A friction plate surface defect image recognition method and system

Abstract

The application provides a friction plate surface defect image recognition method and system, relates to the technical field of industrial automation detection, and actively induces and captures the instability of optical artifacts by introducing a small and controlled relative displacement between the collection of two images, quantifies the change amount of the to-be-analyzed area between the two images, and discriminates according to the stability standard, so that the unstable optical bright spots can be accurately recognized and excluded, only the stable real physical defects are alarmed, the interference problem of optical artifacts on defect detection is fundamentally solved, the false positive rate is greatly reduced, the reliability and accuracy of the automatic detection system are improved, and the effectiveness of product quality control is ensured.

Description

Technical Field

[0001] This application relates to the field of industrial automation inspection technology, and more specifically, to a method and system for image recognition of surface defects on friction plates. Background Technology

[0002] In the automated quality inspection stage of friction pad manufacturing, image recognition systems are widely used to identify surface defects. However, when production lines introduce friction pads using new materials and processes, existing systems face significant challenges. These new friction pads, due to their unique surface coatings, produce a specular reflection effect under fixed industrial light sources, forming highly bright optical spots that are not physically damaged. Traditional inspection methods primarily rely on brightness thresholds in static images to identify defects, failing to effectively distinguish between non-defective bright spots—formed by optical reflection artifacts and whose shape changes slightly with the viewing angle—and real defects such as scratches or pits with fixed physical morphologies. This confusion leads to numerous false alarms, misclassifying normal reflections on qualified products as defects, severely interfering with production efficiency and the accuracy of quality control. Furthermore, coarse exposure adjustments made to adapt to new materials may actually reduce the contrast between genuine minute defects and the background, increasing the risk of missed detections. Therefore, the core technical problem facing existing technologies lies in the lack of an effective mechanism to accurately distinguish between unstable optical artifacts and stable physical defects, resulting in the reliability of the inspection system failing to meet the requirements of high-precision production. Summary of the Invention

[0003] This application provides a method and system for image recognition of surface defects on friction pads, aiming to solve the technical problem of low accuracy in detecting surface defects on friction pads due to the inability to distinguish between optical reflection artifacts and real physical defects in the prior art. On the one hand, this application provides a method for image recognition of surface defects on friction pads, including: Acquire a first image of the friction pad, identify areas with brightness higher than a preset threshold from the first image, and mark them as areas to be analyzed; After the friction pad generates a preset relative displacement with respect to the image acquisition device, a second image of the friction pad is acquired. After aligning the first image and the second image, for the region to be analyzed, calculate its positional change, shape change, and brightness change between the first image and the second image. Based on the position change, shape change, and brightness change, and according to a preset stability discrimination standard, the region to be analyzed is distinguished as a non-defective optical bright spot or a real physical defect, and a distinction result is obtained; wherein, the non-defective optical bright spot is a brightness abnormal region formed by optical reflection illusion and without actual physical shape defects. Based on the differentiation results, only the regions determined to be real physical defects are output with corresponding defect information, while the regions determined to be non-defective optical bright spots are processed to be defect-free. Optionally, the preset threshold is calculated through the following steps: Acquire images of the defect-free friction pads in the current batch, and perform statistical distribution calculations on the grayscale values ​​of all pixels in the defect-free friction pad images to obtain the mean and standard deviation of grayscale values ​​that reflect the inherent optical reflection behavior of the defect-free friction pads under standard lighting conditions. The sum of the mean gray level and the standard deviation of gray level by a preset multiple is determined as the preset threshold. Optionally, the lower limit of the preset relative displacement is determined to enable the non-defective optical bright spot to produce an imaging position shift of at least two pixels on the imaging sensor of the industrial camera, thereby ensuring that the instability characteristics of the optical reflection artifact can be captured by the optical resolution and spatial sampling frequency of the industrial camera. The upper limit of the preset relative displacement is determined to be less than the equivalent diameter of the smallest detectable defect in the real physical defect, thereby ensuring that the real physical defect will not undergo connected domain breakage, boundary topology change or grayscale inversion due to changes in the observation angle during the image acquisition process, so as to maintain its characteristic stability as a solid geometric shape. Optionally, the step of acquiring a first image of the friction pad, identifying regions in the first image with brightness higher than a preset threshold, and marking them as regions to be analyzed includes: Traverse all pixels of the first image, aggregate isolated pixels with gray values ​​greater than a preset threshold using eight-neighbor or four-neighbor connectivity, and mark each connected region formed after aggregation as the region to be analyzed. Optionally, the step of distinguishing the region to be analyzed as a non-defect optical bright spot or a real physical defect based on the position change, the shape change, and the brightness change, and according to a preset stability discrimination criterion, to obtain the distinction result includes: By analyzing samples known to be non-defective optical bright spots and real physical defects, we can identify the distribution characteristics of the changes in position, shape and brightness of these two different types before and after a small displacement, and obtain the change patterns of non-defective optical bright spots and real physical defects. Compare the degree of matching between the positional change, shape change, and brightness change of the region to be analyzed and the non-defective optical bright spot change pattern and the real physical defect change pattern; Based on the degree of matching, the region to be analyzed is distinguished as a non-defect optical bright spot or a real physical defect, and the distinction result is obtained. Optionally, after the step of comparing the degree of matching between the positional change, shape change, and brightness change of the region to be analyzed and the non-defective optical bright spot change pattern and the real physical defect change pattern, the following is included: When the position change, shape change, and brightness change of the region to be analyzed are all within a preset effective matching range, the entropy value of the gray gradient distribution of the region to be analyzed is calculated. Calculate the average intensity of the edge detection operator response in the region to be analyzed; The entropy value of the gray-level gradient distribution is compared with a preset entropy threshold, and the average intensity of the edge detection operator response is compared with a preset intensity threshold. When the entropy value of the gray-scale gradient distribution is lower than the preset entropy threshold and the average intensity of the edge detection operator response is higher than the preset intensity threshold, the region to be analyzed is judged as a real physical defect. When the entropy value of the gray-scale gradient distribution is higher than the preset entropy threshold and the average intensity of the edge detection operator response is lower than the preset intensity threshold, the region to be analyzed is judged as a non-defective optical bright spot. Optionally, the preset entropy threshold is obtained through the following steps: Acquire images of defect-free friction pads in the current batch, and statistically analyze the gray-level gradient distribution entropy value of the defect-free region in the images of the defect-free friction pads to obtain the entropy value distribution range; Based on the entropy value distribution range, a preset entropy value threshold is set to distinguish normal textures from defects or bright spots. Optionally, the step of comparing the entropy value of the gray-level gradient distribution with a preset entropy threshold includes: The image of the defect-free friction pad is divided into multiple sub-regions; The entropy value of the gray-level gradient distribution of each sub-region is calculated to obtain the entropy value distribution range of each sub-region. Based on the entropy value distribution range of the sub-regions, a preset entropy value threshold is set for each sub-region; Determine the sub-region to which the region to be analyzed belongs, and calculate the entropy value of the gray-level gradient distribution of the region to be analyzed; The entropy value of the gray-level gradient distribution of the region to be analyzed is compared with the preset entropy threshold value corresponding to the sub-region. Optionally, the step of calculating the average intensity of the edge detection operator response in the region to be analyzed includes: Obtain the local grayscale distribution information of the region to be analyzed, and adjust the response threshold of the edge detection operator; Edge detection operators in multiple directions are used to perform edge detection on the region to be analyzed, and edge response results are obtained. By fusing edge response results from multiple directions, the average intensity of the fused edge response is calculated. On the other hand, this application provides a friction pad surface defect image recognition system, the system comprising: The image acquisition and region recognition module is used to acquire a first image of the friction pad, identify regions with brightness higher than a preset threshold from the first image, and mark them as regions to be analyzed. The second image acquisition module is used to acquire a second image of the friction pad after the friction pad has generated a preset relative displacement with respect to the image acquisition device. The change calculation module is used to calculate the positional change, shape change, and brightness change of the region to be analyzed between the first image and the second image after aligning the first image and the second image. The defect discrimination module is used to distinguish the region to be analyzed as a non-defect optical bright spot or a real physical defect based on the position change, the shape change and the brightness change, and according to a preset stability discrimination standard, and to obtain a discrimination result. The non-defect optical bright spot is a brightness abnormal region formed by optical reflection illusion and without actual physical shape defects. The information output module is used to output the corresponding defect information only for areas determined to be real physical defects, and to perform defect-free processing on areas determined to be non-defective optical bright spots. The technical solution provided in this application has significant advantages over existing technologies. Existing technologies, when dealing with novel friction pads with specular reflection characteristics, can only judge based on the brightness information of a single static image, failing to distinguish between physically stable real defects and optical bright spots whose shapes change due to variations in the viewing angle, leading to numerous false alarms. The core innovation of this application lies in introducing a temporal domain analysis dimension. By introducing a small, controlled relative displacement between two acquired images, the instability of optical artifacts is actively induced and captured. Real physical defects, as inherent geometric structures of an object's surface, are firmly bound to the object itself in terms of shape and position in the image, exhibiting high stability even under small displacements. In contrast, non-defective optical bright spots are essentially angular effects of light reflection, extremely sensitive to the viewing angle; even a small displacement is enough to cause drastic and irregular changes in their position, shape, or brightness in the image. The method of this application utilizes this fundamental difference. By quantifying the change in the area to be analyzed between two images and judging based on stability criteria, unstable optical bright spots can be accurately identified and eliminated, while only stable real physical defects trigger alarms. This method fundamentally solves the problem of optical artifacts interfering with defect detection, greatly reduces the false alarm rate, improves the reliability and accuracy of automated inspection systems, and ensures the effectiveness of product quality control. Attached Figure Description

