Tungsten steel mold defect detection method and system based on visual detection

Abstract

The application discloses a tungsten steel mold defect detection method and system based on visual detection, and belongs to the field of image processing, and comprises the following steps: heat compensation correction, defect detection analysis, optimized reflection model improvement and tungsten steel characteristic correction analysis. The application inputs the mold surface image into the specified reflection model for illumination processing, simultaneously performs heat compensation correction on the specified reflection model to obtain an optimized reflection model, and obtains a standardized reflection image accordingly, then inputs the standardized reflection image into a defect detection model to output a defect detection result, then obtains corresponding defect detection feedback results to determine whether to improve the optimized reflection model, if yes, tungsten steel characteristic correction analysis is performed, otherwise, reflection component separation is continued based on the optimized reflection model, the accuracy of tungsten steel mold defect detection is improved, and the problem that the visual detection result is inaccurate due to the tungsten steel surface characteristics in the prior art, so that the defect detection accuracy is reduced is solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for detecting defects in tungsten steel molds based on visual inspection. Background Technology

[0002] Currently, the field of metal mold defect detection is rapidly evolving from a traditional model heavily reliant on human experience towards automation and intelligence. Existing mainstream automated detection technologies utilize sensors and machines to replace human eyes, achieving more objective and efficient detection. Through machine vision, such as industrial cameras (linear or area arrays), lenses, light sources, image acquisition cards, computers, and processing software, they achieve image acquisition, image preprocessing, feature extraction, and classification and judgment of metal molds, thus improving the efficiency of defect detection. Simultaneously, to address the bottlenecks of traditional machine vision in complex defect detection, cutting-edge intelligent detection technologies have introduced deep learning models, especially convolutional neural networks. These technologies collect a large number of mold images containing various defects and qualified products, accurately annotating them (e.g., outlining and classifying defect locations), and then train the deep learning model using the annotated dataset. This model automatically learns the deep, abstract features of defects and deploys the trained model on a metal mold detection platform to perform real-time inference on newly acquired mold images, automatically outputting the location, category, and confidence level of defects.

[0003] For example, the Chinese invention patent with announcement number CN118587226B discloses a machine vision-based method and system for quality inspection of mold plates, which includes: constructing a defect detection dataset using a data augmentation method based on conditional generative adversarial networks; detecting defect areas of mold plates using a Defect-ResNet network; preprocessing the defect areas of the mold plates by using defect pixels in the defect areas as seed points, setting growth criteria, and continuously merging adjacent pixels that meet the growth criteria into the defect areas until no new pixels are added, thus obtaining segmented defect areas; extracting contours from the segmented defect areas to obtain a contour matrix composed of contour point coordinates; using a rotating caliper algorithm to find the circumscribed rectangle with the smallest area in the contour matrix and calculating the coordinates of the four vertices of the circumscribed rectangle; and using the formula for calculating the coordinates of the diagonal intersection point to obtain the center coordinates of the circumscribed rectangle as the position coordinates of the defect area.

[0004] For example, the Chinese invention patent with announcement number CN117974665B discloses a computer vision-based online inspection method and device for metal molds, which includes: S1: First, a high-resolution camera is used to take real-time pictures of the metal mold to obtain an image of the mold surface; S2: Then, the obtained mold surface image is optimized through image preprocessing technology; S3: Then, a computer vision algorithm is applied to identify the features of the mold from the image and output the defect detection result.

[0005] Deep learning-based visual inspection is becoming a development trend. It enables machines to automatically learn defect features through convolutional neural networks, demonstrating strong generalization ability and extremely high accuracy in the identification and classification of complex defects. Therefore, based on the existing technology for defect detection of metal molds, it can be generalized to the defect detection of tungsten steel molds, which also has certain reference value.

[0006] The production of tungsten carbide molds is a high-end manufacturing process integrating materials science, precision mechanics, automation control, and information technology. It involves a series of sophisticated equipment working in tandem. For example, an automated loader or feeding system pours precisely measured powder into a high-strength steel die. The upper and lower punches of a molding press simultaneously apply pressure from both ends towards the center under immense pressure. Finally, pressure is maintained and the mold is demolded. Then, in an annealing or sintering furnace at high temperatures, the cobalt binder phase melts, and through "liquid-phase sintering," tungsten carbide particles dissolve and precipitate, tightly bonding together to achieve densification. Simultaneously, equipment such as ball mills and spray drying towers are used to "shape" and "activate" micron-sized powder into high-performance industrial teeth—tungsten carbide molds. Similarly, existing technologies for defect detection of tungsten carbide molds utilize industrial cameras to acquire surface images of the molds, followed by image preprocessing, defect feature extraction, and defect identification and classification. Defect detection is a key technology in the mold manufacturing industry, used to replace manual visual inspection and improve the accuracy and efficiency of defect detection. However, tungsten steel molds (usually tungsten carbide or cemented carbide) have smooth and highly reflective surfaces, so the focus of visual inspection is on reflection suppression and high-precision detection of minute defects. To enhance defect visibility, image preprocessing commonly uses algorithms such as illumination equalization, brightness normalization, median filtering, bilateral filtering (denoising and edge preservation), reflection suppression algorithms (based on polarized light or image fusion), contrast enhancement, and edge enhancement (Sobel or Laplacian). Current reflection suppression algorithms often employ Lambert and Phong models to address reflectivity fluctuations caused by high-temperature environments. Furthermore, to detect various defects such as cracks, chipping, pitting, scratches, dents, impurities, and wear, current mainstream detection algorithms also utilize deep learning methods, such as Faster... R-CNN (Faster Region-Convolutional Neural Network) outputs defect categories and location boxes, while the segmentation network U-Net (U-shaped Network) outputs pixel-level defect regions. By using the above deep learning methods, high robustness of defect detection in tungsten steel molds can be achieved, and stronger adaptability to changes in lighting and complex textures can be obtained.

