion of the shoe upper and avoids misjudging wrinkles as real defects due to displacement and deformation in traditional static imaging.

[0021] 2. This invention uses a difference map to extract candidate regions and combines area, boundary length and consistency coefficient for comprehensive judgment. It can simultaneously constrain the region size, contour shape and the difference in response at two time points, thereby improving the accuracy of identifying real defects such as holes and broken yarns, and reducing false detections and missed detections caused by reflection, wrinkles and noise. Attached Figure Description

[0022] 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:

[0023] Figure 1 This is a flowchart illustrating an intelligent detection method for shoe upper defects based on industrial vision, provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments will be described in detail below with reference to the accompanying drawings.

[0025] A smart detection method for shoe upper defects based on industrial vision includes:

[0026] Acquire the trigger signal of the shoe upper entering the imaging area and start the color industrial camera exposure;

[0027] During the exposure time, a red strobe light source and a blue strobe light source are triggered sequentially to capture a single frame of color image;

[0028] Channel separation is performed on a single-frame color image to obtain a red channel image and a blue channel image;

[0029] The conveying speed of the shoe upper along the conveying direction is obtained. Based on the conveying speed, the time difference between the two light source triggers and the pixel calibration coefficient of the color industrial camera, the blue channel image is displacement compensated along the conveying direction to obtain the compensated blue channel image.

[0030] Calculate the difference map between the red channel image and the compensated blue channel image; extract candidate defect regions based on the difference map, and calculate the area, boundary length, and consistency coefficient of each candidate defect region;

[0031] The preset area threshold, boundary length threshold, and consistency threshold are retrieved as corresponding thresholds; based on the comparison results of area, boundary length, and consistency coefficient with the corresponding thresholds, real defects, wrinkle pseudo-defects, noise, or environmental blocky interference are determined.

[0032] The pulse widths of both the red and blue stroboscopic sources are no greater than [a certain value]. The camera's exposure time is to ;

[0033] Displacement compensation includes: determining the moving distance based on the conveying speed of the shoe upper and the time difference between the two light source triggers, and shifting the blue channel image by the corresponding number of pixels in conjunction with the pixel calibration coefficient of the color industrial camera.

[0034] This implementation method is based on an image acquisition framework using single-frame dual-color time-division sampling and displacement compensation, such as... Figure 1 As shown; during the conveyor belt transport process, the flexible shoe upper will produce elastic deformation such as wrinkles due to the vibration of the sole. Such deformation is difficult to distinguish from real structural damage such as holes in conventional static imaging.

[0035] Two strobe pulses of different colors are introduced within a single exposure cycle to convert the deformation difference in the time dimension into the response difference in the color channel, and the rigid body displacement is offset by velocity feedforward compensation.

[0036] The system monitors the imaging area through a photoelectric trigger sensor, such as the photoelectric sensor in the example. When the shoe upper is detected to be there, the control unit, such as the programmable logic controller in the example, records the trigger moment and sends a hard trigger signal to the color industrial camera to start the exposure cycle.

[0037] The camera's complementary metal-oxide-semiconductor sensor has an inherent global shutter clear and response delay, i.e. physical hysteresis. Within the set exposure time of 5ms to 15ms, the system uses the camera's hardware stroboscopic output interface to capture the hardware enable feedback transition edge of the sensor's underlying photoelectric integration, thus avoiding communication timing jitter caused by the controller's open-loop delay without feedback output.

[0038] The pulse width of both light sources is limited to no more than 100. After the camera completes the exposure, it outputs a single-frame color image. The image processing unit, such as the industrial control computer in the example, performs a channel separation operation on it to extract the red channel image and the blue channel image corresponding to two different sampling times.