[0004] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort. Figure 1 The diagram above illustrates a flowchart of a method for identifying surface defects on a friction pad. Figure 2 The diagram above illustrates a schematic of a friction pad surface defect image recognition system. Figure reference numerals: 100, Friction plate surface defect image recognition system; 10, Image acquisition and region recognition module; 20, Second image acquisition module; 30, Change calculation module; 40, Defect discrimination module; 50, Information output module. Detailed Implementation

[0005] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. In a typical automated production line scenario for friction pads, such as a factory producing automotive brake pads, quality control is a crucial step. At the end of the production line, an automated optical inspection system based on machine vision is typically deployed. This system uses industrial cameras to photograph each friction pad coming off the line and employs image processing algorithms to identify physical defects such as scratches, dents, and impurities on the surface. However, with the development of materials science, factories have begun to use composite materials containing special metal fibers or ceramic particles to manufacture high-performance friction pads. While this new material improves friction performance, it also brings unexpected inspection challenges. Its surface is not completely flat and diffusely reflective at the microscopic level; instead, it contains a large number of randomly distributed microcrystals or fiber cross-sections. Under a fixed ring light source, these microstructures act like countless miniature mirrors, reflecting light directly into the camera lens, thus forming numerous bright spots with extremely high brightness and irregular shapes on the image. These bright spots are not physical damage to the product surface; their position and shape are extremely sensitive to the relative angle between the friction pad and the camera. Traditional detection algorithms rely solely on the brightness threshold of static images, failing to distinguish between non-defective optical bright spots formed by optical reflection artifacts and real physical defects. This leads to numerous false alarms, misclassifying many qualified products as defective ones, severely impacting production efficiency and significantly increasing the burden of manual re-inspection. like Figure 1 The diagram illustrates an exemplary flowchart of a method for recognizing surface defects on a friction pad. This application proposes a method for recognizing surface defects on a friction pad, comprising: S10, acquire a first image of the friction pad, identify areas with brightness higher than a preset threshold from the first image, and mark them as areas to be analyzed; S20, after the friction pad generates a preset relative displacement with respect to the image acquisition device, a second image of the friction pad is acquired; S30, after aligning the first image and the second image, calculate the positional change, shape change, and brightness change of the region to be analyzed between the first image and the second image. S40, based on the position change, the shape change, and the brightness change, and according to a preset stability discrimination standard, the region to be analyzed is distinguished as a non-defective optical bright spot or a real physical defect, and a distinction result is obtained; wherein, the non-defective optical bright spot is a brightness abnormal region formed by optical reflection illusion and without actual physical morphological defects. Among them, true physical defects refer to structurally abnormal areas with stable three-dimensional geometric morphology that objectively exist on the surface of the friction pad. These defects are damage or flaws in the material itself, such as scratches caused by friction during processing, pits or pinholes caused by uneven mixing or shedding of materials, or foreign particles that have fallen into the production environment. Their core characteristic is that these defects are part of the object itself; their shape, size, and position in the object's coordinate system are fixed and will not disappear or undergo fundamental morphological changes due to slight changes in the viewing angle. Non-defective optical bright spots are purely optical phenomena, products of the imaging process, not physical entities on the surface of an object. They arise when the normal direction of a microscopic region on the surface of the friction pad precisely satisfies the specular reflection condition—that is, the incident angle of the light source equals the observation angle of the camera. This region reflects the light directly onto the camera sensor like a mirror, forming a bright spot with a local brightness far exceeding the surrounding area. Such bright spots lack corresponding physical unevenness. Their core characteristic lies in their instability; they are extremely sensitive to the geometric relationship between the light source, the object surface, and the camera. Even a slight change in this geometric relationship—such as moving or slightly rotating the friction pad—can disrupt the original microscopic specular reflection condition, causing the bright spot's position to drastically change, its shape to alter significantly, or even disappear completely. S50, based on the distinction result, output the corresponding defect information only for the areas determined to be real physical defects, and perform defect-free processing on the areas determined to be non-defective optical bright spots. The method proposed in this application is based on the fundamental difference between the stability of real physical defects and the instability of non-defect optical bright spots. The specific implementation process of this method will be explained in detail below. In a specific implementation scenario, friction pads are sequentially fed into the inspection station via a conveyor belt. A high-resolution industrial area scan camera, such as the ACE series camera from Basler GmbH, Germany, is fixedly installed above the inspection station, with a resolution of up to 2048 by 2048 pixels. Coaxial or ring light sources are arranged around the camera to provide uniform illumination to the surface of the friction pads. First, when a friction pad moves to a predetermined position directly below the camera, the conveyor belt pauses, triggering the camera to take its first exposure and acquire the first image of the friction pad. This image is a digital grayscale image, where the grayscale value of each pixel ranges from 0 to 255, representing the brightness of the corresponding location. Subsequently, the image processing unit begins to analyze this first image, aiming to initially filter out all potentially suspicious areas that may be defects or bright spots. This step is achieved through a preset brightness threshold. For example, a fixed grayscale value, such as 210, can be set. The image processing unit will iterate through every pixel in the first image, and any pixel with a grayscale value greater than 210 is considered suspicious. These suspicious pixels often cluster together, forming one or more bright areas. These areas are then marked as areas to be analyzed, and their initial information, such as position and outline, in the first image is recorded. Next, a crucial operation is performed: generating a tiny, controlled, preset relative displacement between the friction plate and the camera. This relative displacement can be achieved in various ways. In one embodiment, the conveyor belt can be moved forward a very small distance, such as 0.05 mm, by precisely controlling the drive motor of the conveyor belt, such as a stepper motor or a servo motor. Since the friction pads are placed on the conveyor belt, they also move by 0.05 mm. In another embodiment, the conveyor belt and friction pads remain stationary, while an industrial camera is mounted on a high-precision motorized translation stage. By controlling the stage's motor, the camera is moved horizontally by the same minute distance, for example, 0.05 millimeters. From the perspective of relative motion, this is equivalent to moving the friction pads. In yet another embodiment, a more ingenious approach can be adopted. A piezoelectric ceramic actuator is mounted below the clamp holding the friction pad. By applying a brief voltage pulse to the actuator, it can produce precise micron-level stretching and contraction, thereby driving the friction pad to produce a tiny and rapid translation or jitter. After completing this preset relative displacement, the camera is immediately triggered to perform a second exposure to obtain a second image of the friction plate. After acquiring two images with slight displacements, the image processing unit enters the core analysis stage. Due to the relative displacement, the second image will have a global translation relative to the first image. To accurately