[0007] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0008] In existing technologies, the high-temperature environment of the molding press, automated loader, and annealing furnace used in the tungsten carbide mold production process may cause changes in the reflectivity of the mold surface, making visual inspection features unstable. Furthermore, the temperature and optical reflectivity of tungsten carbide molds vary greatly under different equipment operating conditions (such as molding, annealing, and handling), leading to unstable visual inspection results. However, the Lambert model commonly used in existing technologies only considers the diffuse reflection of light on the surface and cannot describe the specular reflection characteristics of tungsten carbide. The Phong model assumes that the specular reflection direction is fixed and cannot reflect the temperature-induced reflectivity distribution. Therefore, the surface characteristics of tungsten carbide lead to inaccurate visual inspection results, resulting in a decrease in the accuracy of defect detection. Summary of the Invention

[0009] To address the technical problem of inaccurate visual inspection results due to the surface characteristics of tungsten carbide, which leads to a decrease in defect detection accuracy, this invention provides a visual inspection-based method and system for tungsten carbide mold defect detection. The technical solution is as follows:

[0010] On the one hand, a visual inspection-based defect detection method for tungsten carbide molds is provided. This method includes: acquiring mold surface images of a batch of tungsten carbide molds to be inspected using an industrial camera, inputting these images into a specified reflection model for illumination processing, and simultaneously performing thermal compensation correction on the specified reflection model to obtain a corresponding optimized reflection model, thereby reducing the brightness reference drift of the defect detection model at different temperature stages; after separating the reflection components of the mold surface image according to the optimized reflection model, performing image preprocessing to obtain a standardized reflection image, which is then input into the defect detection model for defect detection analysis and outputting defect detection results; determining whether to re-inspect the batch of tungsten carbide molds to be inspected based on the defect detection results and obtaining corresponding defect detection feedback results, so as to determine whether to improve the optimized reflection model through defect detection performance analysis, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift; if the optimized reflection model is improved, performing tungsten carbide characteristic correction analysis to introduce the influence of tungsten carbide surface characteristics on illumination reflection for correction, and performing fluctuation judgment on the defect detection feedback results after the optimization reflection model is improved to determine whether to continue improving the optimized reflection model; otherwise, continuing to separate the reflection components based on the optimized reflection model.

[0011] On the other hand, a vision-based tungsten carbide mold defect detection system is provided. This system applies vision-based tungsten carbide mold defect detection methods, including: a thermal compensation correction module, a defect detection analysis module, an optimized reflection model improvement module, and a tungsten carbide characteristic correction analysis module. The thermal compensation correction module acquires mold surface images of the batch of tungsten carbide molds to be inspected using an industrial camera and inputs them into a specified reflection model for illumination processing. Simultaneously, it performs thermal compensation correction on the specified reflection model to obtain a corresponding optimized reflection model, thereby reducing the brightness reference drift of the defect detection model at different temperature stages. The defect detection analysis module separates the reflection components of the mold surface image according to the optimized reflection model and then performs image preprocessing to obtain a standardized reflection image for input. The defect detection model performs defect detection analysis and outputs the defect detection results. The optimized reflection model improvement module is used to determine whether to re-inspect the batch of tungsten carbide molds to be inspected based on the defect detection results and obtain the corresponding defect detection feedback results. The defect detection performance analysis is used to determine whether to improve the optimized reflection model, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift. The tungsten carbide characteristic correction analysis module is used to perform tungsten carbide characteristic correction analysis if the optimized reflection model is improved, so as to introduce the influence of tungsten carbide surface characteristics on light reflection for correction, and to determine the fluctuation of the defect detection feedback results after the optimization reflection model is improved to determine whether to continue to improve the optimized reflection model. Otherwise, the reflection component separation is continued based on the optimized reflection model.

[0012] Beneficial effects

[0013] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0014] 1. The present invention provides a visual inspection-based defect detection method for tungsten steel molds. This method acquires surface images of the batch of tungsten steel molds to be inspected using an industrial camera and inputs them into a specified reflection model for illumination processing. Simultaneously, thermal compensation correction is performed on the specified reflection model to obtain a corresponding optimized reflection model. This not only reduces the brightness reference drift of the defect detection model at different temperature stages but also avoids the problem of conventional Lambert or Phong models having constant parameters and being unable to dynamically adapt to temperature changes. It achieves dynamic temperature correction of the parameters in the reflection model, ensuring that the optimized reflection model maintains a consistent brightness reference across different temperature ranges. Then, after separating the reflection components of the mold surface image based on the optimized reflection model, image preprocessing is performed to obtain a standardized reflection image, which is then input into the defect detection model for defect detection analysis and output of defect detection results. This solves the traditional problem of uneven brightness caused by the mixing of diffuse and specular reflections and provides a unified illumination condition for the defect detection model, improving the repeatability of detection results and helping to suppress false detections and missed detections caused by high reflectivity or uneven illumination. The defect detection results determine whether to re-inspect the batch of tungsten carbide molds to be inspected and obtain the corresponding defect detection feedback results. The defect detection performance analysis determines whether to improve the optimized reflection model, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift. This improves the adaptability of the optimized reflection model to different batches of tungsten carbide molds, compensating for the shortcomings of different batches of molds having different surface oxidation levels, processing precision, and reflection characteristics. Finally, if the optimized reflection model is improved, a tungsten carbide characteristic correction analysis is performed to correct the influence of tungsten carbide surface characteristics on light reflection. The fluctuation of the defect detection feedback results after the improvement of the optimized reflection model is judged to determine whether to continue improving the optimized reflection model. Otherwise, reflection component separation is continued based on the optimized reflection model, so that the reflection model can more realistically represent the anisotropy of the tungsten carbide surface and the influence of the oxide layer, improving the physical accuracy of the model and thus improving the accuracy of tungsten carbide mold defect detection. This effectively solves the problem in existing technologies where inaccurate visual inspection results due to the surface characteristics of tungsten carbide lead to a decrease in defect detection accuracy.

[0015] 2. This invention extracts thermal correction coefficients for thermal compensation terms from a pre-set database. These coefficients include correction coefficients for the reference mold surface temperature, diffuse reflection compensation term, specular reflection compensation term, and specular reflection mold temperature. By correcting the output of the specified reflection model through these thermal compensation terms, the brightness consistency at different temperature stages is maintained, thereby improving the stability of the input image of the defect detection model. This fills the gap in the existing technology where temperature compensation lacks a physical separation mechanism. Next, the mold surface temperature distribution is obtained and substituted into the thermal correction function obtained by pre-regression fitting along with the thermal correction coefficients to obtain the corresponding thermal compensation terms. This helps to solve the problem of reflection coefficient drift caused by temperature changes in the existing technology, allowing the temperature correction of diffuse reflection coefficient and specular reflection coefficient to be performed independently, which is more in line with the actual physical characteristics of tungsten steel. Finally, the thermal compensation terms are substituted into the specified reflection model to obtain the corresponding optimized reflection model to correct the influence of light reflection. This enables the model to have adaptive capabilities for different mold materials, surface conditions, and temperature ranges, thereby improving the stability and consistency of defect detection results.

