Online visual detection system for watch metal frame

Abstract

The invention relates to the technical field of surface defect detection, in particular to an online visual detection system for a watch metal frame, which comprises a multi-angle polarization imaging module used for arranging polarization imaging sensors at different angles and acquiring data of a plurality of images of the watch metal frame according to the polarization imaging sensors. And the photometric stereo analysis module is used for carrying out brightness normalization and difference calculation on the data of the plurality of images, solving a local normal vector in combination with an illumination direction matrix, deriving a surface height gradient from a normal vector component, and finally generating a pseudo-color depth map for representing fine height differences of the surface of the watch frame. And the defect feature extraction module is used for acquiring local height difference distribution of the surface of the watch frame according to the pseudo-color depth map, and determining a potential defect area of the surface of the watch frame according to the local height difference distribution. High-precision identification of the defects of the metal frame of the watch is realized through photometric stereo analysis, strong light interference is inhibited, and the shallow defect detection capability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of surface defect detection, in particular to an online visual detection system for a watch metal frame. BACKGROUND

[0002] At present, the appearance detection of the watch metal frame mainly relies on machine vision technology, and in the related technology, two-dimensional imaging and image processing algorithm based on threshold segmentation are mostly used to identify and determine the scratches and depressions on the frame surface. In a small part of high-end detection systems, three-dimensional point cloud measurement means are also combined to supplement the deficiency of two-dimensional images in depth information acquisition, so as to improve the accuracy of defect positioning. However, since the watch metal frame is mostly made of stainless steel, titanium alloy or electroplated material, the surface of such materials generally has certain mirror reflection characteristics, which is easily disturbed by complex lighting environment in actual detection process.

[0003] In summary, the existing detection technology still has many limitations, which are specifically as follows:

[0004] Firstly, the mirror metal is easy to produce halo effect under strong light, and the light spot and reflection often cover the fine scratches on the frame surface, so that the traditional threshold segmentation method produces misjudgment, and the detection result lacks stability.

[0005] Secondly, the two-dimensional imaging technology is difficult to accurately capture the depth information of the shallow scratches or small depressions on the surface, and when three-dimensional point cloud is used for compensation, when the point cloud resolution is insufficient, the slight defect features on the surface are easily weakened or even filtered out in the data smoothing process, resulting in missing defects.

[0006] Thirdly, most of the existing detection systems only rely on fixed threshold and pre-trained model, lack of adaptive updating ability for unlabeled samples in online detection process, and cannot expand the decision boundary in real time, so frequent manual review and intervention are needed, which reduces the detection efficiency.

[0007] Therefore, an online visual detection system for a watch metal frame is proposed to solve the above problems. SUMMARY

[0008] In order to achieve the above purpose, in the present application, an online visual detection system for a watch metal frame is provided, which comprises:

[0009] A multi-angle polarization imaging module is used to arrange polarization imaging sensors at different angles, and to obtain multiple image data of the watch metal frame according to the polarization imaging sensors.

[0010] A photometric stereo analysis module is configured to perform brightness normalization and difference calculation on the plurality of image data, solve local normal vectors in combination with a light direction matrix, and derive surface height gradients from the normal vector components to finally generate a pseudo-color depth map for representing subtle height differences of the watch bezel surface.

[0011] A defect feature extraction module is configured to obtain a local height difference distribution of the watch bezel surface from the pseudo-color depth map, and confirm a potential defect area of the watch bezel surface according to the local height difference distribution.

[0012] An anomaly detection module is configured to establish an anomaly detection model, discriminate the potential defect area of the watch bezel surface according to the anomaly detection model, reconstruct image data of the potential defect area through the anomaly detection model, and calculate a reconstruction error map; and determine whether the potential defect area of the watch bezel surface is a defect according to a comparison result of the reconstruction error map and a preset error threshold.

[0013] An online learning module is configured to dynamically update a discrimination boundary of the anomaly detection model according to newly added unlabelled samples and misjudgment correction results.

[0014] Preferably, the plurality of image data of the watch metal bezel is obtained by arranging polarization imaging sensors at different angles, and specifically includes:

[0015] The polarization imaging sensor is formed by adding a linear polarizer at the incident light path position of the sensor and adding an adjustable polarization filter at the imaging lens position.