compare the changes of the same region to be analyzed in the two images, image alignment must be performed first. The purpose of alignment is to eliminate the influence of this global displacement. A commonly used alignment method is feature point registration. For example, in the non-defective background area of ​​the friction piece in the first image, stable corner points or texture features can be extracted using algorithms such as scale-invariant feature transformation. Then, the positions of these identical feature points can be found in the second image. By calculating the global displacement vector of these feature point pairs, the global translation and rotation between the two images can be accurately obtained. Subsequently, based on this transformation relationship, a geometric transformation is performed on the second image to generate an aligned image that is perfectly aligned with the first image in the coordinate system. Once alignment is complete, the stability features of each region to be analyzed, marked in the first image, can be calculated. The calculation process includes the following aspects: Positional change: In the first image, find the centroid coordinates of a region to be analyzed, denoted as (x1, y1). Then, in the second image after alignment, find the corresponding region that best matches this region near the same position and calculate its centroid coordinates, denoted as (x2, y2). The positional change is the Euclidean distance between these two coordinate points. For a real physical defect, since it is fixed to the friction plate body, its relative position should remain almost unchanged after global alignment, i.e., the positional change is close to zero. However, for a non-defective optical bright spot, because it is sensitive to angles, a small displacement may cause its reflection point to jump a long distance on the object surface, so its positional change will be significantly greater than zero. Shape Change: The stability of shape can be measured in various ways. For example, the area A1 and perimeter P1 of the region to be analyzed in the first image, and the area A2 and perimeter P2 of its corresponding region in the second image, can be calculated. A simple approach is to calculate the rate of change of area, i.e., the absolute value of (A2-A1) / A1. Alternatively, a more robust shape descriptor, such as Hu moments, can be used. Hu moments are a set of seven invariant moments calculated from the second and third central moments of the region, which are invariant to translation, rotation, and scaling. The Hu moment vectors of the region to be analyzed in the two frames can be calculated separately, and then the distance between these two vectors can be calculated as a measure of shape change. The shape of a real physical defect is fixed, and its shape change will be very small. In contrast, the shape of a non-defect optical bright spot can change drastically, such as changing from a circular bright spot to an elliptical bright spot, or breaking into several smaller bright spots, resulting in a large shape change. Brightness variation: Calculate the average gray value I1 of the region to be analyzed in the first image and the average gray value I2 of the corresponding region in the second image. The brightness variation can be the difference or ratio between the two. Since the formation of optical bright spots depends on a precise reflection angle, even a small displacement can disrupt this angle, causing a sharp drop in brightness; therefore, the brightness variation is usually large. In contrast, the contrast of real physical defects, such as a pit, with its surrounding area is mainly determined by its depth and diffuse reflection characteristics, and is not sensitive to small changes in the viewing angle; therefore, its brightness variation will be small. After calculating the changes in position, shape, and brightness of each region to be analyzed, a final distinction is made based on a preset stability criterion. In a basic implementation, this criterion can be a set of simple threshold rules. For example, it can be set that if the position change is greater than 3 pixels, or the area change rate is greater than 20%, or the average grayscale change rate is greater than 30%, then the region to be analyzed is determined to be a non-defective optical bright spot. If all three changes in a region to be analyzed are below their respective thresholds, then it is determined to be a real physical defect. Finally, the output is based on the differentiation results. For all areas determined to be real physical defects, their information, such as their coordinates in the image, defect size, and type, is recorded in the defect log file and highlighted on the monitoring interface, while triggering an alarm signal to notify the production line operator. For all areas determined to be non-defective optical bright spots, they are treated as normal surfaces and treated as defect-free, i.e., ignored directly and not reflected in any reports. The above method effectively filters out optical artifacts caused by the properties of new materials that are prone to misjudgment from the detection results, focusing only on the truly existing physical defects. Compared to existing technologies that rely solely on the brightness of a single static image, this application's method introduces a time dimension and dynamic analysis, utilizing the inherent stability differences between defects and artifacts. This significantly improves the accuracy and reliability of defect detection, avoiding decreased production efficiency and cost waste caused by numerous false alarms. In some embodiments, the preset threshold is calculated through the following steps: Acquire images of the defect-free friction pads in the current batch, and perform statistical distribution calculations on the grayscale values ​​of all pixels in the defect-free friction pad images to obtain the mean and standard deviation of grayscale values ​​that reflect the inherent optical reflection behavior of the defect-free friction pads under standard lighting conditions. The sum of the mean gray level and the standard deviation of gray level by a preset multiple is determined as the preset threshold. This is necessary because different batches of friction pads, and even the same batch at different production times, may have slight fluctuations in the raw material ratios and surface processing techniques. These fluctuations are reflected in the overall surface optical reflectivity, causing changes in the overall average brightness and contrast of the image. If a fixed brightness threshold is used, when a new batch of products is generally too bright, many normal areas may be incorrectly identified as areas to be analyzed, increasing the computational burden. Conversely, when a new batch of products is generally too dark, some less bright but real defects may be missed. The adaptive threshold calculation method proposed in this application effectively solves this problem. In practice, a simple calibration process can be performed before the start of production for each batch of products. The operator selects several (e.g., 50) standard samples confirmed to be defect-free, and allows them to pass sequentially through the inspection station, acquiring 50 images of these defect-free friction pads. Then, the image processing unit aggregates the grayscale values ​​of all pixels in these 50 images, forming a large dataset. For example, if each image is 2000 by 2000 pixels, then the dataset contains 200 million grayscale values. Statistical analysis is performed on this dataset to calculate the average grayscale value of all pixels, i.e., the grayscale mean, and the standard deviation of all pixel grayscale values. The grayscale mean reflects the average brightness level of the current batch of products under standard illumination, while the standard deviation reflects the degree of brightness fluctuation caused by normal surface texture. For example, after calculation, the average grayscale value of the current batch is found to be 145, and the standard deviation of grayscale is 12. Next, a preset multiplier needs to be determined, which is usually set based on experience, for example, 3. Therefore, the final preset threshold used to filter the areas to be analyzed is determined as: 145 + 3 * 12 = 181. This means that any pixel whose brightness exceeds the mean by 3 standard deviations is considered an abnormal bright spot requiring further analysis. This statistically based method allows the threshold to dynamically adapt to the optical characteristics of different batches of products, ensuring the stability and accuracy of the screening of the areas to be analyzed. It is a more intelligent and robust threshold setting strategy. In some embodiments, the lower limit of the preset relative displacement is determined to enable the non-defective optical bright spot to produce an imaging position shift of at least two pixels on the imaging sensor of the industrial