Picture defect test apparatus control system and method

Abstract

The application relates to the technical field of visual control, in particular to a picture defect test equipment control system and method, and the system comprises a difference tensor construction module, a space-time constraint construction module, a strain response judgment module, a coupling focusing mapping module and a boundary stability evaluation module.In the application, a multi-dimensional difference correlation relationship is constructed by extracting pixel gray difference, brightness gradient and chrominance phase angle, and subtle brightness fluctuation and color deviation in a picture can be captured.The superposition of space-time change rates can distinguish the dynamic performance of picture defects, and can effectively identify persistent and transient abnormalities.Combining tensor strain ratio and difference amplitude analysis, abnormal areas can be adaptively identified, and the neglect of traditional methods for complex scenes is avoided.The weighted aggregation of local energy response improves the stability of the defect boundary, ensures the accurate positioning and consistency of the picture defect, and significantly improves the reliability and accuracy of the detection result, especially under dynamic changes and complex lighting environments.

Description

Technical Field

[0001] This invention relates to the field of visual control technology, and in particular to a control system and method for image defect testing equipment. Background Technology

[0002] The field of visual control technology is an important branch of automated and intelligent inspection systems. It primarily involves the detection, identification, and control of target objects using optical imaging, image acquisition, image recognition, and control execution. Its core aspects include visual information acquisition, imaging signal conversion, image feature analysis, and collaborative feedback mechanisms with the control system. This technology utilizes industrial cameras, light sources, and processing units to detect and determine the surface, shape, or motion state of the object being measured. This enables automatic identification and judgment of visual information during manufacturing, assembly, and quality inspection processes, forming a closed-loop control system with the control device.

[0003] The traditional image defect testing equipment control system refers to a system used to detect defects in the output image of a display screen, image projection surface, or imaging device, and to control the testing equipment. This system primarily targets and tests for visual defects in the image, such as bright spots, dark spots, color casts, and broken lines. The traditional image defect testing equipment control system uses a fixed camera unit to capture the displayed image frame by frame. The optical signals are then converted into electrical signals by an image sampling circuit. Defects in the target image area are determined using threshold comparison or pixel difference methods. Simultaneously, control commands drive the testing platform to perform operations such as image switching or brightness adjustment, thereby achieving the detection and control of image defects.

[0004] Traditional image defect detection systems rely on fixed camera units to acquire images frame by frame and simple pixel difference judgment. This method cannot effectively handle complex lighting changes, color gradations, and dynamic textures in the image. Because static threshold settings cannot adapt to different display conditions, they are prone to false positives or false negatives, especially for abnormal or transient defects that appear within a short period. Image sampling and differential processing fail to effectively capture temporal changes, and lag in control response also leads to decreased defect localization accuracy. The lack of accurate judgment of abnormal region boundaries affects the stability and consistency of detection results. Summary of the Invention

[0005] To address the shortcomings of traditional image defect detection systems that rely on fixed camera units for frame-by-frame acquisition and simple pixel difference judgment, which are ineffective in handling complex lighting changes, color gradients, and dynamic textures in images, and the inability of static threshold settings to adapt to different display conditions, leading to misjudgments or missed detections, especially for abnormal or transient defects appearing within a short period, this invention provides a control system and method for image defect testing equipment. The technical solution is as follows:

[0006] On the one hand, a control system for image defect testing equipment is provided, which includes:

[0007] The difference tensor construction module performs multi-dimensional feature analysis on the input image, extracts pixel unit grayscale difference, brightness gradient vector and chromaticity phase angle, performs spatiotemporal correlation and tensor organization, constructs pixel difference tensor structure, and passes it to the spatiotemporal constraint construction module.

[0008] The spatiotemporal constraint construction module receives the pixel difference tensor structure, calculates the change rate of the time dimension tensor, superimposes the spatiotemporal change rate to form a multidimensional constraint matrix, compares the constraint offset with the preset response threshold, clusters the pixel positions that exceed the threshold, generates a set of constraint discontinuous regions, and transmits it to the strain response determination module.

[0009] The strain response determination module receives the set of constrained discontinuous regions, calculates the tensor dimension strain ratio and difference magnitude, inputs it into the hidden Markov model to identify persistent and transient abnormal regions, generates a set of abnormal response region types, and passes it to the coupling focus mapping module.

[0010] The coupling focus mapping module receives the abnormal response region type set, extracts the neighboring tensor blocks for gradient comparison, and when the change rate of the tensor coupling balance value exceeds the preset mutation threshold, the weighted local difference response value is injected into the energy coefficient of the center pixel to form the energy distribution map of the defect region, which is then transmitted to the boundary stability assessment module.