[0016] 3. By acquiring the tungsten steel characteristic data of the batch of tungsten steel molds to be inspected, and based on the tungsten steel characteristic data and the tungsten steel characteristic reference data extracted from the preset database, a tungsten steel characteristic influence index was obtained. This enabled a more accurate quantification of the influence of different grain orientations and oxide layer thickness differences on the diffuse reflection coefficient and specular reflection coefficient of the tungsten steel surface. This solved the problem that existing technologies often neglect the differences in different grain orientations and oxide layer thicknesses on the tungsten steel surface. Then, based on the tungsten steel characteristic influence index, matching was performed on the tungsten steel characteristic correction dataset after training to obtain the corresponding reflection coefficient correction amount, thereby improving the optimized reflection model. This improved the physical accuracy and adaptability of the optimized reflection model, filling the gap in the lack of a dynamic update mechanism for tungsten steel surface characteristic correction in traditional technologies. If the batch of tungsten steel molds to be inspected changes, the tungsten steel characteristic data of the batch of tungsten steel molds to be inspected is reacquired for tungsten steel characteristic-correction reflection model analysis, thus ensuring that the model can more accurately reflect changes in surface characteristics under different times and environments, thereby improving the accuracy of defect detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a vision-based defect detection method for tungsten steel molds provided in this application embodiment;

[0019] Figure 2 A schematic diagram of the visualization of defect detection results provided in the embodiments of this application;

[0020] Figure 3 This is a schematic diagram of the process for performing defect detection performance analysis and judgment provided in an embodiment of this application;

[0021] Figure 4 This is a schematic diagram illustrating the process of improving the optimized reflection model according to an embodiment of this application;

[0022] Figure 5 A schematic diagram of the structure of the visual inspection-based tungsten steel mold defect detection system provided in the embodiments of this application. Detailed Implementation

[0023] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0024] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0025] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0026] like Figure 1 The diagram shows a flowchart of a vision-based defect detection method for tungsten carbide molds provided in this application embodiment. The method includes the following steps:

[0027] Step 1: Acquire mold surface images of the batch of tungsten steel molds to be inspected using an industrial camera, and input them into a specified reflection model for illumination processing. At the same time, perform thermal compensation correction on the specified reflection model to obtain the corresponding optimized reflection model, so as to reduce the brightness reference drift of the defect detection model at different temperature stages.

[0028] It should be explained that the specified reflection model is the Lambert-Phong hybrid model, which simulates or separates the light reflection characteristics of an object's surface by jointly describing diffuse reflection and specular reflection. This provides a physically consistent reflection model for defect detection, brightness correction, and material characterization. The light source direction vector, viewing direction vector, surface normal vector, and reflection parameters (diffuse reflection coefficient, specular reflection coefficient, specular index) are input into the Lambert-Phong hybrid model, and the corresponding surface reflection brightness, diffuse reflection component, and specular reflection component are output.

[0029] Step 2: After separating the reflection components of the mold surface image according to the optimized reflection model, image preprocessing is performed to obtain a standardized reflection image, which is then input into the defect detection model for defect detection analysis and output of defect detection results.

[0030] It should be explained that defect detection models are typically CNN, ResNet, EfficientNet, YOLO, Faster R-CNN, or RetinaNet. The inputs are a standardized reflection image, reflection component features (diffuse reflection image, specular reflection image, brightness gradient, polarization features, etc.), and multi-channel input (RGB, grayscale, polarization channel, depth channel). The outputs are the corresponding defect segmentation map, defect location and bounding box, defect category, defect confidence or score, and statistical indicators (such as defect area, number, distribution, and severity).

[0031] Step 3: Determine whether to re-inspect the batch of tungsten steel molds to be inspected based on the defect detection results and obtain the corresponding defect detection feedback results. Then, through the defect detection performance analysis, determine whether to improve the optimized reflection model, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift.

[0032] Step four: If the optimized reflection model is to be improved, a tungsten steel characteristic correction analysis is performed to correct the influence of tungsten steel surface characteristics on light reflection. The fluctuation of the defect detection feedback results after the optimization reflection model is improved is determined to determine whether to continue to improve the optimized reflection model, thereby reducing the negative correction impact on the optimized reflection model. Otherwise, the reflection component separation is continued based on the optimized reflection model.

[0033] In this embodiment, a defect detection process for tungsten carbide molds based on a reflection model and thermal compensation correction is constructed, achieving coordinated correction of illumination conditions, temperature environment, and material properties. Furthermore, this method, through steps such as optimizing the reflection model, separating reflection components, dynamic feedback updates, and characteristic correction analysis, solves the problem of unstable detection accuracy caused by changes in illumination, temperature drift, and differences in the reflective properties of tungsten carbide materials in traditional visual inspection. It helps to offset brightness drift caused by temperature and ambient light changes, improves the contrast and detectability of surface defects, achieves adaptive coordination between the detection model and the reflection model, and enhances the generalization ability of a specified reflection model for tungsten carbide molds of different material batches and surface conditions. It avoids parameter drift caused by over-correction, thereby improving the detection accuracy and long-term stability of the tungsten carbide mold defect detection system.

[0034] It should be added that the image preprocessing first separates the reflection components of the mold surface image according to the optimized reflection model, obtaining the corresponding diffuse reflection component and specular component; then, the diffuse reflection component is normalized to reduce the surface brightness difference caused by the angle, and specular component is subjected to specular suppression or mask removal to reduce specular false defects; finally, the mold surface image is filtered and denoised to preserve the geometric structure of the defect and suppress noise, resulting in a standardized reflection image after illumination compensation; if the specular highlight is not saturated and contains recoverable texture information, specular highlight suppression (preserving information) is preferentially performed on the specular component; if the specular highlight is saturated or the area is large and unrecoverable, mask removal (combined with re-inspection or inpainting) is used, and the final strategy is selected according to the capability of the downstream detection model and engineering constraints.