camera, thereby ensuring that the instability characteristics of the optical reflection artifact can be captured by the optical resolution and spatial sampling frequency of the industrial camera. The upper limit of the preset relative displacement is determined to be less than the equivalent diameter of the smallest detectable defect in the real physical defect, thereby ensuring that the real physical defect will not undergo connected domain breakage, boundary topology change or grayscale inversion due to changes in the observation angle during the image acquisition process, so as to maintain its characteristic stability as a solid geometric shape. This principle provides a clear physical basis for setting the displacement. The fundamental reason for setting a lower limit for displacement lies in the discrete sampling characteristics of digital images. A camera sensor consists of individual pixel units, with a pixel being the smallest unit of an image. If the distance an optical spot moves between two images is less than the size of one pixel on the camera's imaging plane, then in the digital image, the center of this spot may still fall within the same pixel, making a change in position undetectable. Even a movement of approximately one pixel can lead to ambiguity due to sampling effects. Therefore, to obtain a clear and reliable motion signal, the displacement must be sufficient to allow the spot's imaging position to span at least two pixels on the sensor. For example, suppose the camera's optical system makes 0.01 mm on an object correspond to one pixel on the camera sensor. Then, to meet the lower limit requirement, the relative displacement of the friction pad must be greater than 0.02 mm. This ensures that even the least sensitive optical spot's positional change can be clearly captured. The upper limit for the displacement is set to protect the characteristic stability of real physical defects. While real defects are physically stable, their appearance in an image is affected by the viewing angle. If the relative displacement is too large, causing a significant difference in the viewing angle of the defect during a second shot, problems may arise. For example, a thin scratch, if the viewing angle changes too much, may have its projection in the image broken into several segments, or its edges may become blurred. Even a normally dark pit may appear bright with a bright edge due to a change in the lighting angle, resulting in grayscale inversion. These changes can lead to the misclassification of the real defect as "unstable," causing it to be incorrectly classified as an optical bright spot and missed. To avoid this, the relative displacement must be less than the minimum defect size that the system is required to detect. For example, if the quality standard requires the detection of all defects with a diameter of 0.1 mm or more, then the relative displacement must be set to less than 0.1 mm, such as 0.05 mm. This ensures that even the smallest defects will still overlap in the two images, preserving their connectivity and topological structure in the image, thus maintaining their status as a criterion for judging stable physical features. In summary, by setting reasonable upper and lower limits for displacement, such as choosing a displacement amount between 0.02 mm and 0.1 mm (e.g., 0.05 mm) in the example above, operation can be performed within an optimal window. This effectively excites and detects the instability of optical bright spots while ensuring the characteristic stability of real physical defects, laying a solid foundation for accurate differentiation in the future. In some embodiments, the step of acquiring a first image of the friction pad, identifying regions with brightness higher than a preset threshold from the first image, and marking them as regions to be analyzed includes: Traverse all pixels of the first image, aggregate isolated pixels with gray values ​​greater than a preset threshold using eight-neighbor or four-neighbor connectivity, and mark each connected region formed after aggregation as the region to be analyzed. Specifically, after the image processing unit has traversed all pixels of the first image and found all pixels with gray values ​​higher than a preset threshold, it needs to organize these pixels. For example, in a small area, there may be multiple adjacent pixels that are very bright. This is where connectivity aggregation comes in. Four-neighbor connectivity aggregation refers to considering only the pixels directly adjacent to a central pixel in the four directions (up, down, left, and right). If any of these neighboring pixels also have a grayscale value higher than a threshold, then they are considered connected to the central pixel. Eight-neighbor connectivity aggregation is more lenient, including not only the four neighbors (top, bottom, left, and right), but also the four diagonal neighbors (top left, top right, bottom left, and bottom right). The algorithm starts with an unlabeled pixel that is above a threshold. Like "coloring," it marks all pixels connected to it by four-neighbor or eight-neighbor rules and whose grayscale values ​​are also above the threshold as the same region, for example, "region 1". Then, the algorithm continues to find the next unlabeled pixel above the threshold, repeating this process and marking its connected component as "region 2", and so on, until all pixels above the threshold are assigned to a connected component. In this way, the originally scattered bright pixels are effectively organized into independent, clearly defined regions. Each such region is a region to be analyzed. Subsequent calculations of changes in position, shape, brightness, etc., are all performed using these aggregated connected regions as basic units. In some embodiments, the step of distinguishing the region to be analyzed as a non-defect optical bright spot or a real physical defect based on the position change, the shape change, and the brightness change, and according to a preset stability discrimination criterion, to obtain the distinction result includes: By analyzing samples known to be non-defective optical bright spots and real physical defects, we can identify the distribution characteristics of the changes in position, shape and brightness of these two different types before and after a small displacement, and obtain the change patterns of non-defective optical bright spots and real physical defects. Compare the degree of matching between the positional change, shape change, and brightness change of the region to be analyzed and the non-defective optical bright spot change pattern and the real physical defect change pattern; Based on the degree of matching, the region to be analyzed is distinguished as a non-defect optical bright spot or a real physical defect, and the distinction result is obtained. This essentially introduces an idea based on machine learning or pattern recognition. Simple threshold rules are "one-size-fits-all" and may not handle complex situations where changing features are in a fuzzy area well. Pattern matching-based methods, on the other hand, are more flexible and powerful. In practice, a "learning" or "training" phase is required. This phase necessitates a database containing a large number of labeled samples. Operators can manually classify the regions to be analyzed in a batch of images, clearly identifying which are real defects and which are optical bright spots. For each sample in the database, the changes in position, shape, and brightness before and after a minute displacement are calculated, forming a three-dimensional feature vector. Then, statistical analysis is performed on the feature vectors of these two types of samples to establish two "change patterns". Non-defective optical spot variation patterns: By analyzing the feature vectors of all samples labeled as optical spots, the statistical distribution of variation characteristics of such regions can be obtained. For example, the mean of its positional change may be around 10 pixels, with a standard deviation of 5 pixels; the mean of its area change rate may be 40 percent, with a standard deviation of 20 percent. This statistical distribution constitutes the "variation fingerprint" or "variation pattern" of the optical spot. Real Physical Defect Variation Patterns: Similarly, by analyzing all samples labeled as real defects, their "variation patterns" can also be obtained. These patterns are typically characterized by a mean of positional variation close to 0 and a very small standard deviation (e.g., less than 1 pixel); the mean of area variation rate is also close to 0, and the standard deviation is also very small (e.g., less than 3%). After the "learning" phase is completed, the normal "working" or "discrimination" phase begins. When a new region to be analyzed appears, its three-dimensional variation feature vectors of position, shape, and brightness are calculated. This new feature vector is then compared with two pre-established variation patterns. The comparison can be made