[0011] As a further embodiment of the present invention, the pixel difference tensor structure includes a grayscale difference feature domain, a brightness gradient distribution layer, and a chromaticity angle distribution map; the multidimensional constraint matrix includes a time rate of change component, a spatial clustering component, and a threshold limit group; the abnormal response region type set includes persistent abnormal region categories, transient abnormal region categories, and response intensity levels; and the defect region energy distribution map includes a high energy concentration region, an energy attenuation band, and an energy transition layer.

[0012] As a further aspect of the present invention, the difference tensor construction module includes:

[0013] The grayscale difference extraction submodule obtains the grayscale value of the pixel unit in the input image, calculates the grayscale difference amplitude between adjacent pixels according to the pixel sequence position, marks the grayscale change relationship according to the row and column index and removes abnormal pixel differences, and generates a pixel grayscale difference distribution set.

[0014] The brightness gradient calculation submodule calculates the brightness difference values ​​of adjacent pixels in the horizontal and vertical directions based on the pixel grayscale difference distribution set, superimposes them into a local brightness gradient vector, determines the brightness direction distribution according to the pixel coordinates, and obtains the pixel brightness gradient vector field.

[0015] The chroma phase integration submodule calculates the phase offset angle between chroma channels based on the pixel luminance gradient vector field, performs multi-channel coupling based on the luminance gradient direction and chroma phase angle, maps and reassembles the channel data, and obtains the pixel difference tensor structure.

[0016] As a further aspect of the present invention, the spatiotemporal constraint construction module includes:

[0017] The temporal deformation extraction submodule obtains the pixel difference tensor structure, parses the node coordinate values ​​and strain components, arranges the deformation data of multiple frames in sequence according to the time series identifier, calculates the displacement change rate of adjacent time frames, and generates a time dimension change rate set according to the time step.

[0018] The tensor change calculation submodule, based on the time dimension change rate set, calls the spatial component information of the pixel difference tensor structure, performs superposition calculation on the time change rate and spatial component of multiple dimensions, analyzes the comprehensive change rate value in multiple directions, and establishes a multidimensional constraint matrix.

[0019] The constraint clustering identification submodule calls the multidimensional constraint matrix to calculate the constraint offset corresponding to the pixel, compares the offset with the response threshold, filters out the pixel coordinates whose offset exceeds the threshold, and performs clustering based on the similarity features of the spatial neighborhood to generate a set of constraint discontinuous regions.

[0020] The response threshold is adaptively determined based on the statistical characteristics of the constraint offset of adjacent pixels.

[0021] As a further aspect of the present invention, the strain response determination module includes:

[0022] The strain data acquisition submodule acquires the node strain data of the constrained discontinuous region set, parses the node tensor dimension component values, pairs the strain components with the region index number, adjusts the components of adjacent regions synchronously according to the time series and corrects missing values, and generates a regional strain dataset.

[0023] The strain ratio calculation submodule extracts the principal and secondary direction components based on the regional strain dataset, calculates the principal-secondary ratio and compares it with the average ratio benchmark, filters out regions where the difference exceeds the benchmark, and generates a tensor dimension strain ratio difference set.

[0024] The Hidden Markov Recognition Submodule calls the tensor dimension strain ratio difference set, calculates the temporal difference transition probability to fit the observation sequence, determines the sequence change based on the Hidden Markov state transition characteristics, identifies persistent and transient offset regions, and generates an abnormal response region type set.

[0025] As a further aspect of the present invention, the coupling focusing mapping module includes:

[0026] The neighborhood data parsing submodule obtains the abnormal response region type set, parses the pixel response sequence, extracts the neighborhood tensor block according to the spatial index and establishes gradient reference data, normalizes according to the response weight and calculates the response difference between pixels, summarizes the coupling balance parameters, and generates the initial coupling balance value.

[0027] The neighborhood gradient comparison submodule calls the initial coupling balance value, calculates the gradient change rate for the pixel neighborhood tensor block, compares the pixel gradient change rate with the abrupt threshold, filters the set of pixels whose change rate exceeds the threshold and calculates the local difference response value to obtain the local difference response coefficient.

[0028] The energy distribution generation submodule performs weighted superposition calculation on pixel energy factors based on the local difference response coefficient, constructs energy transfer data and integrates accumulated energy, calculates the superposition intensity of energy intervals, and obtains the energy distribution map of the defect area.

[0029] As a further aspect of the present invention, the initial coupling balance value refers to the initial coupling amount of the pixel neighborhood calculated based on the pixel response difference and normalized weight in the neighborhood tensor block.