[0035] Specifically, the defect detection results include defect type classification, geometric feature extraction, defect level determination, and defect result output and statistics. Defect type classification involves analyzing the acquired mold surface images, automatically identifying and distinguishing different types of surface defects using visual information including but not limited to texture features, brightness anomalies, and reflection structure damage, achieving semantic-level classification of defects. Defect types include, but are not limited to, scratches, pits, cracks, chipped corners, and dents. Geometric feature extraction performs pixel-level geometric analysis on the detected defect areas, extracting the size, shape, and spatial distribution features of the defects, providing accurate measurement data for quantitative assessment and level classification, including length, area, and depth estimation combined with morphological information. Defect level determination quantifies the severity of defects based on their type and geometric features, combined with industry standards or enterprise quality specifications, determining whether they are qualified or unqualified as the judgment result and level score, achieving automated quality control, and comparing size or optical features with thresholds. Defect result output and statistics visualize the defect location coordinates, defect type, and defect level in a standardized reflection image and statistically analyze the defect data, including the number of defects, defect density, and defect distribution.

[0036] like Figure 2 The diagram shown is a visualization of defect detection results provided in this application embodiment. Various defect categories are marked, including silk spots, welding lines, inclusions, and water spots. In addition, defects also include punching holes, crescent gaps, oil spots, waist folding, creases, and rolled pits. Figure 2 The marking not only visualizes the defect categories on the surface of tungsten carbide molds, but also enables the location of defects on the surface of tungsten carbide molds, which helps to manage tungsten carbide molds with defects more efficiently.

[0037] Furthermore, thermal compensation correction is performed on the specified reflection model to obtain the corresponding optimized reflection model. The specific process is as follows:

[0038] The first step is to extract the thermal correction coefficients for thermal compensation items from the preset database. The thermal correction coefficients include the reference mold surface temperature, the correction coefficients for diffuse reflection compensation items, the correction coefficients for specular reflection compensation items, and the correction item for specular reflection mold temperature.

[0039] It should be added that the thermal correction coefficients are all preset by the pre-set testers, who pre-set and store them in the preset database based on historical data and empirical rules.

[0040] The second step is to obtain the temperature distribution on the mold surface and substitute it together with the thermal correction coefficient into the thermal correction function obtained by pre-regression fitting to obtain the corresponding thermal compensation term. The thermal compensation term includes diffuse reflection compensation term and specular reflection compensation term, and the thermal correction function includes diffuse reflection thermal correction function and specular reflection thermal correction function.

[0041] Specifically, diffuse reflection thermal correction function Where a1 represents the diffuse reflection compensation term correction coefficient, T0 represents the reference mold surface temperature, T is the mold surface temperature, and g is the specular reflection thermal correction function. , where a2 represents the correction coefficient for the specular reflection compensation term, and b represents the correction term for the temperature of the specular reflection mold.

[0042] The third step is to substitute the thermal compensation term into the initial specified reflection model to obtain the corresponding optimized reflection model to correct the influence of light reflection.

[0043] Specifically, the expression for optimizing the reflection model is as follows:

[0044] ;

[0045] In the formula, I(x,y) represents the pixel brightness value on the mold surface, T(x,y) represents the temperature distribution on the mold surface, and k d Let L represent the diffuse reflection coefficient, θ represent the incident light intensity, and k represent the incident angle. s denoted by α, where α represents the reflection angle deviation, and n represents the smoothness index (specular index); the diffuse reflection coefficient is obtained by preset testers through calibration or experimental fitting using a standard diffuse reflection plate, the specular reflection coefficient is obtained by preset testers through fitting the grayscale peak of the bright spot region of the specular surface, and the smoothness index is obtained by preset testers through fitting based on the half-width curve of the grayscale distribution.

[0046] In this embodiment, automatic temperature compensation is achieved by optimizing the reflection model, which stabilizes the brightness curve. Furthermore, the specular reflection thermal correction function suppresses high-reflectivity false defects, enabling dynamic compensation for changes in the thermal environment. In addition, by introducing a temperature distribution-based thermal compensation mechanism into the specified reflection model, the reflection characteristics of the tungsten steel mold surface under different thermal states can be dynamically corrected. This method utilizes the diffuse and specular reflection thermal correction functions obtained by regression fitting, combined with preset thermal correction coefficients in the database, to achieve more accurate temperature compensation for reflection parameters. This helps to solve the problems of brightness reference drift, false detection, and model inaccuracy caused by temperature changes in traditional visual inspection, thereby improving the stability and consistency of defect detection results.

[0047] like Figure 3The diagram shown is a flowchart illustrating the defect detection performance analysis and judgment process provided in this application embodiment. The specific logic is as follows: 1) Obtain defect detection feedback data from manual defect re-inspection of the tungsten carbide mold based on the defect detection results; 2) Perform weighted coupling processing on the defect detection feedback data using set feedback weights to obtain the corresponding defect detection performance score; 3) Determine whether the defect detection performance score is greater than the defect detection performance threshold. If the defect detection performance score is greater than the defect detection performance threshold, continue to separate the reflection components based on the current optimized reflection model; otherwise, perform tungsten carbide characteristic-corrected reflection model analysis. Through the above process, adaptive control of the tungsten carbide mold defect detection system is achieved, improving the accuracy of defect detection in tungsten carbide molds.

[0048] Furthermore, a defect detection performance analysis is conducted to determine whether the optimized reflection model needs improvement. The specific process is as follows:

[0049] Obtain defect detection feedback data for manual defect re-inspection of tungsten steel molds based on defect detection results. The defect detection feedback data includes defect detection accuracy rate, defect detection false detection rate, and defect detection missed detection rate.

