An image defect intelligent identification system for industrial detection

Abstract

This invention relates to the field of image analysis technology, and more particularly to an intelligent image defect recognition system for industrial inspection. The system includes a data acquisition module, an anomaly analysis module, a dynamic introduction module, a coupling processing module, and a detection and recognition module. Based on the intrinsic influence coefficient determined by the anomaly analysis module using intrinsic parameters and the environmental influence coefficient determined by environmental parameters, the system calculates a recognition confusion characterization value to determine the defect recognition anomaly tendency of the detected image. For detected images with a weak defect recognition anomaly tendency, dynamic interference parameters and material interference parameters are introduced to determine the dynamic interference enhancement coefficient and the medium attenuation confusion coefficient. The system then combines the recognition confusion characterization value to calculate a defect coupling feature value to distinguish the defect anomaly tendency of the detected image, thus completing the intelligent image defect recognition. This invention combines static and dynamic perspectives to analyze image defects caused by the container itself and the external environment, achieving intelligent defect analysis and improving recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to an intelligent image defect recognition system for industrial inspection. Background Technology

[0002] During the manufacturing process, industrial products are prone to various surface defects due to multiple factors such as raw material quality, processing technology stability, equipment wear, and environmental interference. These defects not only directly affect the appearance quality and service life of the products, but may also lead to functional failures or even safety accidents during use. In recent years, with the rapid development of deep learning technology, visual inspection methods based on deep neural networks have provided a brand-new solution for industrial defect detection. Deep learning methods can automatically learn multi-level defect feature representations through a large amount of training data, without relying on manually designed feature engineering. They significantly surpass traditional methods in terms of detection accuracy, generalization ability, and robustness. However, most existing deep learning models adopt an end-to-end closed detection framework, which is difficult to flexibly adapt to the needs of rapid product model switching and collaborative detection of multiple defect types. Moreover, the high coupling between modules such as image preprocessing, feature extraction, defect classification and localization results in significant deficiencies in the maintainability and scalability of the system.

[0003] Chinese Patent Publication No. CN121685403A discloses an artificial intelligence image recognition system for industrial visual inspection, including an image recognition center, an image acquisition module, a pre-recognition verification module, a multi-dimensional recognition analysis module, an autonomous management module, a composition feedback module, and a back-end response module. This invention repeatedly tests and optimizes parameters of known standard parts to ensure the quality of image acquisition from the source. At the same time, through a dual early warning mechanism of regional deviation and type deviation, it reminds human intervention for review, which significantly reduces the false judgment rate. Furthermore, by monitoring the false judgment detection rate and sample update volume, it dynamically evaluates the performance degradation of the defect recognition model, realizing the system's self-diagnosis and guidance for optimization. It also transforms abstract defect coordinates into an intuitive defect interference distribution map, helping operation and management personnel to quickly locate weak links or problem sources in the production process.

[0004] Chinese Patent Publication No. CN112017182A discloses an industrial-grade intelligent surface defect detection method. This method constructs and trains a Siamese Generative Adversarial Network (GAN), which repairs the input image into a "normal sample" using an improved GAN. The output is then compared with manually labeled positive samples using a Siamese CNN; differences indicate defects. This invention's Siamese GAN requires no large number of samples or data augmentation, solving the common problem of limited sample sizes in industrial products. It reduces overfitting caused by limited or zero-sample deep learning, making defect detection possible for products with few defect samples and during new product development. The method utilizes cross-alignment loss function (CA) and distribution alignment loss function (DA) to strengthen the relationship between the outputs of the two networks, achieving better classification and recognition results. The use of an attention mechanism and GPU hardware improves model training speed, enabling rapid industrial deployment.

[0005] However, the following problems still exist in the existing technology. When identifying defects in transparent or semi-transparent containers, the scale of the defects is more easily confused with optical background noise because the containers themselves are multiple refractive bodies, which reduces the accuracy of image defect identification. Summary of the Invention

[0006] Therefore, the present invention provides an intelligent image defect recognition system for industrial inspection, which solves the problem that when identifying defects in transparent or semi-transparent bottles, the scale of the defects is more easily confused with optical background noise because the bottles themselves are multiple refractive bodies, thus reducing the accuracy of image defect recognition.

[0007] To achieve the above objectives, the present invention provides an intelligent image defect recognition system for industrial inspection, comprising: The data acquisition module is used to acquire the intrinsic parameters of the container corresponding to the detection image, including the refractive index of the container wall and the container wall thickness, and to acquire the environmental parameters of the container, including the light source projection angle and background stray light. An anomaly analysis module is used to calculate the identification confusion characterization value based on the intrinsic influence coefficient determined by the intrinsic parameters and the environmental influence coefficient determined by the environmental parameters, and to determine the defect identification anomaly tendency of the detected image. The dynamic input module is used to collect dynamic interference parameters and material interference parameters for detection images that show abnormal tendencies in weak defect identification. The dynamic interference parameters include the apparent density of residual bubbles in the container and the value of liquid surface fluctuation. The material interference parameters include the light transmittance of the material and the concentration of the material. A coupling processing module, which is used to determine the dynamic interference enhancement coefficient based on the dynamic interference parameters and The medium attenuation confusion coefficient determined by the material interference parameters is combined with the identification confusion characterization value to calculate the defect coupling feature value, so as to distinguish the defect anomaly tendency of the detection image. The detection and recognition module is used to determine whether to mark defects based on whether preset conditions of the detected image are met.