[0039] The system reads the conveyor speed fed back in real time from the encoder. Combining the time difference between the two light source triggers Calculate the physical movement distance of the shoe upper. ,Right now And based on the camera's pixel calibration coefficient Convert the physical distance into pixel translation. ,Right now ;

[0040] Camera pixel calibration factor During the system debugging phase, this information is acquired in advance using the following method: A standard calibration board with known physical spacing characteristics is placed within the imaging field of view of a stationary conveyor belt. The camera is controlled to perform static imaging, and the number of pixels corresponding to this spacing characteristic in the image is measured. This number of pixels is then divided by the actual physical spacing to calculate the pixel translation amount. The system performs reverse displacement compensation on the blue channel image along the transport direction;

[0041] when When the value is not an integer, the system follows the preset interpolation rules. In this embodiment, bilinear interpolation is specifically used. The system calculates the pixel values ​​of the four adjacent integer pixels around the target sub-pixel position and performs a weighted summation using the product of the horizontal and vertical coordinate distances of the target point relative to the diagonal pixels as the weight to perform sub-pixel translation. If the translation interpolation window exceeds the effective grid of the sensor at the image edge, the system triggers the edge truncation strategy and automatically extracts the effective integer pixel value closest to the out-of-bounds coordinates for constant extension filling, thereby preventing the default zero value filling from causing anomalies.

[0042] At the same time, the system generates an effective mask corresponding to the translation amount. During the subsequent difference calculation, the edge area outside the mask is zeroed or cropped to ensure that the images before and after compensation can be effectively compared in the real physical overlap area.

[0043] After compensation, the system performs pixel-by-pixel difference calculation between the red channel image and the compensated blue channel image to generate a difference map. In this embodiment, the difference map is constructed using pixel-by-pixel difference magnitude, so that subsequent threshold extraction is based on the magnitude of the pixel difference rather than the positive or negative direction.

[0044] Candidate defect regions are extracted based on the difference map, and the area, boundary length and consistency coefficient are calculated to finally classify and determine real defects, false wrinkles or noise.

[0045] To ensure the continuity of the process, if the photoelectric trigger sensor causes signal jitter due to unevenness at the edge of the shoe upper, resulting in repeated triggering, the system will automatically activate a shielding period equal to the exposure time after the first trigger, ignoring the redundant signals during this period.

[0046] If the encoder fails to transmit the speed data of the current cycle due to line interference during the displacement compensation stage, the system will extract the average speed of the previous three valid cycles as the substitute value for the current cycle to perform compensation calculation, so as to avoid interruption of the detection process due to the lack of single-point data.

[0047] On a continuously operating flexible flyknit shoe upper inspection production line, the conveyor belt transports the shoe upper at a constant speed; when the shoe upper enters the camera's field of view, the programmable logic controller receives the photoelectric sensor signal and immediately sends a hardware trigger level command to the camera to start a 10ms exposure cycle;

[0048] The camera sensor has an inherent low-level clear response delay. After the light source controller captures the hardware feedback edge of the camera sensor's low-level integration, it avoids the risk of underexposure caused by physical timing delays in the time link.

[0049] After receiving the single-frame image transmitted from the camera, the industrial control computer extracts the red and blue channels and calculates the physical displacement within the 8ms time difference between the two light pulses using the current speed of the conveyor belt. After converting it into pixel values, the blue channel image is translated and aligned, thereby eliminating the interference of the rigid body motion of the conveyor belt and focusing the response of the difference map purely on the local deformation and defects of the shoe surface itself.

[0050] Furthermore, extracting candidate defect regions based on the difference map includes: extracting pixels with pixel values ​​greater than a preset difference threshold from the difference map, and combining adjacent connected pixels into candidate defect regions;

[0051] The comparison results include: when the area and boundary length of the candidate defect region meet the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, it is judged as a real defect; when the area of ​​the candidate defect region meets the corresponding threshold and the consistency coefficient is greater than the corresponding threshold, it is judged as a wrinkle pseudo-defect; when the area of ​​the candidate defect region does not meet the corresponding threshold, it is judged as noise; when the area of ​​the candidate defect region meets the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, but the boundary length does not meet the corresponding threshold, it is judged as an environmental blocky interference.