[0050] By assigning weights to the defect detection feedback data, the corresponding defect detection performance score is obtained. The defect detection performance score is used to quantify the performance of the defect detection model in detecting defects in tungsten steel molds. Specifically, the defect detection performance score is obtained by multiplying the defect detection accuracy rate, the difference between the value 1 and the defect detection false positive rate, and the difference between the value 1 and the defect detection false negative rate with the corresponding feedback weights, and then performing a coupling operation. The feedback weights include the defect detection accuracy rate weight, the defect detection false positive rate weight, and the defect detection false negative rate weight, and the sum of the three is 1. They are usually preset by the testers based on the importance of each defect detection feedback data and stored in a preset database in advance.

[0051] The defect detection performance score is compared with the preset defect detection performance threshold. The specific determination method is as follows: if the defect detection performance score is greater than the defect detection performance threshold, the reflection component separation is continued based on the current optimized reflection model; otherwise, the tungsten steel characteristic-corrected reflection model analysis is performed to improve the optimized reflection model. The defect detection performance threshold represents the minimum value that limits the performance of the defect detection model to be qualified, and is generally preset by the preset test personnel.

[0052] In this embodiment, through the design of the above steps, the system introduces a performance feedback closed-loop mechanism based on the detection results during the defect detection process of tungsten steel molds, thereby realizing adaptive dynamic optimization of the reflection model and the detection model. Secondly, by setting feedback weights to weighted couple different types of detection indicators to obtain a unified defect detection performance score, the system achieves normalization and quantification of multi-dimensional performance parameters, facilitating dynamic comparison and judgment of model performance. This mechanism can dynamically adjust the weights of different indicators according to production requirements or detection scenarios, such as increasing the weight of the false negative rate in high reliability scenarios, thereby making the system's performance evaluation more targeted and controllable. By judging the defect detection performance score against a preset performance threshold, the system achieves automatic decision control of model performance. This closed-loop feedback structure enables the system to have the ability to self-evaluate, self-correct, and self-optimize, helping to prevent the problem of decreased detection accuracy over time or batch fluctuations. This realizes the intelligent and adaptive control of the tungsten steel mold defect detection system, improves the robustness and long-term stability of the defect detection model in complex environments, ensures the continuous high accuracy of detection results, and provides reliable data-driven and model self-optimization support for high-precision quality control of tungsten steel molds.

[0053] It should be noted that the specific process of performing the tungsten carbide property-corrected reflection model analysis is as follows:

[0054] First, obtain the tungsten steel characteristic data of the batch of tungsten steel molds to be tested. The tungsten steel characteristic data includes the real part of the refractive index, the degree of linear polarization, the polarization angle, and the spectral reflectance.

[0055] Specifically, the real part of the refractive index is obtained using an ellipsometer, the degree of linear polarization and the polarization angle are obtained using a polarization camera, and the spectral reflectance is obtained using a spectrophotometer.

[0056] Next, based on the tungsten steel characteristic data and the tungsten steel characteristic reference data extracted from the preset database, a tungsten steel characteristic influence index is obtained to quantify the influence of different grain orientations and oxide layer thickness differences on the diffuse reflection coefficient and specular reflection coefficient of tungsten steel surface.

[0057] Among them, the tungsten steel property influence index α W The specific constraint expression is as follows:

[0058] ;

[0059] In the formula, i represents the real part of the refractive index, P represents the degree of linear polarization, and θ PR(λ) represents the polarization angle, R(λ) represents the spectral reflectance, where λ represents the wavelength, and ω1, ω2, ω3, and ω4 represent the weighting factors corresponding to each tungsten steel characteristic data. These factors reflect the relative importance of each tungsten steel characteristic data in the reflection behavior of the tungsten steel surface. They are generally tungsten steel characteristic reference data read from a preset database and are preset by the testers based on historical data and experience rules.

[0060] It should be explained that by integrating the spectral reflectance function R(λ) over a certain wavelength range, R(λ) represents the reflectance of light at a specific wavelength λ on the tungsten steel surface. It represents the total reflectance of the tungsten steel surface within the specified wavelength range, and the upper and lower limits of the integration are determined by the actual wavelength range. For example, the value range of λ is from 400nm to 700nm.

[0061] It should be noted that the real part of the refractive index affects the reflection intensity; an increase in refractive index enhances specular reflection. Therefore, the real part of the refractive index is positively correlated with specular reflection intensity, meaning that a higher refractive index corresponds to stronger specular reflection. The degree of linear polarization affects the polarization characteristics of specular reflection, significantly influencing the specular reflection coefficient. Generally, a higher degree of linear polarization results in a greater influence on the tungsten steel's properties, indicating higher overall reflectivity. The degree of linear polarization is positively correlated with overall reflectivity; a higher degree of linear polarization implies stronger polarization characteristics, thus affecting specular reflection. Changes in the polarization angle affect the specular reflection distribution on the surface, influencing the intensity and directionality of specular reflection. The influence of the polarization angle on the tungsten steel's properties is indicated by... The influence of the number depends on the specific value of the tungsten steel property influence index. The cosine value of the polarization angle varies between -1 and 1, affecting the overall reflectivity. When the cosine value of the polarization angle is positive, the change in the polarization angle will increase the overall reflectivity, and when the cosine value of the polarization angle is negative, the change in the polarization angle will decrease the overall reflectivity. Therefore, the influence of the polarization angle on the overall reflectivity is non-linear and depends on the value of the corresponding cosine function. Spectral reflectivity affects the intensity and wavelength dependence of diffuse reflection and specular reflection. The integral term in the formula represents the total reflectivity of the tungsten steel surface within a specified wavelength range. Spectral reflectivity is positively correlated with overall reflectivity. The larger the spectral reflectivity, the larger the integral value, and the higher the overall reflectivity.

[0062] Then, based on the tungsten steel property influence index, the corresponding reflection coefficient correction amount is obtained by matching in the tungsten steel property correction dataset after training, so as to improve the optimized reflection model. The reflection coefficient correction amount includes diffuse reflection correction amount and specular reflection correction amount.

[0063] It should be added that the tungsten steel characteristic influence index is input into the tungsten steel characteristic correction dataset, and the reflection coefficient correction amount is output. The tungsten steel characteristic correction dataset is obtained by pre-training by professional technicians according to a set training process. The training process is as follows: the tungsten steel characteristic influence index obtained in the historical time period and the reflection coefficient correction amount set by the professional technicians based on empirical rules are input into the initial dataset constructed by the logistic regression algorithm, the cross-entropy loss function is used as the optimization criterion, and the training is completed through the scikit-learn framework to obtain the corresponding tungsten steel characteristic correction dataset, which reflects the mapping relationship between the tungsten steel characteristic influence index and the reflection coefficient correction amount. In actual use, the real-time obtained tungsten steel characteristic influence index is input into the tungsten steel characteristic correction dataset to output the corresponding reflection coefficient correction amount.