[0008] Further, the anomaly analysis module determines the intrinsic influence coefficient, including, The refractive factor is used to determine the ratio of the refractive index of the container wall to that of a reference container wall. The wall thickness factor is used to determine the ratio of the container wall thickness to the reference container wall thickness. The square root of the product of the refractive factor and the wall thickness factor is used to determine the intrinsic influence coefficient.

[0009] Furthermore, the anomaly analysis module determines the environmental impact coefficient, including, The projection factor is used to determine the angular deviation between the projection angle of the light source and the projection angle of the reference light source, and the ratio of the angular deviation to the reference deviation is determined as the projection factor. Used to determine the average brightness value of a preset dark area and the average brightness value of a preset bright area in the detected image; The ratio of the average brightness value of the preset dark area to the average brightness value of the preset bright area is used to determine the interference intensity. The ratio of the interference intensity to the reference interference intensity is used to determine the optical factor; The square root of the product of the projection factor and the light factor is used to determine the environmental impact coefficient.

[0010] Furthermore, the anomaly analysis module calculates and identifies confusion representation values, including: The intrinsic influence factor is used to determine the ratio of the intrinsic influence coefficient to the benchmark intrinsic influence coefficient. The ratio of the environmental impact coefficient to the benchmark environmental impact coefficient is used to determine the environmental impact factor; The weighted sum of the intrinsic impact factor and the environmental impact factor is used to determine the confusion characterization value.

[0011] Furthermore, the anomaly analysis module determines the defect recognition anomaly tendency of the detected image, wherein, If the identification confusion characterization value is greater than the identification confusion characterization value threshold, then the defect identification abnormality tendency of the detected image is determined to be a strong abnormality tendency. If the identification confusion characterization value is less than or equal to the identification confusion characterization value threshold, then the defect identification anomaly tendency of the detected image is determined to be a weak anomaly tendency.

[0012] Furthermore, the coupling processing module determines the dynamic interference enhancement coefficient, including, The bubble factor is used to determine the ratio of the apparent density of residual bubbles in the container to the reference apparent density. Used to construct a fluctuation time history curve of liquid level fluctuation values ​​relative to sampling time, and to determine the root mean square value of liquid level fluctuation values ​​in the fluctuation time history curve; The fluctuation factor is used to determine the ratio of the root mean square value to the benchmark root mean square value. The square root of the product of the bubble factor and the fluctuation factor is used to determine the dynamic disturbance enhancement coefficient.

[0013] Further, the coupling processing module determines the dielectric attenuation confusion coefficient, including, The concentration factor is used to determine the ratio of the concentration of the substance to the concentration of the reference substance. Used to obtain the transmittance values ​​of the material in each preset sub-region of the detection field of view; The ratio of the standard deviation to the mean of each transmittance value is used to determine the transmittance uniformity value. The ratio of the transmittance uniformity value to the reference transmittance uniformity value is used to determine the uniformity deviation factor; The square root of the product of the concentration factor and the uniformity deviation factor is used to determine the medium attenuation confusion coefficient.

[0014] Further, the coupling processing module calculates defect coupling characteristic values, including, The ratio of the dynamic interference enhancement coefficient to the reference dynamic interference enhancement coefficient is used to determine the dynamic influence factor; The ratio of the dielectric attenuation confusion coefficient to the reference dielectric attenuation confusion coefficient is used to determine the dielectric influence factor. The ratio of the identification confusion characterization value to the benchmark identification confusion characterization value is used to determine the confusion influence factor; The weighted sum of the dynamic influence factor, the media influence factor, and the confusion influence factor is used to determine the defect coupling characteristic value.

[0015] Furthermore, the coupling processing module distinguishes the defect and anomaly tendencies of the detected image, wherein, If the defect coupling feature value is greater than the defect coupling feature value threshold, then the defect anomaly tendency of the detected image is distinguished as an interference defect; If the defect coupling feature value is less than or equal to the defect coupling feature value threshold, then the defect abnormality tendency of the detected image is distinguished as a normal defect.

[0016] Furthermore, the detection and identification module determines whether to mark defects, wherein, If the detected image meets the preset conditions, then it is determined to mark the defect; If the detected image does not meet the preset conditions, then it is determined that defect marking will not be performed; The preset conditions are that the defect coupling feature value of the detected image is less than or equal to the defect coupling feature value threshold, and the recognition confusion characterization value of the detected image is less than or equal to the recognition confusion characterization value threshold.