[0016] The polarization imaging sensor is arranged in a ring shape and is arranged in an equiangular distribution as a whole. The polarization imaging sensor obtains the plurality of image data of the watch metal bezel by sequentially activating light sources at different angles and synchronously collecting images. The image data includes reflected light intensity data, polarization direction difference data, and reference light intensity data relied on by brightness comparison.

[0017] Preferably, the brightness normalization and difference calculation are performed on the plurality of image data, the local normal vectors are solved in combination with the light direction matrix, and the surface height gradients are derived from the normal vector components to finally generate a pseudo-color depth map for representing subtle height differences of the watch bezel surface, and specifically includes:

[0018] The brightness comparison is performed by performing difference calculation on the brightness values of the images collected at different light angles at the same pixel position, and performing normalization processing on the brightness difference result in combination with the reference light intensity data, and specifically includes:

[0019]

[0020] In the formula, I(x, y, θ) represents the brightness value of the image collected at the pixel position (x, y) and the light angle θ. represents the brightness value of the image collected at the pixel position (x, y) and the light angle θ. the original luminance value of the pixel position under the i-th illumination angle; the original luminance value of the pixel position under the i-th illumination angle; the reference illumination intensity corresponding to the i-th illumination angle; the reference illumination intensity corresponding to the i-th illumination angle; the reference illumination intensity corresponding to the i-th illumination angle.

[0021] According to the normalized luminance vector and the illumination direction matrix, the local normal vector is calculated, specifically:

[0022]

[0023] In the formula, the reference illumination intensity corresponding to the i-th illumination angle; the reference illumination intensity corresponding to the i-th illumination angle, the generalized vector containing the albedo and normal information.

[0024] According to the normal vector component, the local gradient of the surface height is calculated, specifically:

[0025]

[0026] In the formula, the gradient of the surface height in the direction; the gradient of the surface height in the direction; the gradient of the surface height in the direction; the gradient of the surface height in the direction; the gradient of the surface height in the direction; the gradient of the surface height in the direction; the surface height function; the pseudo-color depth map is generated by mapping the luminance gradient vector distribution and the local normal vector estimation value into height change data, and generating the pseudo-color depth map in a pseudo-color coding manner according to the height change data, and the pseudo-color depth map is used to represent the subtle height difference of the watch bezel surface.

[0027] Preferably, the local height difference distribution of the watch bezel surface is obtained according to the pseudo-color depth map, and the potential defect area of the watch bezel surface is confirmed according to the local height difference distribution, specifically including:

[0028] The local height difference distribution is obtained by selecting a pixel neighborhood in the pseudo-color depth map, calculating the difference between the maximum value and the minimum value of the surface height function in the neighborhood, and combining the residual value of the local fitting plane, and according to the calculation result of the residual value, the height difference parameter distribution is obtained, which is used to represent the micro relief of the watch bezel surface.

[0029] The local height difference parameter distribution is compared with a preset height difference threshold value, and if the height difference exceeds the preset height difference threshold value, the corresponding pixel neighborhood is marked as a potential defect area; the height difference threshold value is set by calculating the statistical distribution of the normal sample surface height difference, and the mean value of the normal sample height difference is added by three times the standard deviation as the height difference threshold value.

[0030] Preferably, an anomaly detection model is established, and the potential defect area on the surface of the watch bezel is discriminated according to the anomaly detection model, the image data of the potential defect area is reconstructed through the anomaly detection model, and a reconstruction error map is calculated; whether the potential defect area on the surface of the watch bezel is a defect is determined according to the comparison result of the reconstruction error map and the preset error threshold value, specifically including:

[0031] The anomaly detection model is established based on a convolutional autoencoder architecture, and the image data and the pseudo-color depth map are taken as input data; the anomaly detection model compresses the features of the input data through an encoder, and reconstructs the input data through a decoder, and the output is a corresponding reconstruction image and a reconstruction error map.

[0032] The anomaly detection model discriminates the potential defect area on the surface of the watch bezel in the following process:

[0033] The potential defect area confirmed by the local height difference distribution is input into the anomaly detection model, the reconstruction error map of the potential defect area is obtained, the reconstruction error map is compared with the preset error threshold value, and if the reconstruction error exceeds the preset error threshold value, the determination result is output as a defect area.