[0064] Finally, if the batch of tungsten carbide molds to be tested changes, the tungsten carbide characteristic data of the batch of tungsten carbide molds to be tested are re-acquired for tungsten carbide characteristic-corrected reflection model analysis.

[0065] In this embodiment, a dynamic reflection model correction method based on tungsten carbide surface characteristic data is proposed. This method quantifies the influence of physical properties such as grain orientation, oxide layer thickness, and surface roughness on the reflection coefficient, and combines this with preset correction data in a database to more accurately correct the reflection model of the tungsten carbide mold surface. By acquiring the physical characteristic data of each batch of tungsten carbide molds and calculating the tungsten carbide characteristic influence index, this method achieves personalized correction for each batch, eliminating errors in traditional reflection models and improving the accuracy and reliability of defect detection. When the production batch of tungsten carbide molds changes, new tungsten carbide characteristic data can be dynamically acquired for recalibration of the reflection model, ensuring that detection accuracy is unaffected by changes in time and environment. Furthermore, compared to existing technologies, this method, through more accurate reflection coefficient correction and a dynamic update mechanism, effectively solves the detection error problems caused by tungsten carbide surface non-uniformity, material differences, and batch variations, thereby improving the accuracy of defect detection in tungsten carbide molds.

[0066] In addition to performing tungsten carbide property-corrected reflection model analysis, the process also includes determining the conformity of the batch of tungsten carbide molds to be tested.

[0067] The defect detection accuracy rate is compared with the lowest defect detection pass accuracy rate read from the preset database. The lowest defect detection pass accuracy rate is generally set manually by preset testers based on historical data and experience rules, and stored in the preset database.

[0068] If the defect detection accuracy rate is greater than the minimum defect detection pass accuracy rate, it means that the current batch of tungsten carbide molds to be inspected is qualified, and then defect detection will continue for the next batch of tungsten carbide molds.

[0069] If the defect detection accuracy rate is not greater than the minimum defect detection pass accuracy rate, it means that the current batch of tungsten carbide molds to be inspected is unqualified, and the current batch of tungsten carbide molds to be inspected will be re-inspected.

[0070] In this embodiment, an automatic quality judgment and re-inspection mechanism based on a detection accuracy threshold is introduced into the tungsten carbide mold defect detection process. This achieves intelligent closed-loop quality control of the detection process. By comparing the defect detection accuracy of the current batch with the lowest defect detection pass accuracy recorded in the preset database, quantitative judgment of the detection result quality is achieved. A traceable, standardized detection pass threshold is established, avoiding deviations caused by subjective human judgment, thus ensuring the consistency and objectivity of the detection process. This helps to avoid omissions and misidentifications caused by model misjudgment or changes in lighting environment. Furthermore, the introduction of this judgment and re-inspection logic forms a two-layer protection structure of "standard threshold judgment and intelligent re-inspection closed loop" in terms of detection quality assurance. On the one hand, more precise threshold control ensures the passability of each batch of detection results. On the other hand, the automatic re-inspection mechanism ensures the immediate correction and reconfirmation of abnormal batches, thereby significantly reducing the risk of missed and false detections. This helps to improve the quality control level and automation of the tungsten carbide mold defect detection process, achieving comprehensive optimization of detection efficiency, accuracy, and stability.

[0071] like Figure 4 The diagram shows a flowchart illustrating the improvement of the optimized reflection model provided in this application embodiment. The specific logic is as follows: The reflection coefficient correction value is input into the optimized reflection model to correct the diffuse reflection coefficient and specular reflection coefficient respectively; the batch of tungsten carbide molds to be inspected is re-inspected, and the defect detection performance score after re-inspection is obtained and recorded as the re-inspection defect detection performance score; if the re-inspection defect detection performance score is greater than the defect detection performance threshold, reflection component separation is continued based on the improved optimized reflection model; if the re-inspection defect detection performance score is not greater than the defect detection performance threshold, the improvement direction is determined and optimized; through the above process, further optimization of the optimized reflection model is achieved, ensuring the stability and reliability of defect detection in tungsten carbide molds.

[0072] Furthermore, the specific methods for improving the optimized reflection model are as follows:

[0073] The reflection coefficient correction value is input into the optimized reflection model to correct the diffuse reflection coefficient and specular reflection coefficient respectively; the tungsten steel mold of the batch to be inspected is re-inspected, and the defect detection performance score after re-inspection is obtained and recorded as the re-inspection defect detection performance score; if the re-inspection defect detection performance score is greater than the defect detection performance threshold, the reflection component separation is continued based on the improved optimized reflection model; otherwise, the improvement direction is determined and optimized.

[0074] In this embodiment, a dynamic closed-loop mechanism is established, encompassing the reflection model, detection performance, and model optimization. This mechanism enables bidirectional linkage optimization of reflection model accuracy and detection performance. Furthermore, through a performance feedback-driven iterative update mechanism for the reflection model, the system possesses continuous optimization capabilities. It can dynamically adjust model parameters based on actual detection performance, ensuring that the reflection model maintains high accuracy and robustness over the long term. The solution provided in this embodiment helps address the problem of traditional reflection models being static and unable to self-correct with changes in the detection environment. It not only achieves self-evaluation, self-correction, and self-optimization in the defect detection process of tungsten steel molds but also improves the system's detection accuracy, stability, and intelligence level.

[0075] As a further embodiment, the specific process for determining the direction of improvement and optimization is as follows:

[0076] The change in defect detection performance is obtained by comparing the re-inspection defect detection performance score with the original defect detection performance score.

[0077] If the change in defect detection performance exceeds the set value, reflection component separation will continue based on the improved optimized reflection model; the set value is generally 1.