[0017] Compared with existing technologies, this invention sets up a data acquisition module, an anomaly analysis module, a dynamic introduction module, a coupling processing module, and a detection and recognition module. These modules are used to calculate the recognition confusion characterization value based on the intrinsic influence coefficient determined by the anomaly analysis module using intrinsic parameters and the environmental influence coefficient determined by environmental parameters. This determines the anomaly tendency of the detected image in defect recognition. For detected images with a weak anomaly tendency, dynamic interference parameters and material interference parameters are introduced to determine the dynamic interference enhancement coefficient and the medium attenuation confusion coefficient. Combined with the recognition confusion characterization value, a defect coupling feature value is calculated to distinguish the anomaly tendency of the detected image and complete intelligent image defect recognition. This invention combines static and dynamic perspectives to analyze image defects caused by the container itself and the external environment, achieving intelligent defect analysis and improving recognition accuracy.

[0018] In particular, by analyzing the intrinsic parameters of the container corresponding to the detected image and the environmental parameters of the container, the intrinsic influence coefficient and environmental influence coefficient are calculated. From the perspectives of optical distortion of the container itself and degradation of external imaging conditions, the confusion tendency of image defect identification is analyzed. In reality, deviations in the refractive index of the container wall can cause unexpected deflection of the imaging optical path. At the same time, changes in the container wall thickness can introduce additional optical path difference. The combined effect of these two factors can distort the morphology, position, and contrast of defects in the image, causing blurring of the boundary between defect features and the normal background, thereby inducing the risk of misidentifying normal structures as defects. Furthermore, the shift in the projection angle of the light source will change the shadow direction and brightness distribution of the defect in the image, causing the same defect to present drastically different image features at different angles. The intrusion of background stray light directly increases the brightness of the dark areas of the image and reduces the overall contrast, causing minor defects to be missed and some stray light spots to be mislabeled as real defects. Based on this, the present invention considers analyzing the optical confusion effect of the container's intrinsic properties and environmental conditions on the detection image, providing a data basis for subsequent calculation of confusion characterization values, so as to further determine the abnormal tendency of defects and perform defect marking, thereby improving the efficiency and accuracy of defect identification.

[0019] In particular, by comprehensively calculating the identification confusion characterization value using both intrinsic influence coefficient and environmental influence coefficient, the abnormal tendency of defect identification in the detected image is determined. In practice, relying solely on either the intrinsic influence coefficient or the environmental influence coefficient may lead to inaccurate prediction of defect identification risk. For example, when there are significant deviations in the refractive index or wall thickness of the container wall, but images are acquired under ideal light source angles and extremely low stray light conditions, relying solely on the environmental influence coefficient may mistakenly lead to the assumption that the current image quality is good and there is no risk of confusion. In reality, the intrinsic distortion of the container has already caused considerable deformation and shift of the defect features in the image. Conversely, when the light source projection angle deviates significantly from the reference and the background stray light is strong, but the container itself is a standard part with uniform wall thickness and standard refractive index, relying solely on the intrinsic influence coefficient may also lead to the assumption that there is no confusion. In reality, the poor lighting conditions have produced shadows, light spots, or contrast anomalies that are highly similar to real defects. Based on this, the present invention considers calculating the identification confusion characterization value using both the intrinsic influence coefficient and the environmental influence coefficient to clarify the superimposed effect when both deteriorate together. Simultaneously, it provides data and theoretical basis for subsequent classification analysis, thereby improving the efficiency and accuracy of defect identification.