[0034] If the reconstruction error does not exceed the preset error threshold value, the determination result is output as a non-defect area.

[0035] If the anomaly detection model cannot effectively discriminate the potential defect area, the current potential defect area is marked as a to-be-confirmed area and stored in an online learning module for dynamically updating the model discrimination boundary in combination with subsequent newly added unannotated samples and misjudgment correction results.

[0036] The preset error threshold value is set by calculating the statistical distribution of the reconstruction error of the normal sample, and the mean value of the normal sample reconstruction error is added by two times the standard deviation as the default preset error threshold value, and the preset error threshold value is updated according to the online learning module.

[0037] An online visual detection system for a watch metal bezel also includes a review module, the review module is used for receiving samples determined as to-be-confirmed areas, providing comparison results of original images and pseudo-color depth maps, and recording determination results after manual review, and feeding back the review results to the online learning module to dynamically update the discrimination boundary of the anomaly detection model.

[0038] Preferably, according to the newly collected unmarked samples and the misjudgment correction results, the discrimination boundary of the anomaly detection model is dynamically updated, specifically including:

[0039] The newly collected unmarked samples are obtained by real-time storage through image data and pseudo-color depth maps; and the misjudgment correction results are obtained by manually reviewing the samples that are determined as the to-be-confirmed areas by the anomaly detection model and taking the review results as the correction labels.

[0040] The anomaly detection model performs incremental training according to the newly collected unmarked samples and the misjudgment correction results, and dynamically updates the default preset error threshold. After the newly collected normal samples and the correction results are introduced, the mean and the standard deviation are recalculated, the default preset error threshold is dynamically updated, and the dynamically updated preset error threshold is applied to the anomaly detection model.

[0041] The application has the following beneficial effects:

[0042] 1. In view of the problem that the mirror metal produces halos under strong light in the background art, causing the fine scratches to be covered, the application sets a polarizer and an adjustable polarizing filter at the positions of the incident light path and the imaging light path of the polarization imaging sensor, suppresses the mirror reflection light and the halo effect, so that the surface fine scratches and the slight depressions remain visible in the image data, avoids the misjudgment of the traditional threshold segmentation under the strong reflection background, and improves the stability of the detection result.

[0043] 2. In view of the problem that two-dimensional imaging is difficult to capture shallow defects and the insufficient resolution of three-dimensional point clouds causes the defects to be smoothed out, the application performs brightness normalization and difference calculation on multi-angle image data through photometric stereo analysis, combines with an illumination direction matrix to solve local normal vectors, and then derives surface height gradients from the normal vector components to generate a pseudo-color depth map that can represent fine height differences. In cooperation with the defect feature extraction module, the pseudo-color depth map is analyzed for local height difference distribution to accurately identify shallow scratches and slight depressions, overcoming the defects of insufficient two-dimensional detection accuracy and excessive three-dimensional point cloud smoothing.

[0044] 3. In view of the problem that the existing detection system lacks adaptive updating mechanism and cannot effectively process unmarked samples and dynamic decision boundaries, the application introduces an anomaly detection module and an online learning module, uses a convolutional autoencoder architecture to reconstruct potential defect areas and output a reconstruction error map, and uses the comparison result between the reconstruction error and a preset threshold as the discrimination basis. On this basis, the discrimination boundary of the anomaly detection model is dynamically updated through the newly collected unmarked samples and the misjudgment correction results formed by manual review, the anomaly detection model is adaptively and iteratively updated, and the frequency of manual intervention is reduced.

[0045] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is an overall framework diagram of an online visual inspection system for watch metal frames provided in an embodiment of this application. Detailed Implementation

[0048] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0049] Please see Figure 1 , Figure 1 This is an overall framework diagram of an online visual inspection system for watch metal frames provided in an embodiment of this application.

[0050] In this embodiment, an online visual inspection system for the metal frame of a watch includes:

[0051] The multi-angle polarization imaging module is used to deploy polarization imaging sensors at different angles and acquire multiple image data of the watch's metal frame based on these sensors. Specifically, this includes:

[0052] The polarization imaging sensor is formed by adding a linear polarizer at the incident light path position of the sensor and an adjustable polarization filter at the imaging lens position.