[0078] If the change in defect detection performance equals the set value, then reflection component separation continues based on the improved optimized reflection model, and the change in defect detection performance obtained in the next defect detection is monitored to see if it is greater than the set value. If the change in defect detection performance obtained in the next defect detection is greater than the set value, then reflection component separation continues based on the improved optimized reflection model; otherwise, the improvement of the optimized reflection model is stopped and an error is reported to the inspection personnel's terminal.

[0079] If the change in defect detection performance is less than the set value, then stop improving the optimized reflection model and report an error to the inspection personnel's terminal.

[0080] In this embodiment, through an improved decision-making mechanism driven by "defect detection performance change", the system achieves a more refined, adaptive, and error-proof reflection model iteration process. This mechanism not only helps improve the stability of the reflection model in complex environments but also sustainably maintains the dynamic optimal state of detection performance, preventing the reflection model from drifting over time. In addition, this scheme strengthens the anomaly warning capability of the detection system, enabling timely error reporting when model optimization is ineffective or performance degrades, ensuring the safety and reliability of quality control. The method provided in this embodiment helps improve the intelligence, controllability, and long-term stability of the tungsten steel mold defect detection system, ensuring that the detection process does not experience continuous false detections or missed detections due to abnormal model updates.

[0081] like Figure 5 The diagram shown is a structural schematic of a visual inspection-based tungsten steel mold defect detection system provided in this application embodiment, including: a thermal compensation correction module, a defect detection analysis module, an optimized reflection model improvement module, and a tungsten steel property correction analysis module.

[0082] The thermal compensation correction module is used to acquire mold surface images of the batch of tungsten steel molds to be inspected based on an industrial camera, and input them into a specified reflection model for illumination processing. At the same time, thermal compensation correction is performed on the specified reflection model to obtain the corresponding optimized reflection model, so as to reduce the brightness reference drift of the defect detection model at different temperature stages.

[0083] The defect detection and analysis module is used to separate the reflection components of the mold surface image according to the optimized reflection model, and then perform image preprocessing to obtain a standardized reflection image, which is then input into the defect detection model for defect detection and analysis and outputs the defect detection results.

[0084] The optimized reflection model improvement module is used to determine whether to re-inspect the batch of tungsten steel molds to be inspected based on the defect detection results and to obtain the corresponding defect detection feedback results. The defect detection performance analysis is then used to determine whether to improve the optimized reflection model, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift.

[0085] The tungsten steel characteristic correction analysis module is used to perform tungsten steel characteristic correction analysis if the optimized reflection model is improved, so as to introduce the influence of tungsten steel surface characteristics on light reflection for correction, and to determine the fluctuation of the defect detection feedback results after the improvement of the optimized reflection model to determine whether to continue to improve the optimized reflection model, thereby reducing the negative correction impact on the optimized reflection model; otherwise, the reflection component separation is continued based on the optimized reflection model.

[0086] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0087] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0088] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units, located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0089] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0090] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the solution, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions 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. A method for defect detection in tungsten carbide molds based on visual inspection, characterized in that, Includes the following steps: The mold surface image of the batch of tungsten steel molds to be inspected is acquired by an industrial camera and input into a specified reflection model for illumination processing. At the same time, the specified reflection model is thermally compensated and corrected to obtain the corresponding optimized reflection model, so as to reduce the brightness reference drift of the defect detection model at different temperature stages. The specific process for performing thermal compensation correction on the specified reflection model to obtain the corresponding optimized reflection model is as follows: The thermal correction coefficients for thermal compensation terms are extracted from a preset database. These thermal correction coefficients include the reference mold surface temperature, the diffuse reflection compensation term correction coefficient, the specular reflection compensation term correction coefficient, and the specular reflection mold temperature correction term. It should be added that the thermal correction coefficients are all preset by the pre-set testers, who pre-set and store them in the preset database based on historical data and empirical rules; The surface temperature distribution of the mold is obtained and substituted together with the thermal correction coefficient into the thermal correction function obtained by pre-regression fitting to obtain the corresponding thermal compensation term. The thermal compensation term includes diffuse reflection compensation term and specular reflection compensation term. The thermal correction function includes diffuse reflection thermal correction function and specular reflection thermal correction function. Substituting the thermal compensation term into the specified reflection model yields the corresponding optimized reflection model to correct for the effects of light reflection. After separating the reflection components of the mold surface image according to the optimized reflection model, the image preprocessing is then performed to obtain a standardized reflection image, which is then input into the defect detection model for defect detection analysis and output of defect detection results. The defect detection results determine whether to re-inspect the batch of tungsten carbide molds to be inspected and obtain the corresponding defect detection feedback results. The defect detection performance analysis is then used to determine whether to improve the optimized reflection model, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift. If the optimized reflection model is improved, a tungsten steel characteristic correction analysis is performed to correct the influence of tungsten steel surface characteristics on light reflection. The fluctuation of the defect detection feedback results after the optimization reflection model is improved is determined to determine whether to continue to improve the optimized reflection model. Otherwise, the reflection component separation is continued based on the optimized reflection model.

2. The method for detecting defects in tungsten carbide molds based on visual inspection as described in claim 1, characterized in that: The specific process for obtaining the standardized reflectance image is as follows: Based on the optimized reflection model, the reflection component of the mold surface image is separated to obtain the corresponding diffuse reflection component and specular component. The diffuse component is grayscale normalized to reduce the surface brightness difference caused by the angle, and the specular component is specular suppression or mask removal to reduce specular artifacts. The mold surface image was filtered and denoised to preserve the geometry of the defects and suppress noise, resulting in a standardized reflection image.

3. The visual inspection-based defect detection method for tungsten carbide molds as described in claim 1, characterized in that: The defect detection results include defect type classification, geometric feature extraction, defect level determination, and defect result output and statistics. The defect type classification is used to analyze the acquired mold surface images. By automatically identifying and distinguishing different types of surface defects, including texture features, brightness anomalies, and reflection structure damage, the semantic level classification of defects is achieved. The defect types include scratches, pits, cracks, chipped corners, and dents. The geometric feature extraction is used to perform pixel-level geometric analysis on the detected defect area, extract the size, shape and spatial distribution features of the defect, and provide measurement data for quantitative assessment and classification. The defect level determination is used to quantitatively assess the severity of a defect based on its type and geometric characteristics, thereby determining whether it is qualified or unqualified as the judgment result and level score. The defect result output and statistics refer to visualizing the image results of defect location coordinates, defect type and defect level in a standardized reflection image and statistically analyzing the defect data, which includes the number of defects, defect density and defect distribution.