[0020] In particular, for detection images exhibiting a tendency towards weak defect identification anomalies, dynamic interference parameters and material interference parameters are introduced. The aggravating effects of time-varying disturbances and medium optical degradation on existing static confusion tendencies are analyzed from two perspectives, providing a data foundation for subsequent calculation of defect coupling feature values. In reality, even when the intrinsic characteristics of the container and imaging environment are ideal, resulting in a low-risk identification confusion characterization value, the continuous fluctuations of the liquid surface and the random tumbling of residual bubbles during production line operation can still introduce transient bright and dark stripes, local distortions, or motion blur in the image. These time-varying artifacts are easily misjudged as cracks or foreign objects. Furthermore, the uneven distribution of material transmittance and regional differences in material concentration can create mottled patterns in the image. The brightness gradient or blurred boundary of the image is highly similar to the morphological characteristics of real defects. Furthermore, the dynamic and media interference mentioned above are not independent effects. When the detection image has static confusion due to the refractive index deviation of the container wall or the angle shift of the light source, the liquid surface fluctuation will amplify the random deflection of light, and the bubbles will produce secondary refraction in the already distorted background. The uneven transmittance area will cause the small defects to be completely missed due to the reduced signal-to-noise ratio. Based on this, the present invention introduces the above parameters for images with weak anomaly tendency to compensate for the coverage blind spot of static confusion assessment, expands the acquisition range of dynamic and media parameters from images with strong anomaly tendency to cover images with weak anomaly tendency, performs targeted defect analysis, and improves the efficiency and accuracy of defect identification. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the intelligent image defect recognition system for industrial inspection, as described in an embodiment of the invention. Figure 2 This is a logic block diagram for determining the abnormal tendency of defect recognition in the detected image according to an embodiment of the invention; Figure 3 A logic block diagram for distinguishing the defect and abnormal tendencies of the detected image in an embodiment of the invention; Figure 4 This is a logic block diagram illustrating whether to perform defect marking in an embodiment of the invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Please see Figure 1 The diagram shown is a structural schematic of an intelligent image defect recognition system for industrial inspection according to an embodiment of the invention. The intelligent image defect recognition system for industrial inspection according to an embodiment of the invention includes: The data acquisition module is used to acquire the intrinsic parameters of the container corresponding to the detection image, including the refractive index of the container wall and the container wall thickness, and to acquire the environmental parameters of the container, including the light source projection angle and background stray light. An anomaly analysis module is used to calculate the identification confusion characterization value based on the intrinsic influence coefficient determined by the intrinsic parameters and the environmental influence coefficient determined by the environmental parameters, and to determine the defect identification anomaly tendency of the detected image. The dynamic input module is used to collect dynamic interference parameters and material interference parameters for detection images that show abnormal tendencies in weak defect identification. The dynamic interference parameters include the apparent density of residual bubbles in the container and the value of liquid surface fluctuation. The material interference parameters include the light transmittance of the material and the concentration of the material. A coupling processing module, which is used to determine the dynamic interference enhancement coefficient based on the dynamic interference parameters and The medium attenuation confusion coefficient determined by the material interference parameters is combined with the identification confusion characterization value to calculate the defect coupling feature value, so as to distinguish the defect anomaly tendency of the detection image. The detection and recognition module is used to determine whether to mark defects based on whether preset conditions of the detected image are met.

[0026] Specifically, there are no restrictions on the method of obtaining intrinsic parameters. For example, in practice, the refractive index of the container wall can be measured by an ellipsometer, and the container wall thickness can be obtained by a laser displacement sensor. Of course, those skilled in the art can also adopt other methods, as long as the required data is obtained. This will not be elaborated further.

[0027] Specifically, there are no restrictions on the method of obtaining environmental parameters. For example, in implementation, the projection angle can be recorded in real time by an angle encoder installed on the light source shaft, and the average gray value of the background image under no-light conditions can be used as the quantification value of background stray light. Of course, those skilled in the art can also adopt other methods, as long as the required data is obtained, which will not be elaborated here.

[0028] Specifically, there are no restrictions on the method of acquiring dynamic interference parameters. For example, in practice, the pixel ratio of residual bubbles in the container can be counted by the continuous frame image difference method to calculate the apparent density, and the liquid surface fluctuation value can be extracted by the laser beam deflection method. Of course, those skilled in the art can also adopt other methods, as long as the required data is obtained, which will not be elaborated here.

[0029] Specifically, there are no restrictions on the method of obtaining the material interference parameters. For example, in practice, the ratio of incident light intensity to transmitted light intensity can be determined by using a photodiode sensor to measure the incident light intensity and transmitted light intensity, which is the material transmittance. The concentration of the medium in the container can be measured by using an online density refractometer. Of course, those skilled in the art can also use other methods, as long as the required data is obtained, which will not be elaborated here.

[0030] It is understandable that several transparent or semi-transparent containers, such as medical glass bottles, plastic beverage bottles, or cosmetic ampoules, exist on the production line. During high-speed transport, the detection system needs to identify defects such as scratches, bubbles, and foreign objects inside and outside the bottle in real time. However, factors such as differences in the refractive index and thickness of the container wall, changes in the projection angle of the light source, background stray light interference, as well as residual bubbles inside the bottle, liquid surface fluctuations, and fluctuations in liquid transmittance and concentration can all lead to false or missed defect detections. Based on this, this invention introduces intrinsic parameters and environmental parameters to calculate and identify confusion characterization values, and combines dynamic and material interference parameters to generate defect coupling feature values, thereby effectively distinguishing between real and false defects and ensuring the accuracy of detection. Specifically, the anomaly analysis module determines the intrinsic influence coefficient, including, The refractive factor is used to determine the ratio of the refractive index of the container wall to that of a reference container wall. The wall thickness factor is used to determine the ratio of the container wall thickness to the reference container wall thickness. The square root of the product of the refractive factor and the wall thickness factor is used to determine the intrinsic influence coefficient.

[0031] Specifically, the refractive index of the reference container wall is calculated in advance. The refractive indices of the container walls of several qualified containers are obtained in advance, and the average of the refractive indices of each container wall is determined as the refractive index of the reference container wall.

[0032] Specifically, the baseline container wall thickness is calculated in advance. The wall thickness of several qualified containers is obtained in advance, and the average of the wall thicknesses of each container is determined as the baseline container wall thickness.