[0052] The area is the total number of pixels contained in the candidate defect region, the boundary length is the contour length of the candidate defect region, and the consistency coefficient is the average pixel difference between the red channel image and the compensated blue channel image within the candidate defect region.

[0053] After determining whether a defect is a real defect, a crease, or noise, the defect detection result is output. Specifically, when a defect is determined to be a real defect, a defect marker is generated in the image and output; when a defect is determined to be a crease, no marker is generated; and when a defect is determined to be noise, the candidate defect area is discarded directly.

[0054] After obtaining the difference map, this implementation method performs binarization judgment based on the difference intensity. When faced with large areas of slight wrinkles, the cumulative deformation may misjudge them as real damage. By introducing a multi-dimensional feature combination of area, boundary length and channel consistency coefficient, and by quantifying the geometric scale of the candidate region and the statistical results of the response at two time points, the method can effectively separate real and fake defects.

[0055] The system iterates through the difference map generated in the previous steps, extracts pixels whose pixel values ​​exceed the preset difference threshold, and combines spatially adjacent and connected pixels to construct candidate defect areas. Here, pixel values ​​exceeding the preset difference threshold means that the value representing the magnitude of pixel difference in the difference map is greater than the difference threshold, thereby avoiding the omission of abnormal responses in the other direction while only retaining a single difference direction.

[0056] For each candidate defect region, the system counts the total number of pixels it contains as the area feature and calls the 8-connected chain code traversal algorithm to track the region edge. By determining the connectivity between adjacent edge pixels one by one, the connection step between pixels in the horizontal and vertical directions is recorded as 1, and the connection step in the diagonal direction is recorded as 1.414. The system then globally accumulates these steps to obtain the outer contour scale of the region as the boundary length feature, in order to overcome the jagged deformation interference caused by the discrete pixel grid.

[0057] Simultaneously, the system extracts the pixel values ​​of the red channel image and the compensated blue channel image within this region, calculates the average of the absolute values ​​of their differences, and defines this average as the consistency coefficient. Specifically, based on the total number of pixels in the candidate region. and the first in the region The pixel value of each pixel in the red channel image and the compensated blue channel image. and The calculation is performed according to the following formula:

[0058]

[0059] in, This indicates summing over all pixels within the candidate region;

[0060] This consistency coefficient is used as a judgment parameter that matches the current two-color time-division sampling timing, compensation direction and threshold calibration calibrator. The consistency coefficient is based on the preset classification rules and sample statistical results, and no longer gives a monotonous interpretation of physical superiority or inferiority, thereby avoiding making inferences about the judgment direction that deviate from the calibration benchmark based solely on the literal meaning.

[0061] In the judgment stage, the system compares the extracted features with the corresponding preset thresholds; when the area and boundary length of the candidate region both meet the corresponding thresholds and the consistency coefficient is not greater than the corresponding threshold, the system judges it as a real defect and generates defect marking data at the corresponding coordinates of the original image for output.

[0062] When the area of ​​a candidate region meets the threshold, but the consistency coefficient is greater than the corresponding threshold, the system judges it as a wrinkle pseudo-defect and does not generate any mark.

[0063] When the area of ​​a candidate region meets the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, but the boundary length does not meet the corresponding threshold, it indicates that the shape of the region is too smooth and lacks real broken jagged edges. The system judges it as an environmental blocky interference and does not generate a label. When the area of ​​a candidate region does not reach the threshold, the system judges it as environmental noise and directly discards the data of the region from memory.

[0064] During feature extraction, if the difference map is interfered with by stray light in the field, causing small breaks in the candidate regions that should be connected, the system will call the morphological dilation operation before performing connected component combination. In this embodiment, a cross-shaped structural element of size 3×3 or 5×5 is used to scan the image edge to repair the difference map, so as to bridge the small breaks that are less than the corresponding pixel spacing and ensure the integrity of the defective region.