[0053] It should be noted that the linear polarizer is installed at the incident light path position of the polarization imaging sensor, and its polarization direction is adjustable relative to the emission direction of the light source, with an adjustable range of 0° to 90°. The adjustable polarization filter is installed at the front end of the imaging lens, and its polarization angle can be continuously adjusted via an electrically rotating bracket to obtain optimal contrast under different detection environments. The combination of the linear polarizer and the adjustable polarization filter can effectively reduce the specular component generated by the strong reflection from the metal surface, thereby enhancing the scattering component corresponding to surface defects in the image, thus keeping fine scratches or shallow depressions visible in the image data. The polarization imaging sensor is not only suitable for stainless steel and titanium alloys, but also for chrome-plated or nickel-plated metal surfaces with strong reflective properties.

[0054] The polarization imaging sensors are arranged in a ring, with the overall arrangement at equal angles. The polarization imaging sensors acquire multiple image data of the watch's metal frame by sequentially activating light sources at different angles and simultaneously acquiring images. The image data includes reflected light intensity data, polarization direction difference data, and reference illumination intensity data on which brightness comparison depends.

[0055] It should be noted that the number of ring-shaped sensors should be no less than four, with an angular interval of 360° divided by the number of sensors. The radius of the ring-shaped sensors should be selected between 10mm and 50mm based on the diameter of the watch bezel to ensure full coverage of the bezel surface. Through the ring-shaped arrangement and equal-angle distribution, a multi-directional uniform illumination and imaging effect is achieved.

[0056] It should be noted that the light source is an LED point light source, and the emission angle of each light source is consistent with the direction of the corresponding polarization imaging sensor. The switching of the light source is determined by the controller according to the position and the LED point light source is lit sequentially according to the triggering sequence. At the same time, the polarization imaging sensor receives the trigger signal and synchronously acquires images to ensure that each image corresponds to a specific illumination angle.

[0057] It should be noted that the reflected light intensity data in the image data is obtained from the original acquisition results of the sensor; the polarization direction difference data is calculated by adjusting the filter angle and recording the brightness difference at different angles; the reference light intensity data is acquired in real time by a light monitoring probe synchronized with the light source, and is used to perform normalization processing in the subsequent brightness comparison process.

[0058] The photometric stereo analysis module performs brightness normalization and difference calculations on multiple image data, solves for local normal vectors by combining the illumination direction matrix, and derives the surface height gradient from the normal vector components. Finally, it generates a pseudo-color depth map to represent subtle height differences on the watch bezel surface, specifically including:

[0059] The brightness comparison is performed by calculating the difference between the brightness values ​​of images acquired at the same pixel location under different illumination angles, and then normalizing the brightness difference results by combining them with reference illumination intensity data. The calculation formula is as follows:

[0060]

[0061] In the formula, Indicates the first Pixel position under different lighting angles The original brightness value; Indicates the first The reference illumination intensity corresponding to each illumination angle; This represents the brightness value after normalization to a reference illumination intensity, used to eliminate the influence of differences in the intensity of different light sources.

[0062] It should be noted that the purpose of the brightness normalization process is to eliminate the influence of differences in the intensity of different light sources, making images from different angles comparable under the same intensity benchmark. The normalized brightness values ​​are stored in matrix form and used as input for photometric stereo analysis.

[0063] It should be noted that the illumination direction matrix consists of circularly arranged light source direction vectors, with each row corresponding to a unit vector of a light source direction. The matrix dimension is ,in The value of m is the number of light sources; in this embodiment, the value of m is not less than 4 to ensure the solvability of the system of equations.

[0064] The local normal vector is calculated based on the normalized brightness vector and the illumination direction matrix:

[0065]

[0066] In the formula, This represents the brightness value after normalization to the reference illuminance. Represents the illumination direction matrix. This represents a generalized vector that contains information about albedo and normal.

[0067] It should be noted that generalized vectors Includes albedo and unit normal vector To obtain the unit normal vector, the product relationship needs to be determined by... After normalization, the calculation formula is as follows:

[0068]

[0069] The normalization step is used to ensure the consistency of the normal vector direction, which is not affected by the amplitude of local brightness changes.