[0033] Specifically, the anomaly analysis module determines the environmental impact coefficient, including: The projection factor is used to determine the angular deviation between the projection angle of the light source and the projection angle of the reference light source, and the ratio of the angular deviation to the reference deviation is determined as the projection factor. Used to determine the average brightness value of a preset dark area and the average brightness value of a preset bright area in the detected image; The ratio of the average brightness value of the preset dark area to the average brightness value of the preset bright area is used to determine the interference intensity. The ratio of the interference intensity to the reference interference intensity is used to determine the optical factor; The square root of the product of the projection factor and the light factor is used to determine the environmental impact coefficient.

[0034] Specifically, the projection angle of the reference light source is calculated in advance. The angle sensor is obtained in advance when the light source is perpendicularly incident under several calibrated working conditions, and the average value of each reading of the angle sensor is determined to be the projection angle of the reference light source.

[0035] Specifically, the preset dark area is the region in the detection image that is expected to be not directly illuminated by a light source, such as the shadow area of ​​a container or the edge of a non-illuminated side in the implementation.

[0036] Specifically, in implementation, the preset dark area is the bright area in the detection image that is expected to be directly illuminated by the light source, for example, the area directly opposite the light source in implementation.

[0037] Specifically, the baseline interference intensity is calculated in advance. The historical interference intensity under several correctly identified defect conditions is obtained in advance, and the average value of each historical interference intensity is determined as the baseline interference intensity.

[0038] Specifically, by analyzing the intrinsic parameters of the container corresponding to the detected image and the environmental parameters of the container, the intrinsic influence coefficient and environmental influence coefficient are calculated. From the perspectives of optical distortion of the container itself and degradation of external imaging conditions, the confusion tendency in image defect identification is analyzed. In reality, deviations in the refractive index of the container wall can cause unexpected deflections in the imaging optical path. Simultaneously, changes in the container wall thickness introduce additional optical path differences. The combined effect of these two factors distorts the morphology, location, and contrast of defects in the image, blurring the boundary between defect features and the normal background. This induces the misidentification of normal structures as defects. Furthermore, the shift in the projection angle of the light source can alter the shadow direction and brightness distribution of defects in the image, causing the same defect to exhibit drastically different image features at different angles. The intrusion of background stray light directly increases the brightness of dark areas in the image and reduces the overall contrast, leading to the omission of minute defects and the mislabeling of some stray light spots as real defects. Based on this, the present invention considers analyzing the optical confusion effect exerted on the detection image by the intrinsic properties of the container and environmental conditions, providing a data basis for subsequent calculation of confusion characterization values, so as to further determine the abnormal tendency of defects and perform defect labeling, thereby improving the efficiency and accuracy of defect identification.

[0039] Specifically, the anomaly analysis module calculates and identifies confusion representation values, including: The intrinsic influence factor is used to determine the ratio of the intrinsic influence coefficient to the benchmark intrinsic influence coefficient. The ratio of the environmental impact coefficient to the benchmark environmental impact coefficient is used to determine the environmental impact factor; The weighted sum of the intrinsic impact factor and the environmental impact factor is used to determine the confusion characterization value.

[0040] Specifically, the baseline intrinsic influence coefficient is calculated in advance. The historical intrinsic influence coefficients under several correctly identified defect conditions are obtained in advance, and the average of each historical intrinsic influence coefficient is determined as the baseline intrinsic influence coefficient.

[0041] Specifically, the baseline environmental impact coefficient is calculated in advance. The historical environmental impact coefficients under several correctly identified defect conditions are obtained in advance, and the average of each historical environmental impact coefficient is determined as the baseline environmental impact coefficient.

[0042] Specifically, the sum of the weight coefficients of the intrinsic influence factor and the environmental influence factor is 1. When configuring the weights, considering that the influence of intrinsic parameters on imaging distortion and optical path deflection and the influence of environmental parameters on image contrast and background noise are similar in degree to those of typical detection scenarios, and that both can individually lead to confusion in defect identification, the weight coefficients of the intrinsic influence factor and the environmental influence factor are both determined to be 0.5.

[0043] Specifically, the anomaly analysis module determines the abnormal tendency of defect recognition in the detected image, wherein, If the identification confusion characterization value is greater than the identification confusion characterization value threshold, then the defect identification abnormality tendency of the detected image is determined to be a strong abnormality tendency. If the identification confusion characterization value is less than or equal to the identification confusion characterization value threshold, then the defect identification anomaly tendency of the detected image is determined to be a weak anomaly tendency.

[0044] Specifically, the confusion characterization threshold represents a boundary where there is a tendency to identify abnormalities. It is calculated in advance by obtaining historical confusion characterization values ​​under several correct identification defect conditions. The product of the mean of each historical confusion characterization value and the confusion accuracy is determined as the confusion characterization threshold. The confusion accuracy is selected in the interval [0,1]. In practice, in order to reduce the false trigger rate, the confusion accuracy is determined to be 0.9.