by calculating the statistical distance from the vector to the distribution centers of the two patterns, such as Mahalanobis distance. Mahalanobis distance considers the correlation between features and is a more efficient measure than Euclidean distance. If the vector is closer to the Mahalanobis distance of the "non-defect optical bright spot variation pattern," it is classified as an optical bright spot; conversely, if it is closer to the Mahalanobis distance of the "real physical defect variation pattern," it is classified as a real defect. By using this pattern matching-based approach, the discrimination process no longer relies on rigid thresholds, but is based on statistical patterns from a large amount of real data, which significantly improves the accuracy of the distinction and the ability to adapt to complex situations. In practical industrial applications, although analyzing the changes in position, shape, and brightness of the area under analysis before and after a minute displacement can distinguish most real physical defects from non-defect optical bright spots, ambiguity may arise in certain critical situations. For example, a very small, low-contrast real scratch may change its image features after displacement, causing its change feature vector to partially overlap with the change pattern of the optical bright spot. Conversely, a relatively "stable" optical bright spot caused by a small surface depression may not change much, thus confusing it with the change pattern of a real defect. When the change feature vector of an area under analysis matches both the "change pattern of non-defect optical bright spots" and the "change pattern of real physical defects" within a preset effective matching range, i.e., it cannot be clearly classified into either category, additional discrimination criteria are needed to break the deadlock. In some embodiments, after the step of comparing the degree of matching between the positional change, the shape change, and the brightness change of the region to be analyzed and the non-defective optical spot change pattern and the real physical defect change pattern, the following is included: When the position change, shape change, and brightness change of the region to be analyzed are all within a preset effective matching range, the entropy value of the gray gradient distribution of the region to be analyzed is calculated. Calculate the average intensity of the edge detection operator response in the region to be analyzed; The entropy value of the gray gradient distribution is compared with a preset entropy threshold, and the average intensity of the edge detection operator response is compared with a preset intensity threshold. When the entropy value of the gray-scale gradient distribution is lower than the preset entropy threshold and the average intensity of the edge detection operator response is higher than the preset intensity threshold, the region to be analyzed is judged as a real physical defect. When the entropy value of the gray-scale gradient distribution is higher than the preset entropy threshold and the average intensity of the edge detection operator response is lower than the preset intensity threshold, the region to be analyzed is judged as a non-defective optical bright spot. The essence of this step is to introduce a secondary analysis of the static texture features within the region, and its physical basis lies in: Real-world physical defects, such as scratches or dents, typically have relatively clear and well-defined boundaries. This is because they represent abrupt changes in the physical structure of the material. Consequently, in an image, the grayscale values ​​at their edges change drastically, causing edge detection operators (such as Sobel, Canny, or Laplacian operators) to produce strong responses at those locations. Simultaneously, the texture within the defective region tends to be more uniform or homogeneous than the surrounding normal surface (e.g., the brightness inside a dark dent may be very consistent), resulting in lower complexity in its grayscale gradient distribution, i.e., lower information entropy values. Non-defective optical bright spots, as a light reflection phenomenon, often have blurred and gradually changing boundaries, lacking a clear physical outline. Therefore, the response of their edge detection operators is usually weak. At the same time, the brightness distribution within the bright spot region may not be uniform, but rather superimposed with the inherent random texture of the underlying friction plate, presenting a more complex and disordered gray-level gradient distribution, which leads to a high information entropy value. In practice, when a region to be analyzed is determined to be a "fuzzy" case, the following operations will be performed: 1. Calculate the entropy of the gray-level gradient distribution: First, calculate the gray-level gradient of all pixels within the region. The gradient can be obtained by applying the Sobel operator to the image. Then, statistically analyze the distribution of all gradient values ​​within the region, forming a gradient histogram. Finally, calculate its Shannon entropy based on this histogram. The higher the entropy value, the more chaotic and unpredictable the gradient distribution. 2. Calculate the average intensity of the edge detection operator response: Apply the Canny edge detection operator to the region to obtain a binary edge image. Then, calculate the average value of all pixel values ​​in the edge image. Since edge pixels are 1 (or 255) and non-edge pixels are 0, this average intensity value is proportional to the density and intensity of the edges within the region. 3. Secondary discrimination: The calculated entropy value and edge intensity are compared with preset thresholds. For example, if a region has a low entropy value (indicating simple internal texture) and a high edge intensity (indicating clear boundaries), it will be ultimately judged as a real physical defect. Conversely, if the entropy value is high and the edge intensity is low, it will be judged as an optical bright spot. In some embodiments, the preset entropy threshold is obtained through the following steps: Acquire images of the defect-free friction pads in the current batch, and statistically analyze the gray-level gradient distribution entropy value of the defect-free region in the images of the defect-free friction pads to obtain the entropy value distribution range; Based on the entropy value distribution range, a preset entropy value threshold is set to distinguish normal textures from defects or bright spots. This method ensures the adaptability of the entropy threshold. When processing a new batch of products, several known defect-free samples are first analyzed. A large number of normal surface areas without any bright spots or defects are randomly selected from these sample images, and the grayscale gradient distribution entropy value of each normal area is calculated. By statistically analyzing these entropy values, the "entropy baseline" and its distribution range of the normal surface texture of the current batch of products can be obtained (e.g., a mean of 5.2 and a standard deviation of 0.4). The preset entropy threshold can be set based on this distribution; for example, it can be set as the mean minus a multiple of the standard deviation (e.g., 5.2 - 2 * 0.4 = 4.4). Thus, any area with an entropy value below 4.4 is considered to have an internal texture that is "too simple" than a normal surface, and is therefore highly likely to be a real defect. In some embodiments, the step of comparing the entropy value of the gray-level gradient distribution with a preset entropy threshold includes: The image of the defect-free friction pad is divided into multiple sub-regions; The entropy value of the gray-level gradient distribution of each sub-region is calculated to obtain the entropy value distribution range of each sub-region. Based on the entropy value distribution range of the sub-regions, a preset entropy value threshold is set for each sub-region; Determine the sub-region to which the region to be analyzed belongs, and calculate the entropy value of the gray-level gradient distribution of the region to be analyzed; The entropy value of the gray-level gradient distribution of the region to be analyzed is compared with the preset entropy threshold value corresponding to the sub-region. In practice, a standard, defect-free image can be virtually divided into a grid, such as a 5x5 grid, resulting in 25 sub-regions. During the calibration phase, the entropy distribution of normal texture within each sub-region is individually calculated, and a specific entropy threshold is set for that sub-region. This results in an "entropy threshold map." During actual detection, when a blurry region to be analyzed appears, its coordinates are first used to determine which sub-region it belongs to, and then the specific threshold corresponding to that sub-region is applied for comparison. This locally adaptive thresholding strategy greatly improves the accuracy of secondary discrimination when dealing with non-uniform textured surfaces. In some embodiments, the step of calculating the average intensity of the edge detection operator response of the region to be analyzed includes: Obtain the local grayscale distribution information of the region to be analyzed, and adjust the response threshold of the edge detection operator; Edge detection operators in multiple directions are used to perform edge detection on the region to be analyzed, and edge response results are obtained. By fusing edge response results from multiple directions, the average intensity of the fused edge response is calculated. This step aims to overcome two major weaknesses of traditional edge detection: sensitivity to changes in lighting and sensitivity to edge orientation. 