[0070] Calculate the local gradient of the surface height based on the normal vector components:

[0071]

[0072] In the formula, Indicates the surface height at Gradient of direction; Indicates the surface height at Gradient of direction; The unit normal vectors are respectively in Component of direction; The surface height function is represented by the pseudo-color depth map, which maps the brightness gradient vector distribution and the local normal vector estimate to height change data, and generates a depth map based on the height change data using a pseudo-color encoding method, to represent subtle height differences on the watch bezel surface.

[0073] It should be noted that the surface height function The reconstruction was achieved using a numerical integration method, employing least squares integration, and a height reference point was set at the boundary conditions to avoid cumulative integration errors; the gradient field was analyzed... By performing global integration, a continuous surface height distribution is obtained.

[0074] It should be noted that the pseudo-color encoding process uses a linear mapping rule to reconstruct the surface height values. Normalization to The range is mapped to the RGB color space according to the height value, thus forming a pseudo-color depth map that is more sensitive to the human eye, so as to intuitively represent the subtle height differences on the border surface.

[0075] The defect feature extraction module is used to obtain the local height difference distribution on the surface of the watch bezel based on the pseudo-color depth map, and to identify potential defect areas on the watch bezel surface based on the local height difference distribution, specifically including:

[0076] The local height difference distribution is obtained by selecting a pixel neighborhood in the pseudo-color depth map, calculating the difference between the maximum and minimum values ​​of the surface height function within the neighborhood, and combining it with the residual value of the local fitting plane. Based on the calculation result of the residual value, the height difference parameter distribution is obtained. The height difference parameter distribution is used to represent the microscopic undulations of the watch bezel surface.

[0077] It should be noted that the selected range of the pixel neighborhood can be set to a small area composed of several pixels according to the detection accuracy requirements. The neighborhood is calculated to suppress the influence of isolated noise and ensure the stability of local height differences.

[0078] By comparing the local height difference parameter distribution with a preset height difference threshold, if the height difference exceeds the preset height difference threshold, the corresponding pixel neighborhood is marked as a potential defect area. The height difference threshold is calculated and set by the statistical distribution of the height difference on the surface of normal samples, and the mean of the height difference of normal samples plus three times the standard deviation is used as the height difference threshold.

[0079] It should be noted that the normal sample height difference refers to the height difference distribution obtained by statistically calculating the local pixel neighborhood on the surface of a watch metal bezel without obvious defects using a pseudo-color depth map. The calculated height difference distribution mainly reflects the minute textures, processing marks, and natural undulations produced on the material surface under normal processing or polishing processes, excluding scratches, dents, or other abnormal defects. The normal sample height difference is obtained as follows: within the bezel surface area confirmed to be defect-free by manual or historical inspection, the height difference values ​​of multiple neighborhoods are extracted using the same neighborhood size and calculation method as for defect detection, and these values ​​are formed into a statistical distribution set. By calculating the mean and standard deviation of the current set, a reference range reflecting the normal surface undulation characteristics is obtained. The reference range serves as the threshold benchmark for subsequent potential defect area determination. The definition of normal sample height difference ensures adaptability between different process batches and different metal materials. For materials with strong reflective properties such as stainless steel, titanium alloy, and electroplated alloy, as long as the bezel surface has undergone standard process treatment and is confirmed to be without abnormalities, it can be used as a normal sample for statistical analysis. By dynamically collecting normal samples from different batches and continuously updating the statistical distribution, it can be ensured that the height difference threshold always matches the actual production conditions.

[0080] The anomaly detection module is used to establish an anomaly detection model, identify potential defect areas on the watch bezel surface based on the model, reconstruct image data of the potential defect areas using the anomaly detection model, and calculate a reconstruction error map. Based on the comparison between the reconstruction error map and a preset error threshold, it determines whether a potential defect area on the watch bezel surface is indeed a defect. Specifically, this includes:

[0081] The anomaly detection model is based on a convolutional autoencoder architecture. It takes image data and pseudo-color depth maps as input data. The anomaly detection model compresses the features of the input data through an encoder and reconstructs the input data through a decoder. The output is the corresponding reconstructed image and reconstruction error map.