4. The visual inspection-based defect detection method for tungsten carbide molds as described in claim 1, characterized in that: The process of determining whether to improve the optimized reflection model through defect detection performance analysis is as follows: Obtain defect detection feedback data for manual defect re-inspection of tungsten steel molds based on defect detection results. The defect detection feedback data includes defect detection accuracy rate, defect detection false detection rate, and defect detection missed detection rate. By setting feedback weights, the defect detection feedback data is weighted and coupled to obtain the corresponding defect detection performance score. The defect detection performance score is used to quantify the performance of the defect detection model in detecting defects in tungsten steel molds. The defect detection performance score is compared with the preset defect detection performance threshold. The specific determination method is as follows: if the defect detection performance score is greater than the defect detection performance threshold, the reflection component separation is continued based on the current optimized reflection model; otherwise, the tungsten steel characteristic-corrected reflection model analysis is performed to improve the optimized reflection model. The defect detection performance threshold represents the minimum value that limits the performance of the defect detection model to be considered acceptable.

5. The visual inspection-based defect detection method for tungsten carbide molds as described in claim 4, characterized in that: The specific process for performing the tungsten steel property-corrected reflection model analysis is as follows: Obtain tungsten steel characteristic data of the batch of tungsten steel molds to be tested, including the real part of refractive index, degree of linear polarization, polarization angle and spectral reflectance; Based on tungsten steel property data and tungsten steel property reference data extracted from a preset database, a tungsten steel property influence index is obtained to quantify the degree of influence of different grain orientations and oxide layer thickness differences on the diffuse reflection coefficient and specular reflection coefficient of tungsten steel surface. Based on the tungsten steel property influence index, the corresponding reflection coefficient correction amount is obtained by matching in the tungsten steel property correction dataset after training, so as to improve the optimized reflection model. The reflection coefficient correction amount includes diffuse reflection correction amount and specular reflection correction amount. If the batch of tungsten carbide molds to be tested changes, the tungsten carbide characteristic data of the batch of tungsten carbide molds to be tested should be re-acquired for tungsten carbide characteristic-corrected reflection model analysis.

6. The visual inspection-based defect detection method for tungsten carbide molds as described in claim 5, characterized in that: In addition to performing tungsten carbide characteristic-corrected reflection model analysis, the process also includes determining the conformity of the batch of tungsten carbide molds to be tested. The defect detection accuracy rate is compared with the lowest defect detection pass accuracy rate read from a preset database: If the defect detection accuracy rate is greater than the minimum defect detection pass accuracy rate, it means that the current batch of tungsten carbide molds to be inspected is qualified, and defect detection will continue for the next batch of tungsten carbide molds. If the defect detection accuracy rate is not greater than the minimum defect detection pass accuracy rate, it means that the current batch of tungsten carbide molds to be inspected is unqualified, and the current batch of tungsten carbide molds to be inspected will be re-inspected.

7. The visual inspection-based defect detection method for tungsten carbide molds as described in claim 4, characterized in that: The specific methods for improving the optimized reflection model are as follows: The reflection coefficient correction value is input into the optimized reflection model to correct the diffuse reflection coefficient and specular reflection coefficient respectively; The batch of tungsten carbide molds to be inspected is re-inspected, and the defect detection performance score after the re-inspection is obtained and recorded as the re-inspection defect detection performance score. If the re-detection defect detection performance score is greater than the defect detection performance threshold, then reflection component separation will continue based on the improved optimized reflection model. If the re-inspection defect detection performance score is not greater than the defect detection performance threshold, then the direction of improvement will be determined and optimized.

8. The method for detecting defects in tungsten carbide molds based on visual inspection as described in claim 7, characterized in that: The specific process for determining and optimizing the direction of improvement is as follows: The change in defect detection performance is obtained by comparing the re-inspection defect detection performance score with the defect detection performance score. If the change in defect detection performance exceeds the set value, then reflection component separation will continue based on the improved optimized reflection model. If the change in defect detection performance equals the set value, then reflection component separation continues based on the improved optimized reflection model, and the change in defect detection performance obtained in the next defect detection is monitored to see if it is greater than the set value. If the change in defect detection performance obtained in the next defect detection is greater than the set value, then reflection component separation continues based on the improved optimized reflection model; otherwise, the improvement of the optimized reflection model is stopped and an error is reported to the detection personnel's terminal. If the change in defect detection performance is less than the set value, then stop improving the optimized reflection model and report an error to the inspection personnel's terminal.

9. A visual inspection-based tungsten carbide mold defect detection system, employing the visual inspection-based tungsten carbide mold defect detection method as described in any one of claims 1-8, characterized in that, include: Thermal compensation correction module, defect detection and analysis module, optimized reflection model improvement module, and tungsten steel property correction and analysis module; The thermal compensation correction module is used to acquire mold surface images of the batch of tungsten steel molds to be inspected based on an industrial camera, and input them into a specified reflection model for illumination processing. At the same time, thermal compensation correction is performed on the specified reflection model to obtain a corresponding optimized reflection model, so as to reduce the brightness reference drift of the defect detection model at different temperature stages. The defect detection and analysis module is used to separate the reflection components of the mold surface image according to the optimized reflection model, and then perform image preprocessing to obtain a standardized reflection image, which is then input into the defect detection model for defect detection and analysis and outputs the defect detection results. The optimized reflection model improvement module is used to determine whether to re-inspect the batch of tungsten steel molds to be inspected based on the defect detection results and to obtain the corresponding defect detection feedback results. The defect detection performance analysis is then used to determine whether to improve the optimized reflection model, thereby improving the accuracy of the optimized reflection model in handling brightness reference drift. The tungsten steel characteristic correction and analysis module is used to perform tungsten steel characteristic correction analysis to correct the influence of tungsten steel surface characteristics on light reflection if the optimized reflection model is improved, and to determine the fluctuation of the defect detection feedback results after the optimization reflection model is improved to determine whether to continue to improve the optimized reflection model; otherwise, it continues to separate the reflection components based on the optimized reflection model.

Images