1. Local Threshold Adjustment: Before performing edge detection on a region to be analyzed, instead of using a fixed global threshold, the gray-scale statistical characteristics (such as mean and variance) of the region and its surrounding small area are analyzed first. Based on this local information, the parameters of the edge detection operator (such as the high and low thresholds in the Canny operator) are dynamically adjusted. This allows edge detection to better adapt to situations with uneven local illumination or fluctuations in the brightness of the material itself. 2. Multi-directional Detection and Fusion: Instead of using a single direction-insensitive operator, a set of operators sensitive to specific directions is employed, such as four Kirsch operators most sensitive to edges in the horizontal, vertical, +45°, and -45° directions, respectively. These four operators are used to process the region separately, resulting in four directional edge response maps. These four response maps are then fused, with the maximum response value for each pixel across the four maps serving as the final response value. This fused edge map comprehensively reflects edge information in all directions within the region, avoiding missed detections due to the specific direction of defect edges. Finally, the average intensity is calculated based on this more comprehensive fused edge response map, resulting in a more robust and reliable edge feature metric. This application also proposes an image recognition system for surface defects of friction pads, such as... Figure 2 As shown, a friction pad surface defect image recognition system 100 includes: The image acquisition and region recognition module 10 is used to acquire a first image of the friction pad, identify regions with brightness higher than a preset threshold from the first image, and mark them as regions to be analyzed. The second image acquisition module 20 is used to acquire a second image of the friction pad after the friction pad has generated a preset relative displacement with respect to the image acquisition device. The change calculation module 30 is used to calculate the position change, shape change and brightness change of the region to be analyzed between the first image and the second image after aligning the first image and the second image. The defect discrimination module 40 is used to distinguish the region to be analyzed as a non-defect optical bright spot or a real physical defect based on the position change, the shape change and the brightness change, and according to a preset stability discrimination standard, and to obtain a discrimination result. The non-defect optical bright spot is a brightness abnormal region formed by optical reflection illusion and without actual physical shape defects. The information output module 50 is used to output the corresponding defect information only for areas determined to be real physical defects, and to perform defect-free processing on areas determined to be non-defective optical bright spots. This system functionalizes and modularizes each step in the aforementioned method, providing a clear architecture for the engineering implementation of the method. The image acquisition and region recognition module corresponds in hardware to the industrial camera, lens, light source, and control circuitry for triggering the camera to take pictures. In software, it corresponds to the algorithm program responsible for receiving image data and performing preliminary brightness threshold segmentation and connected component analysis. The output of this module is a series of labeled regions to be analyzed and their initial information in the first image. The core of the second image acquisition module is the actuator and its driver that control the relative displacement, such as a controller for the conveyor belt stepper motor or a motion controller for the camera translation stage. Upon receiving the signal indicating completion of the first image acquisition, this module immediately drives the actuator to produce a preset minute displacement, and then triggers the camera to take a second picture, acquiring the second image. The change calculation module is a purely software algorithm module. It receives a first image, a second image, and a list of regions to be analyzed as input. Internally, this module first executes an image alignment algorithm, and then, for each region to be analyzed, calls the corresponding algorithm library function to calculate its position, shape, and brightness changes between the two aligned images. Finally, it outputs a data structure containing the change feature vectors of all regions to be analyzed. The defect discrimination module is the core of the entire system's decision-making. It receives the feature vector output from the change calculation module and makes a judgment based on a preset stability criterion. This criterion can be a threshold rule hard-coded in the software or a pre-trained machine learning model (such as a support vector machine, decision tree, or neural network). When encountering ambiguous cases, this module also calls its internal secondary discrimination submodule to calculate entropy and edge strength to make a final decision. The output of this module is the final classification result for each region to be analyzed: a real physical defect or a non-defect optical bright spot. The information output module is responsible for processing the output results of the defect identification module. It filters out all areas judged as optical bright spots and only outputs the information of those judged as real physical defects in a formatted manner. The output format can be various, such as writing defect information to a local database or log file, sending alarm information to the central monitoring system via the network, or directly highlighting the defect location on a monitor connected to the system, accompanied by an audible and visual alarm. The above methods can also be implemented using computer programs. Therefore, this application also provides an electronic device, including a processor and a memory. The memory stores a computer program, and the processor executes the steps of the friction pad surface defect image recognition method described in any of the foregoing embodiments by calling the computer program stored in the memory. This electronic device can be an industrial personal computer equipped with an image acquisition card, an embedded system, or a dedicated digital signal processor. The processor is responsible for executing the instructions stored in the memory, which constitute the software that implements all the above-described algorithmic logic. In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some communication interfaces, devices, or modules, and may be electrical, mechanical, or other forms. Furthermore, the modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for image recognition of surface defects on a friction plate, characterized in that, include: Acquire a first image of the friction pad, identify areas with brightness higher than a preset threshold from the first image, and mark them as areas to be analyzed; After the friction pad generates a preset relative displacement with respect to the image acquisition device, a second image of the friction pad is acquired. After aligning the first image and the second image, for the region to be analyzed, calculate its positional change, shape change, and brightness change between the first image and the second image. Based on the position change, shape change, and brightness change, and according to a preset stability discrimination standard, the region to be analyzed is distinguished as a non-defective optical bright spot or a real physical defect, and a distinction result is obtained; wherein, the non-defective optical bright spot is a brightness abnormal region formed by optical reflection illusion and without actual physical shape defects. Based on the differentiation results, only the regions determined to be real physical defects are output with corresponding defect information, while the regions determined to be non-defective optical bright spots are processed to be defect-free.