[0082] The process by which the anomaly detection model identifies potential defect areas on the surface of the watch bezel is as follows.

[0083] The potential defect area, identified by the local height difference distribution, is input into the anomaly detection model to obtain the reconstruction error map of the potential defect area. The reconstruction error map is then compared with a preset error threshold. If the reconstruction error exceeds the preset error threshold, the result is output as a defect area.

[0084] If the reconstruction error does not exceed the preset error threshold, the output judgment result is a non-defect area.

[0085] If the anomaly detection model cannot effectively identify potential defect areas, the current potential defect area is marked as an area to be confirmed and stored in the online learning module. This module is used to dynamically update the model's discrimination boundary by combining newly collected unlabeled samples and misjudgment correction results.

[0086] The preset error threshold is set by calculating the statistical distribution of the reconstruction error of normal samples. The mean of the reconstruction error of normal samples plus twice the standard deviation is used as the default preset error threshold. The preset error threshold is updated synchronously according to the online learning module.

[0087] It should be noted that the term "normal sample" refers to a set of image data on the surface of the watch's metal bezel that has been manually inspected or verified using historical data, confirming the absence of scratches, dents, wear, or other abnormal defects. This image data set includes both raw image data acquired by the multi-angle polarization imaging module and pseudo-color depth maps generated by the photometric stereo analysis module. Normal samples reflect the typical light reflection characteristics and height distribution characteristics exhibited by the bezel surface under standard processing techniques and normal usage conditions. Normal samples are obtained from two sources: first, defect-free bezel images confirmed by manual sampling on the production line, used for initial model training and threshold benchmark calculation; second, samples identified as non-defective areas by the anomaly detection model during online inspection and confirmed by manual verification, used for subsequent dynamic updates of the threshold.

[0088] It should be noted that the reconstruction error of normal samples refers to the numerical value obtained by measuring the difference between the input image and the reconstructed image after normal samples are input into the convolutional autoencoder model. The value obtained by the difference measure reflects the degree of fit of the anomaly detection model to the normal surface pattern. Since normal samples do not contain anomalous features, the reconstruction error has a low mean and a limited range of fluctuation in its statistical distribution. By statistically analyzing the mean and standard deviation of the reconstruction error of normal samples, a data benchmark can be provided for setting the preset error threshold.

[0089] An online visual inspection system for watch metal frames also includes a verification module. The verification module is used to receive samples identified as areas to be confirmed, provide comparison results of the original image and pseudo-color depth map, record the judgment results after manual verification, and feed the verification results back to the online learning module to dynamically update the discrimination boundary of the anomaly detection model.

[0090] It should be noted that the samples of the areas to be confirmed received by the verification module come from the output of the anomaly detection module. When the anomaly detection module cannot make an effective judgment on a certain area, it will automatically package the original image data of that area and the corresponding pseudo-color depth map and transmit them to the verification module.

[0091] It should be noted that the verification module displays the original image and the pseudo-color depth map side by side or overlaid through image comparison, allowing the verifiers to intuitively observe the subtle differences in the border surface under different data representations, thereby improving the accuracy of the verification.

[0092] It should be noted that the judgment results of manual review are recorded in the form of structured data, which may include binary classification results, namely defective and non-defective. If the label results are refined, they may include specific label categories for scratches, dents, and wear. The judgment results of manual review are stored together with the corresponding sample data.

[0093] It should be noted that when the review results are fed back to the online learning module, they include both the image data itself and the manually labeled judgment tags. The online learning module uses the feedback samples to incrementally train the anomaly detection model, thereby dynamically correcting the judgment boundaries so that subsequent similar defects can be identified more accurately.

[0094] The online learning module is used to dynamically update the discrimination boundary of the anomaly detection model based on newly collected unlabeled samples and misjudgment correction results. Specifically, it includes:

[0095] The newly acquired unlabeled samples are obtained through real-time storage of image data and pseudo-color depth maps; the misjudgment correction results are obtained by manually reviewing the samples identified as areas to be confirmed by the anomaly detection model and using the review results as correction labels.