[0030] The local difference response coefficient refers to the quantitative parameter of difference calculated based on the difference in gradient change rate between the target pixel and its neighboring pixels, combined with the initial coupling equilibrium value.

[0031] The mutation threshold refers to a threshold parameter determined based on the statistical characteristics of the global gradient change rate.

[0032] As a further embodiment of the present invention, the boundary stability assessment module receives the energy distribution map of the defect region, identifies high-response areas through multi-layer threshold scanning, detects the continuity of spatial energy gradient direction and temporal stability, marks effective defect regions, uses Kalman filtering algorithm to suppress noise, and outputs defect feature integration results.

[0033] The integrated defect feature results include an effective defect boundary set, a stability evaluation parameter set, and a noise suppression coefficient set.

[0034] As a further aspect of the present invention, the boundary stability assessment module includes:

[0035] The energy data receiving submodule acquires the energy distribution map of the defect area, performs grayscale channel separation on the image pixels and extracts the energy amplitude, calculates and normalizes the energy difference between adjacent pixels, and scans the energy distribution layer by layer according to the multi-layer energy threshold interval and counts the response density to generate a multi-layer energy response dataset.

[0036] The energy gradient detection submodule, based on the multi-layer energy response dataset, calculates the pixel energy difference and quantizes the gradient direction, extracts the time series energy change and calculates the gradient direction offset, determines the range of the directional continuous region according to the gradient direction stability condition, and obtains the spatial energy continuity parameter set.

[0037] The noise suppression integration submodule calls the spatial energy continuity parameter set, performs sliding smoothing on the time-series energy points and calculates the prediction error, estimates the difference between the predicted and observed values ​​based on the Kalman filter update equation, adjusts the outlier weights to correct the aggregated energy data, and generates the defect feature integration result.

[0038] On the other hand, a screen defect testing equipment control method, which is executed based on the aforementioned screen defect testing equipment control system, includes the following steps:

[0039] S1: Perform multi-dimensional feature analysis on the input screen, extract pixel unit grayscale difference, brightness gradient vector and chromaticity phase angle, perform spatiotemporal correlation and tensor organization, and construct pixel difference tensor structure;

[0040] S2: Call the pixel difference tensor structure, calculate the change rate of the time dimension tensor, superimpose the spatiotemporal change rate to form a multidimensional constraint matrix, compare the constraint offset with the response threshold, cluster the positions of pixels exceeding the threshold, and generate a set of constraint discontinuous regions.

[0041] S3: Receive the set of constrained discontinuous regions, calculate the tensor dimension strain ratio and difference magnitude, input the hidden Markov model to identify persistent and transient abnormal regions, and generate a set of abnormal response region types.

[0042] S4: Call the abnormal response region type set, extract the neighborhood tensor block for gradient comparison, and when the change rate of the tensor coupling balance value exceeds the mutation threshold, inject the weighted local difference response value into the energy coefficient of the center pixel to form the energy distribution map of the defect region.

[0043] S5: Call the energy distribution map of the defect area, identify high response areas through multi-layer threshold scanning, detect the continuity of spatial energy gradient direction and temporal stability, mark the effective defect area, use the Kalman filter algorithm to suppress noise, and output the integrated result of defect features.

[0044] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0045] By comprehensively extracting pixel grayscale differences, brightness gradients, and chromaticity phase angles, and constructing multidimensional difference correlations, subtle brightness fluctuations and color shifts in the image can be accurately captured. The superposition of spatiotemporal change rates allows for the differentiation of dynamic appearance of image defects, effectively identifying persistent and transient anomalies. Combining tensor strain ratio and difference amplitude analysis enables adaptive identification of abnormal regions, avoiding the neglect of complex scenes by traditional methods. Weighted aggregation of local energy responses enhances the stability of defect boundaries, ensuring accurate localization and consistent judgment of image defects, significantly improving the reliability and accuracy of detection results, especially under dynamic and complex lighting conditions. Attached Figure Description

[0046] 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 the accompanying drawings without creative effort.

[0047] Figure 1 This is a system schematic diagram of the present invention;

[0048] Figure 2 This is a schematic diagram of the system framework of the present invention;

[0049] Figure 3 This is a flowchart of the difference tensor construction module in this invention;

[0050] Figure 4 This is a flowchart of the spatiotemporal constraint construction module in this invention;

[0051] Figure 5 This is a flowchart of the strain response determination module in this invention;

[0052] Figure 6 This is a flowchart of the coupling focus mapping module in this invention;

[0053] Figure 7 This is a flowchart of the boundary stability assessment module in this invention;

[0054] Figure 8 This is a flowchart of the method of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0056] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0057] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0058] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0059] 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.