[0045] Specifically, the identification confusion characterization value is calculated by combining the intrinsic influence coefficient and the environmental influence coefficient to determine the abnormal tendency of defect identification in the detected image. In reality, relying solely on either the intrinsic influence coefficient or the environmental influence coefficient may lead to inaccurate prediction of defect identification risk. For example, when there are significant deviations in the refractive index or wall thickness of the container wall, but images are acquired under ideal light source angles and extremely low stray light conditions, relying solely on the environmental influence coefficient may mistakenly lead to the assumption that the current image quality is good and there is no risk of confusion. In reality, the intrinsic distortion of the container has already caused considerable deformation and shift of the defect features in the image. Conversely, when the light source projection angle deviates significantly from the reference and the background stray light is strong, but the container itself is a standard part with uniform wall thickness and standard refractive index, relying solely on the intrinsic influence coefficient may also lead to the assumption that there is no confusion. In reality, the poor lighting conditions have produced shadows, light spots, or contrast anomalies that are highly similar to real defects. Based on this, this invention considers calculating the identification confusion characterization value by combining the intrinsic influence coefficient and the environmental influence coefficient to clarify the superimposed effect when both deteriorate together. Simultaneously, it provides data and theoretical basis for subsequent classification analysis, thereby improving the efficiency and accuracy of defect identification.

[0046] Specifically, the coupling processing module determines the dynamic interference enhancement coefficient, including, The bubble factor is used to determine the ratio of the apparent density of residual bubbles in the container to the reference apparent density. Used to construct a fluctuation time history curve of liquid level fluctuation values ​​relative to sampling time, and to determine the root mean square value of liquid level fluctuation values ​​in the fluctuation time history curve; The fluctuation factor is used to determine the ratio of the root mean square value to the benchmark root mean square value. The square root of the product of the bubble factor and the fluctuation factor is used to determine the dynamic disturbance enhancement coefficient.

[0047] Specifically, the baseline apparent density is calculated in advance. The historical apparent density under several correctly identified defect conditions is obtained in advance, and the average of each historical apparent density is determined as the baseline apparent density.

[0048] Specifically, there is no limit to the exact duration of the sampling time. For example, it can be determined to be 1 second in practice. Of course, those skilled in the art can also determine it according to the actual situation, as long as it is reasonable. This will not be elaborated further.

[0049] Specifically, the baseline root mean square value is calculated in advance. The historical root mean square values ​​of the fluctuation time history curves under several correctly identified defect conditions are obtained in advance, and the average of each historical root mean square value is determined as the baseline root mean square value.

[0050] Specifically, the coupling processing module determines the media attenuation confusion coefficient, including, The concentration factor is used to determine the ratio of the concentration of the substance to the concentration of the reference substance. Used to obtain the transmittance values ​​of the material in each preset sub-region of the detection field of view; The ratio of the standard deviation to the mean of each transmittance value is used to determine the transmittance uniformity value. The ratio of the transmittance uniformity value to the reference transmittance uniformity value is used to determine the uniformity deviation factor; The square root of the product of the concentration factor and the uniformity deviation factor is used to determine the medium attenuation confusion coefficient.

[0051] Specifically, the reference substance concentration is calculated in advance. Several historical substance concentrations of the test medium at the standard preparation concentration, such as the normal concentration required by the target process, are obtained in advance, and the average of each historical substance concentration is determined as the reference substance concentration.

[0052] Specifically, the detection field of view is the imaging area of ​​the container covered by the camera in a single acquisition. In practice, it usually corresponds to the main detection range of the container body or bottom.

[0053] Specifically, there are no restrictions on the division of the preset sub-regions. For example, in implementation, they can be divided into equal parts. Of course, those skilled in the art can also determine the division based on the actual situation, as long as it is reasonable. This will not be elaborated further.

[0054] Specifically, the baseline transmittance is calculated in advance. The historical transmittance of the test medium at the standard concentration is obtained in advance, and the average of each historical transmittance is determined as the baseline transmittance.

[0055] Specifically, the coupling processing module calculates defect coupling characteristic values, including, The ratio of the dynamic interference enhancement coefficient to the reference dynamic interference enhancement coefficient is used to determine the dynamic influence factor; The ratio of the dielectric attenuation confusion coefficient to the reference dielectric attenuation confusion coefficient is used to determine the dielectric influence factor. The ratio of the identification confusion characterization value to the benchmark identification confusion characterization value is used to determine the confusion influence factor; The weighted sum of the dynamic influence factor, the media influence factor, and the confusion influence factor is used to determine the defect coupling characteristic value.

[0056] Specifically, the baseline dynamic interference enhancement coefficient is calculated in advance. The historical dynamic interference enhancement coefficients under several correctly identified defect conditions are obtained in advance, and the average of each historical dynamic interference enhancement coefficient is determined as the baseline dynamic interference enhancement coefficient.