[0096] It should be noted that the newly collected unlabeled samples are automatically extracted and stored in real time by the system during the detection process. The storage location is a dedicated sample cache library, which supports the corresponding storage of image data and pseudo-color depth maps, ensuring that each sample can be fully accessed. The dedicated sample cache library is periodically synchronized with the long-term database to ensure the traceability of data in subsequent model training.

[0097] The anomaly detection model performs incremental training based on newly collected unlabeled samples and misjudgment correction results, and dynamically updates the default preset error threshold. After introducing new normal samples and correction results, the mean and standard deviation are recalculated and the default preset error threshold is dynamically updated. The dynamically updated preset error threshold is then applied to the anomaly detection model.

[0098] It should be noted that the preset error threshold is updated using a sliding window statistical update mechanism. Under the sliding window statistical method, the system only counts the reconstruction error of normal samples in the most recent period to adapt to changes in process conditions, thereby ensuring that the preset error threshold always matches the latest detection environment.

[0099] It should be noted that the dynamically updated preset error threshold will be applied to the subsequent detection process immediately after training is completed; the updated anomaly detection model can directly apply the new discrimination threshold in new detection tasks to more accurately identify potential defect areas, thereby gradually reducing the occurrence of false positives and false negatives.

[0100] This completes the development of an online visual inspection system for watch metal bezels.

[0101] It should be understood that in the various embodiments of this application, the order of the above-mentioned units / modules does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0102] Those skilled in the art will recognize that the algorithms or steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An online visual inspection system for watch metal bezels, characterized in that, include: The multi-angle polarization imaging module is used to deploy polarization imaging sensors at different angles and acquire multiple image data of the watch's metal frame based on the polarization imaging sensors. The photometric stereo analysis module is used to perform brightness normalization and difference calculation on multiple image data, solve the local normal vector by combining the illumination direction matrix, and derive the surface height gradient from the normal vector components, and finally generate a pseudo-color depth map to represent the subtle height differences on the surface of the watch bezel. The defect feature extraction module is used to obtain the local height difference distribution on the surface of the watch bezel based on the pseudo-color depth map, and to identify potential defect areas on the surface of the watch bezel based on the local height difference distribution. The anomaly detection module is used to establish an anomaly detection model, identify potential defect areas on the surface of the watch bezel based on the anomaly detection model, reconstruct image data of potential defect areas through the anomaly detection model, and calculate the reconstruction error map. Based on the comparison between the reconstructed error map and the preset error threshold, it is determined whether the potential defect area on the surface of the watch bezel is a defect. The online learning module is used to dynamically update the discrimination boundary of the anomaly detection model based on newly collected unlabeled samples and misjudgment correction results.

2. The online visual inspection system for watch metal bezels as described in claim 1, characterized in that, Polarization imaging sensors are deployed at different angles to acquire multiple image data of the watch's metal frame, specifically including: The polarization imaging sensor is formed by adding a linear polarizer at the incident light path position of the sensor and an adjustable polarization filter at the imaging lens position. The polarization imaging sensors are arranged in a ring, with the overall arrangement at equal angles. The polarization imaging sensors acquire multiple image data of the watch's metal frame by sequentially activating light sources at different angles and simultaneously acquiring images. The image data includes reflected light intensity data, polarization direction difference data, and reference illumination intensity data on which brightness comparison depends.

3. The online visual inspection system for watch metal bezels as described in claim 2, characterized in that, Brightness normalization and difference calculations are performed on multiple image data. Local normal vectors are solved using the illumination direction matrix, and the surface height gradient is derived from the normal vector components. Finally, a pseudo-color depth map is generated to represent subtle height differences on the watch bezel surface. Specifically, this includes: The brightness comparison is achieved by calculating the difference between the brightness values ​​of images acquired at the same pixel location under different illumination angles, and then normalizing the brightness difference results by combining them with reference illumination intensity data. Specifically: ； In the formula, Indicates the first Pixel position under different lighting angles The original brightness value; Indicates the first The reference illumination intensity corresponding to each illumination angle; This represents the brightness value after normalization to a reference illuminance, used to eliminate the influence of differences in the intensity of different light sources; Based on the normalized brightness vector and the illumination direction matrix, the local normal vector is calculated as follows: ； In the formula, This represents the brightness value after normalization to the reference illuminance. Represents the illumination direction matrix. This represents a generalized vector that contains information about albedo and normal. The local gradient of the surface height is calculated based on the normal vector components, specifically as follows: ； In the formula, Indicates the surface height at Gradient of direction; Indicates the surface height at Gradient of direction; The unit normal vectors are respectively in Component of direction; The pseudo-color depth map represents the surface height function; it maps the brightness gradient vector distribution and the local normal vector estimate to height change data, and generates a pseudo-color depth map based on the height change data using a pseudo-color encoding method. The pseudo-color depth map is used to represent subtle height differences on the surface of the watch bezel.