[0065] The specific mathematical mapping relationship of this morphological dilation operation is defined as grayscale dilation:

[0066]

[0067] in, To input a difference image, and These are the spatial coordinates of the image pixels. This is the domain of the cross-shaped structural element. and Define local coordinates within the domain of the structuring element, and replace the original center pixel value by extracting the local maximum pixel value within the area covered by the structuring element.

[0068] When calculating the consistency coefficient, in order to avoid the interference of edge artifacts caused by image translation interpolation, the system calls morphological erosion operation to peel off edge pixels in the candidate region. Because some candidate regions are too small, the number of core effective pixels after erosion becomes zero. At this time, the geometric scale of the candidate region is already in the non-physical artifact zone of the interpolation grid fluctuation. If the system directly divides by zero, it will cause abnormal calculation of the matrix data structure. Directly discarding the region may miss the detection of tiny real damage.

[0069] Therefore, the system triggers an abnormal degradation calculation mechanism, no longer using the mean calculation formula, but directly extracting the extreme value of the red-blue channel difference from the original pixel cluster before erosion of the tiny region, that is, the largest absolute value of the difference between individual pixels in the region. At this time, the system forcibly changes the formula for calculating the consistency coefficient from the mean to a maximum value formula to replace the consistency result in the original formula; while blocking the division-by-zero anomaly, it ensures the continuity of mathematical evaluation and prevents the omission of high-frequency tiny defects.

[0070]

[0071] After the above-mentioned flyknit upper completes channel compensation and generates a difference map, the industrial control computer scans the difference map and delineates multiple abnormal connected regions; one of the connected regions is generated by the stretching of the upper mesh, its area reaches the threshold, and the consistency coefficient calculated for this region is higher than the corresponding threshold. The system identifies it as a wrinkle pseudo-defect and ignores it according to the judgment rules.

[0072] Another connected domain originates from a broken yarn hole in the fabric. Its area and boundary length meet the corresponding threshold, and its consistency coefficient is not greater than the corresponding threshold. The system confirms it as a real defect, writes the coordinate information into the shared state record, and drives the rejection mechanism at the back end of the production line to perform the waste separation action.

[0073] The difference threshold, area threshold, boundary length threshold, and consistency threshold are determined based on the statistical results of good shoe upper samples, real defective shoe upper samples, and wrinkled shoe upper samples;

[0074] The 95th percentile of the difference distribution of good quality shoe upper samples was used as the difference threshold, the 95th percentile of the area distribution of real defect samples was used as the area threshold, the 95th percentile of the boundary length distribution of real defect samples was used as the boundary length threshold, and the 95th percentile of the consistency coefficient distribution of wrinkled shoe upper samples was used as the consistency threshold.

[0075] This implementation uses a fixed value as the judgment threshold. When the production line switches to shoe uppers with different materials or reflectivity, the fixed threshold will cause the detection system to miss or misjudge. By dynamically setting the benchmark parameters based on the actual operating parameters and the quantiles of the sample statistical distribution, the system's adaptability to different scenarios can be improved.

[0076] For the displacement compensation process, the system directly reads the actual conveying speed recorded by the encoder and multiplies it by the precise time difference between the two light source triggers to calculate the moving distance used for image translation, ensuring that the compensation amount is strictly synchronized with the physical state of the production line.

[0077] Regarding the threshold setting in the judgment process, the system reads the image data of pre-collected good shoe upper samples, real defective shoe upper samples, and wrinkled shoe upper samples during the initialization phase, and performs the same difference map and feature calculation as conventional detection.

[0078] The system performs distribution statistics on the calculation results to ensure the consistency of the judgment logic: In order to avoid calculation anomalies in the underlying data structure, when constructing the data structure of the difference distribution of good product samples, it does not perform undifferentiated full expansion of millions of underlying pixels in a single image. Otherwise, directly applying the 95th percentile would result in a fixed 5% in each good product image, i.e., a large number of false positive difference pixels, causing subsequent connected component operations to fail. Instead, the good product image is divided into a spatial grid of a preset scale, and only the local maxima of the pixel difference in each grid are extracted and entered into the distribution pool.