2. The method for image recognition of surface defects of friction plates according to claim 1, characterized in that, The preset threshold is calculated through the following steps: Acquire images of the defect-free friction pads in the current batch, and perform statistical distribution calculations on the grayscale values ​​of all pixels in the defect-free friction pad images to obtain the mean and standard deviation of grayscale values ​​that reflect the inherent optical reflection behavior of the defect-free friction pads under standard lighting conditions. The sum of the mean gray level and the standard deviation of gray level by a preset multiple is determined as the preset threshold.

3. The method for image recognition of surface defects of friction plates according to claim 1, characterized in that, The lower limit of the preset relative displacement is determined to be such that the non-defective optical bright spot produces an imaging position shift of at least two pixels on the imaging sensor of the industrial camera, thereby ensuring that the instability characteristics of the optical reflection artifact can be captured by the optical resolution and spatial sampling frequency of the industrial camera. The upper limit of the preset relative displacement is determined to be less than the equivalent diameter of the smallest detectable defect in the real physical defect, thereby ensuring that the real physical defect will not undergo connected domain breakage, boundary topology change or grayscale inversion due to changes in the observation angle during the image acquisition process, so as to maintain its characteristic stability as a solid geometric shape.

4. The method for image recognition of surface defects of friction plates according to claim 1, characterized in that, The step of acquiring a first image of the friction pad, identifying areas with brightness higher than a preset threshold from the first image, and marking them as areas to be analyzed includes: Traverse all pixels of the first image, aggregate isolated pixels with gray values ​​greater than a preset threshold using eight-neighbor or four-neighbor connectivity, and mark each connected region formed after aggregation as the region to be analyzed.