4. The online visual inspection system for watch metal bezels as described in claim 3, characterized in that, The local height difference distribution on the watch bezel surface is obtained based on the pseudo-color depth map. This local height difference distribution is then used to identify potential defect areas on the watch bezel surface, specifically including: The local height difference distribution is obtained by selecting a pixel neighborhood in the pseudo-color depth map, calculating the difference between the maximum and minimum values ​​of the surface height function within the neighborhood, and combining it with the residual value of the local fitting plane. Based on the residual value calculation result, the height difference parameter distribution is obtained. The height difference parameter distribution is used to represent the microscopic undulations of the watch bezel surface. By comparing the local height difference parameter distribution with a preset height difference threshold, if the height difference exceeds the preset height difference threshold, the corresponding pixel neighborhood is marked as a potential defect area. The height difference threshold is calculated and set by the statistical distribution of the height difference on the surface of normal samples, and the mean of the height difference of normal samples plus three times the standard deviation is used as the height difference threshold.

5. The online visual inspection system for watch metal bezels as described in claim 1, characterized in that, An anomaly detection model is established to identify potential defect areas on the watch bezel surface. Image data of the potential defect areas are reconstructed using the anomaly detection model, and a reconstruction error map is calculated. Based on the comparison between the reconstructed error map and the preset error threshold, it is determined whether the potential defect area on the watch bezel surface is a defect, specifically including: The anomaly detection model is based on a convolutional autoencoder architecture. It takes image data and pseudo-color depth map as input data. The anomaly detection model compresses the features of the input data through the encoder and reconstructs the input data through the decoder. The output is the corresponding reconstructed image and reconstruction error map. The process by which the anomaly detection model identifies potential defect areas on the watch bezel surface is as follows: Input the potential defect area identified by the local height difference distribution into the anomaly detection model, obtain the reconstruction error map of the potential defect area, compare the reconstruction error map with the preset error threshold, and if the reconstruction error exceeds the preset error threshold, output the judgment result as a defect area. If the reconstruction error does not exceed the preset error threshold, the output judgment result is a non-defect area; If the anomaly detection model cannot effectively identify potential defect areas, the current potential defect area is marked as an area to be confirmed and stored in the online learning module. This module is used to dynamically update the model's discrimination boundary by combining newly collected unlabeled samples and misjudgment correction results. The preset error threshold is set by calculating the statistical distribution of the reconstruction error of normal samples. The mean of the reconstruction error of normal samples plus twice the standard deviation is used as the default preset error threshold. The preset error threshold is updated synchronously according to the online learning module.

6. The online visual inspection system for watch metal bezels as described in claim 1, characterized in that, It also includes a verification module, which receives samples identified as areas to be confirmed, provides comparison results between the original image and the pseudo-color depth map, records the judgment results after manual verification, and feeds the verification results back to the online learning module to dynamically update the discrimination boundary of the anomaly detection model.

7. The online visual inspection system for watch metal bezels as described in claim 5, characterized in that, Based on newly collected unlabeled samples and misjudgment correction results, the discrimination boundary of the anomaly detection model is dynamically updated, specifically including: The newly acquired unlabeled samples are obtained through real-time storage of image data and pseudo-color depth maps; the misjudgment correction results are obtained by manually reviewing the samples identified as areas to be confirmed by the anomaly detection model and using the review results as correction labels. The anomaly detection model performs incremental training based on newly collected unlabeled samples and misjudgment correction results, and dynamically updates the default preset error threshold. After introducing new normal samples and correction results, the mean and standard deviation are recalculated and the default preset error threshold is dynamically updated. The dynamically updated preset error threshold is then applied to the anomaly detection model.

Images