[0079] The local maximum difference features of the dimensionality-reduced good samples are sorted in ascending order from smallest to largest, and their 95th percentiles are extracted as the global difference threshold and upper limit threshold during runtime, thereby eliminating the consumption of a large amount of local artifact data on the system's computing power at the computational level.

[0080] The area and boundary length of the real defect samples are sorted in descending order from largest to smallest to construct an inverse cumulative distribution. The 95th percentile of each is extracted in descending order. The value is essentially equivalent to the smallest 5th percentile limit in the ascending distribution. These are used as the area threshold and boundary length threshold, respectively, as the lower limit reference values. When making formal judgments, a comparison direction not less than the descending 95th percentile threshold is adopted. This ensures that the calculation logic can retain and detect smaller defects in 95% of the real defect samples, avoiding missed detections due to quantile inversion.

[0081] Given that the consistency coefficient of the crease pseudo-defect is too large due to local elastic deformation, the system sorts the consistency coefficients of the crease shoe upper samples in descending order from large to small, and extracts the 95th percentile of its descending distribution, which is the lower bound of the consistency coefficient characterizing the crease feature, as the consistency threshold.

[0082] In subsequent judgments, the consistency coefficient of the candidate region must not be greater than this threshold in order to be judged as a real defect. The lower bound of the above judgment threshold is used to filter out interference and ensure that wrinkles above this lower bound are judged as false defects. After the above threshold is calculated, it is uniformly updated to the system's configuration parameter table for the judgment module to call directly.

[0083] To ensure consistency in statistical standards, the calibration phase for good products, genuine defects, and wrinkles uses the same trigger timing, exposure parameters, displacement compensation methods, and connected component extraction rules as the actual operation. This avoids distortion of quantile thresholds due to inconsistencies between the calibration and operation processes. At the same time, the system writes the material type, lighting configuration, and conveying speed range of each batch of samples into the configuration parameter table, so that the corresponding set of thresholds can be called when changing materials, instead of using a single threshold across different materials.

[0084] During the sample statistics phase, the system counts the total number of input samples for each category in real time. If the number of input samples for a certain category, such as real defective shoe upper samples, is insufficient, the specific judgment rule is that the total number of current input samples for that category is less than 20% of the preset standard sample size, resulting in the statistical confidence of the 95th percentile not meeting the preset requirements. The system will then stop the calculation of the current single sample set and instead extract the benchmark distribution data of the same material fabric from the historical database. The data will be fused with the current small number of samples according to the preset weights. In the fusion calculation, the weight of the current sample is specifically set as the ratio of the current number of samples to the preset standard sample size, and the remaining weights are all allocated to the historical benchmark data.

[0085] In this process, to eliminate the probability step faults generated at the boundaries when heterogeneous data is forcibly weighted and stitched together, the system further calls the kernel density estimation algorithm to reconstruct the continuous probability density function of the fused discrete histogram. The probability density function is defined as follows:

[0086]

[0087] in, For the reconstructed continuous probability density function, For the total sample size, For sample feature variables, For the first Feature values ​​of each sample Using the standard Gaussian kernel function:

[0088]

[0089] in, The independent variable of the kernel function; adaptive bandwidth Using empirical formulas Calculate, where, The standard deviation of the sample is used to output a smooth and consistent alternative threshold, preventing the threshold setting from being distorted due to severe bias or missing local samples.

[0090] If the quantile fluctuation of the same material in multiple consecutive batches exceeds the preset range, specifically defined as the absolute value of the relative deviation between the 95th quantile of the current batch and the mean of the quantiles of the previous three batches is greater than 15%, the system retains the threshold group of the previous stable batch and issues a review prompt to avoid misjudgment during operation due to individual abnormal batch samples directly overwriting the configuration table.