5. The method for image recognition of surface defects of friction plates according to claim 1, characterized in that, The step of distinguishing the region to be analyzed as a non-defect optical bright spot or a real physical defect based on the position change, the shape change, and the brightness change, and according to a preset stability discrimination standard, to obtain the distinction result includes: By analyzing samples known to be non-defective optical bright spots and real physical defects, we can identify the distribution characteristics of the changes in position, shape and brightness of these two different types before and after a small displacement, and obtain the change patterns of non-defective optical bright spots and real physical defects. Compare the degree of matching between the positional change, shape change, and brightness change of the region to be analyzed and the non-defective optical bright spot change pattern and the real physical defect change pattern; Based on the degree of matching, the region to be analyzed is distinguished as a non-defect optical bright spot or a real physical defect, and the distinction result is obtained.

6. The method for image recognition of surface defects of friction plates according to claim 5, characterized in that, Following the step of comparing the degree of matching between the positional change, shape change, and brightness change of the region to be analyzed and the non-defective optical spot change pattern and the real physical defect change pattern, the following is included: When the position change, shape change, and brightness change of the region to be analyzed are all within a preset effective matching range, the entropy value of the gray gradient distribution of the region to be analyzed is calculated. Calculate the average intensity of the edge detection operator response in the region to be analyzed; The entropy value of the gray-level gradient distribution is compared with a preset entropy threshold, and the average intensity of the edge detection operator response is compared with a preset intensity threshold. When the entropy value of the gray-scale gradient distribution is lower than the preset entropy threshold and the average intensity of the edge detection operator response is higher than the preset intensity threshold, the region to be analyzed is judged as a real physical defect. When the entropy value of the gray-scale gradient distribution is higher than the preset entropy threshold and the average intensity of the edge detection operator response is lower than the preset intensity threshold, the region to be analyzed is judged as a non-defective optical bright spot.

7. The method for image recognition of surface defects of friction plates according to claim 6, characterized in that, The preset entropy threshold is obtained through the following steps: Acquire images of defect-free friction pads in the current batch, and statistically analyze the gray-level gradient distribution entropy value of the defect-free region in the images of the defect-free friction pads to obtain the entropy value distribution range; Based on the entropy value distribution range, a preset entropy value threshold is set to distinguish normal textures from defects or bright spots.

8. The method for image recognition of surface defects of friction plates according to claim 7, characterized in that, The step of comparing the entropy value of the gray-level gradient distribution with a preset entropy threshold includes: The image of the defect-free friction pad is divided into multiple sub-regions; The entropy value of the gray-level gradient distribution of each sub-region is calculated to obtain the entropy value distribution range of each sub-region. Based on the entropy value distribution range of the sub-regions, a preset entropy value threshold is set for each sub-region; Determine the sub-region to which the region to be analyzed belongs, and calculate the entropy value of the gray-level gradient distribution of the region to be analyzed; The entropy value of the gray-level gradient distribution of the region to be analyzed is compared with the preset entropy threshold value corresponding to the sub-region.

9. The method for image recognition of surface defects of friction plates according to claim 6, characterized in that, The step of calculating the average intensity of the edge detection operator response of the region to be analyzed includes: Obtain the local grayscale distribution information of the region to be analyzed, and adjust the response threshold of the edge detection operator; Edge detection operators in multiple directions are used to perform edge detection on the region to be analyzed, and edge response results are obtained. By fusing edge response results from multiple directions, the average intensity of the fused edge response is calculated.

10. A friction pad surface defect image recognition system, characterized in that, The system includes: The image acquisition and region recognition module is used to acquire a first image of the friction pad, identify regions with brightness higher than a preset threshold from the first image, and mark them as regions to be analyzed. The second image acquisition module is used to acquire a second image of the friction pad after the friction pad has generated a preset relative displacement with respect to the image acquisition device. The change calculation module is used to calculate the positional change, shape change, and brightness change of the region to be analyzed between the first image and the second image after aligning the first image and the second image. The defect discrimination module is used to distinguish the region to be analyzed as a non-defect optical bright spot or a real physical defect based on the position change, the shape change and the brightness change, and according to a preset stability discrimination standard, and to obtain a discrimination result. The non-defect optical bright spot is a brightness abnormal region formed by optical reflection illusion and without actual physical shape defects. The information output module is used to output the corresponding defect information only for areas determined to be real physical defects, and to perform defect-free processing on areas determined to be non-defective optical bright spots.

Images