[0057] Specifically, the reference medium attenuation confusion coefficient is calculated in advance. Several historical medium attenuation confusion coefficients under the condition of correctly identifying defects are obtained in advance, and the mean of each historical medium attenuation confusion coefficient is determined as the reference medium attenuation confusion coefficient.

[0058] Specifically, the identification confusion characterization value corresponding to the benchmark intrinsic influence coefficient and the benchmark environmental influence coefficient is the benchmark identification confusion characterization value.

[0059] Specifically, the sum of the weight coefficients of the dynamic impact factor, the media impact factor, and the confusion impact factor is 1. When configuring the weights, it is considered that the identification confusion characterization value comprehensively analyzes the main confusion contribution and is the core indicator for judging the abnormal tendency of defects. Dynamic interference and media interference are auxiliary correction factors. Therefore, the weight coefficients of the dynamic impact factor and the media impact factor are both determined to be 0.3, and the weight coefficient of the confusion impact factor is 0.4.

[0060] Specifically, the coupling processing module distinguishes the defect and abnormal tendencies of the detected image, wherein, If the defect coupling feature value is greater than the defect coupling feature value threshold, then the defect anomaly tendency of the detected image is distinguished as an interference defect; If the defect coupling feature value is less than or equal to the defect coupling feature value threshold, then the defect abnormality tendency of the detected image is distinguished as a normal defect.

[0061] Specifically, the defect coupling feature value threshold represents a boundary between interfering defects and normal defects. It is calculated in advance by obtaining historical defect coupling feature values ​​under several correctly identified defect conditions. The product of the mean of each historical defect coupling feature value and the defect accuracy is determined as the defect coupling feature value threshold. The defect accuracy is determined in the interval [0,1]. In practice, in order to improve the robustness of distinguishing between real defects and pseudo defects, the defect accuracy is determined to be 0.9.

[0062] Specifically, for detection images exhibiting a tendency towards weak defect identification anomalies, dynamic interference parameters and material interference parameters are introduced. The aggravating effects of time-varying disturbances and medium optical degradation on existing static confusion tendencies are analyzed from two perspectives, providing a data foundation for subsequent calculation of defect coupling feature values. In reality, even when the intrinsic characteristics of the container and imaging environment are ideal, resulting in a low-risk identification confusion characterization value, continuous fluctuations in the liquid surface and random tumbling of residual bubbles during production line operation can still introduce transient bright and dark stripes, local distortions, or motion blur in the image. These time-varying artifacts are easily misjudged as cracks or foreign objects. Furthermore, the uneven distribution of material transmittance and regional differences in material concentration can form spots in the image. The brightness gradient or blurred boundary of the defect is highly similar to the morphological characteristics of the real defect. Furthermore, the above-mentioned dynamic and medium interferences do not act independently. When the detection image has static confusion due to the refractive index deviation of the container wall or the angle shift of the light source, the liquid surface fluctuation will amplify the random deflection of light, the bubbles will produce secondary refraction in the already distorted background, and the uneven transmittance area will cause the small defects to be completely missed due to the reduced signal-to-noise ratio. Based on this, the present invention introduces the above parameters for images with weak anomaly tendency to compensate for the coverage blind spot of static confusion assessment, expands the collection range of dynamic and medium parameters from images with strong anomaly tendency to cover images with weak anomaly tendency to perform targeted defect analysis, and improves the efficiency and accuracy of defect identification.

[0063] Specifically, the detection and identification module determines whether to mark defects, whereby... If the detected image meets the preset conditions, then it is determined to mark the defect; If the detected image does not meet the preset conditions, then it is determined that defect marking will not be performed; The preset conditions are that the defect coupling feature value of the detected image is less than or equal to the defect coupling feature value threshold, and the recognition confusion characterization value of the detected image is less than or equal to the recognition confusion characterization value threshold.

[0064] Specifically, there are no restrictions on the method of defect marking. For example, in implementation, it can be the output of defect coordinates, or any other marking method, as long as the desired effect is achieved. This will not be elaborated further.

[0065] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent image defect recognition system for industrial inspection, characterized in that, include: The data acquisition module is used to acquire the intrinsic parameters of the container corresponding to the detection image, including the refractive index of the container wall and the container wall thickness, and to acquire the environmental parameters of the container, including the light source projection angle and background stray light. An anomaly analysis module is used to calculate the identification confusion characterization value based on the intrinsic influence coefficient determined by the intrinsic parameters and the environmental influence coefficient determined by the environmental parameters, and to determine the defect identification anomaly tendency of the detected image. The dynamic input module is used to collect dynamic interference parameters and material interference parameters for detection images that show abnormal tendencies in weak defect identification. The dynamic interference parameters include the apparent density of residual bubbles in the container and the value of liquid surface fluctuation. The material interference parameters include the light transmittance of the material and the concentration of the material. The coupling processing module is used to determine the dynamic interference enhancement coefficient based on the dynamic interference parameters and The medium attenuation confusion coefficient determined by the material interference parameters is combined with the identification confusion characterization value to calculate the defect coupling feature value, so as to distinguish the defect anomaly tendency of the detection image. The detection and recognition module is used to determine whether to mark defects based on whether preset conditions of the detected image are met.

2. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The anomaly analysis module determines the intrinsic influence coefficient, including: The refractive factor is used to determine the ratio of the refractive index of the container wall to that of a reference container wall. The wall thickness factor is used to determine the ratio of the container wall thickness to the reference container wall thickness. The square root of the product of the refractive factor and the wall thickness factor is used to determine the intrinsic influence coefficient.

3. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The anomaly analysis module determines the environmental impact coefficient, including: The projection factor is used to determine the angular deviation between the projection angle of the light source and the projection angle of the reference light source, and the ratio of the angular deviation to the reference deviation is determined as the projection factor. Used to determine the average brightness value of a preset dark area and the average brightness value of a preset bright area in the detected image; The ratio of the average brightness value of the preset dark area to the average brightness value of the preset bright area is used to determine the interference intensity. The ratio of the interference intensity to the reference interference intensity is used to determine the optical factor; The square root of the product of the projection factor and the light factor is used to determine the environmental impact coefficient.

4. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The anomaly analysis module calculates and identifies confusion characterization values, including: The intrinsic influence factor is used to determine the ratio of the intrinsic influence coefficient to the benchmark intrinsic influence coefficient. The ratio of the environmental impact coefficient to the benchmark environmental impact coefficient is used to determine the environmental impact factor; The weighted sum of the intrinsic impact factor and the environmental impact factor is used to determine the confusion characterization value.

5. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The anomaly analysis module determines the defect recognition anomaly tendency of the detected image, wherein, If the identification confusion characterization value is greater than the identification confusion characterization value threshold, then the defect identification abnormality tendency of the detected image is determined to be a strong abnormality tendency. If the identification confusion characterization value is less than or equal to the identification confusion characterization value threshold, then the defect identification anomaly tendency of the detected image is determined to be a weak anomaly tendency.

6. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The coupling processing module determines the dynamic interference enhancement coefficient, including: The bubble factor is used to determine the ratio of the apparent density of residual bubbles in the container to the reference apparent density. Used to construct a fluctuation time history curve of liquid level fluctuation values ​​relative to sampling time, and to determine the root mean square value of liquid level fluctuation values ​​in the fluctuation time history curve; The fluctuation factor is used to determine the ratio of the root mean square value to the benchmark root mean square value. The square root of the product of the bubble factor and the fluctuation factor is used to determine the dynamic disturbance enhancement coefficient.

7. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The coupling processing module determines the medium attenuation confusion coefficient, including, The concentration factor is used to determine the ratio of the concentration of the substance to the concentration of the reference substance. Used to obtain the transmittance values ​​of the material in each preset sub-region of the detection field of view; The ratio of the standard deviation to the mean of each transmittance value is used to determine the transmittance uniformity value. The ratio of the transmittance uniformity value to the reference transmittance uniformity value is used to determine the uniformity deviation factor; The square root of the product of the concentration factor and the uniformity deviation factor is used to determine the medium attenuation confusion coefficient.

8. The intelligent image defect recognition system for industrial inspection according to claim 1, characterized in that, The coupling processing module calculates the defect coupling characteristic values, including: The ratio of the dynamic interference enhancement coefficient to the reference dynamic interference enhancement coefficient is used to determine the dynamic influence factor; The ratio of the dielectric attenuation confusion coefficient to the reference dielectric attenuation confusion coefficient is used to determine the dielectric influence factor. The ratio of the identification confusion characterization value to the benchmark identification confusion characterization value is used to determine the confusion influence factor; The weighted sum of the dynamic influence factor, the media influence factor, and the confusion influence factor is used to determine the defect coupling characteristic value.

9. The intelligent image defect recognition system for industrial inspection according to claim 5, characterized in that, The coupling processing module distinguishes the defect and abnormal tendencies of the detected image, wherein, If the defect coupling feature value is greater than the defect coupling feature value threshold, then the defect anomaly tendency of the detected image is distinguished as an interference defect; If the defect coupling feature value is less than or equal to the defect coupling feature value threshold, then the defect abnormality tendency of the detected image is distinguished as a normal defect.

10. The intelligent image defect recognition system for industrial inspection according to claim 9, characterized in that, The detection and identification module determines whether to mark defects, wherein... If the detected image meets the preset conditions, then it is determined to mark the defect; If the detected image does not meet the preset conditions, then it is determined that defect marking will not be performed; The preset condition is that the defect coupling feature value of the detected image is less than or equal to the defect coupling feature value threshold, and the recognition confusion characterization value of the detected image is less than or equal to the recognition confusion characterization value threshold.

Images