[0091] When a new synthetic leather shoe upper is introduced to the production line, the operator controls the conveyor belt to run a batch of standard samples, including good products, scratches and holes, and squeezed wrinkles; the system automatically records the current actual speed of the conveyor belt and the time difference of the light source to determine the moving distance, and performs quantile statistics on the characteristic data of this batch of samples;

[0092] Because the reflective properties of synthetic leather are different from those of woven fabric, the background difference of good quality synthetic leather is greater than the preset benchmark value. The system extracts the 95th percentile and automatically raises the difference threshold. It also writes the area threshold, boundary length threshold and consistency threshold corresponding to the material into the configuration table. Thus, during formal batch testing, it can filter out the interference signals caused by the normal texture reflection of synthetic leather surface under the same calibration.

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent detection of defects in shoe uppers based on industrial vision, characterized in that, include: Acquire the trigger signal of the shoe upper entering the imaging area and start the color industrial camera exposure; During the exposure time, a red strobe light source and a blue strobe light source are triggered sequentially to capture a single frame of color image; The single-frame color image is subjected to channel separation to obtain a red channel image and a blue channel image; The conveying speed of the shoe upper along the conveying direction is obtained. Based on the conveying speed, the time difference between the two light source triggers and the pixel calibration coefficient of the color industrial camera, the blue channel image is displacement compensated along the conveying direction to obtain the compensated blue channel image. Calculate the difference between the red channel image and the compensated blue channel image; Candidate defect regions are extracted based on the difference map, and the area, boundary length, and consistency coefficient of each candidate defect region are calculated. The preset area threshold, boundary length threshold, and consistency threshold are retrieved as corresponding thresholds. Based on the comparison results of the area, boundary length, and consistency coefficient with the corresponding thresholds, the real defects, wrinkle pseudo-defects, noise, or environmental blocky interference are determined.

2. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, The pulse widths of both the red and blue stroboscopic sources are no greater than [a certain value]. The camera's exposure time is to .

3. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, The displacement compensation includes: determining the moving distance based on the conveying speed of the shoe upper and the time difference between the two light source triggers, and shifting the blue channel image by the corresponding number of pixels in conjunction with the pixel calibration coefficient of the color industrial camera.

4. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, Extracting candidate defect regions from the difference map includes: extracting pixels with pixel values ​​greater than a preset difference threshold from the difference map, and combining adjacent connected pixels into candidate defect regions.

5. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, The comparison results include: when the area and boundary length of the candidate defect region meet the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, it is determined to be a real defect; when the area of ​​the candidate defect region meets the corresponding threshold and the consistency coefficient is greater than the corresponding threshold, it is determined to be a wrinkle pseudo-defect; when the area of ​​the candidate defect region does not meet the corresponding threshold, it is determined to be noise; when the area of ​​the candidate defect region meets the corresponding threshold and the consistency coefficient is not greater than the corresponding threshold, but the boundary length does not meet the corresponding threshold, it is determined to be an environmental blocky interference.

6. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, The area is the total number of pixels contained in the candidate defect region, the boundary length is the contour length of the candidate defect region, and the consistency coefficient is the average pixel difference between the red channel image and the compensated blue channel image within the candidate defect region.

7. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, After determining whether a defect is a real defect, a crease, or noise, the defect detection result is output, specifically including: when a defect is determined to be a real defect, a defect marker is generated in the image and output; when a defect is determined to be a crease, no marker is generated; when a defect is determined to be noise, the candidate defect area is directly discarded.

8. The intelligent detection method for shoe upper defects based on industrial vision according to claim 1, characterized in that, The difference threshold, area threshold, boundary length threshold, and consistency threshold are determined based on the statistical results of good quality shoe upper samples, genuine defective shoe upper samples, and wrinkled shoe upper samples.

9. The intelligent detection method for shoe upper defects based on industrial vision according to claim 8, characterized in that, The 95th percentile of the difference distribution of good quality shoe upper samples was used as the difference threshold, the 95th percentile of the area distribution of real defect samples was used as the area threshold, the 95th percentile of the boundary length distribution of real defect samples was used as the boundary length threshold, and the 95th percentile of the consistency coefficient distribution of wrinkled shoe upper samples was used as the consistency threshold.

Images