[0060] This invention provides a control system for an image defect testing device, such as... Figure 1-2 The diagram shown illustrates the control system for the image defect testing equipment. This system includes:

[0061] The difference tensor construction module performs multi-dimensional feature analysis on the input image, extracts pixel unit grayscale difference, brightness gradient vector and chromaticity phase angle, performs spatiotemporal correlation and tensor organization, constructs pixel difference tensor structure, and passes it to the spatiotemporal constraint construction module.

[0062] The spatiotemporal constraint construction module receives the pixel difference tensor structure, calculates the change rate of the time dimension tensor, superimposes the spatiotemporal change rate to form a multidimensional constraint matrix, compares the constraint offset with the preset response threshold, clusters the positions of pixels exceeding the threshold, generates a set of constraint discontinuous regions, and passes it to the strain response determination module.

[0063] The strain response determination module receives a set of constrained discontinuous regions, calculates the tensor dimension strain ratio and difference magnitude, inputs it into a hidden Markov model to identify persistent and transient anomaly regions, generates a set of anomaly response region types, and passes it to the coupling focusing mapping module.

[0064] The coupling focus mapping module receives the abnormal response region type set, extracts the neighboring tensor blocks for gradient comparison, and when the change rate of the tensor coupling balance value exceeds the preset mutation threshold, the weighted local difference response value is injected into the energy coefficient of the center pixel to form the energy distribution map of the defect region, which is then transmitted to the boundary stability assessment module.

[0065] The boundary stability assessment module receives the energy distribution map of the defect region, identifies high-response areas through multi-layer threshold scanning, detects the continuity of spatial energy gradient direction and temporal stability, marks the effective defect region, uses the Kalman filter algorithm to suppress noise, and outputs the integrated defect feature result.

[0066] The pixel difference tensor structure includes a gray-level difference feature domain, a brightness gradient distribution layer, and a chromaticity angle distribution map. The multidimensional constraint matrix includes a time rate of change component, a spatial clustering component, and a threshold limit set. The abnormal response region type set includes persistent abnormal region categories, transient abnormal region categories, and response intensity levels. The defect region energy distribution map includes a high-energy concentration region, an energy attenuation band, and an energy transition layer. The defect feature integration result includes an effective defect boundary set, a stability evaluation parameter set, and a noise suppression coefficient set.

[0067] Specifically, such as Figure 2 , 3 As shown, the difference tensor construction module includes:

[0068] The grayscale difference extraction submodule obtains the grayscale value of the pixel unit in the input image, calculates the grayscale difference amplitude between adjacent pixels according to the pixel sequence position, marks the grayscale change relationship according to the row and column index and removes abnormal pixel differences, and generates a pixel grayscale difference distribution set.

[0069] Based on the input image captured from the surface of the conveyor belt of the feeding machine, extract one of them. The 8-bit grayscale values ​​of a pixel region, and the grayscale value matrix of that region, for example, is... ,in and Here are the row and column indices of the pixels, and the grayscale value range is... to Next, based on the pixel sequence from left to right and top to bottom, calculate the absolute value of the grayscale difference between each pixel and its right-hand neighbor and its bottom neighbor. For example, for the pixel in the second row and second column... Its value is Then calculate its relationship with the right pixel. (value is) The horizontal grayscale difference, i.e. Simultaneously calculate its relationship with the pixels below. (value is) The vertical grayscale difference, i.e. Then, based on the row and column indices of the pixels, the difference is labeled with its position, forming a set of change relationships containing positional information. For example, the horizontal grayscale difference... Marked on coordinates and Between, the vertical grayscale difference Marked on coordinates and In between, it is necessary to remove abnormal pixel differences caused by image noise or minor foreign objects (such as dust). The determination of whether something is abnormal is based on a preset grayscale difference threshold, which is set with reference to the... The statistical analysis results of the intact conveyor belt sample images were used to calculate the average and standard deviation of the gray-level differences between adjacent pixels in the sample. For example, the average value was... The standard deviation is The threshold can then be set as the average plus three times the standard deviation, i.e. Anything greater than The grayscale difference was judged as "high" and considered abnormal. In the example, the vertical grayscale difference... Much larger Therefore, this difference is removed without affecting the horizontal grayscale difference. Normal change values, by traversing the entire After applying this rejection rule to the region, the non-abnormal grayscale difference amplitude values ​​are finally integrated and processed to generate a pixel grayscale difference distribution set.

[0070] The brightness gradient calculation submodule calculates the brightness difference values ​​of adjacent pixels in the horizontal and vertical directions based on the pixel gray-level difference distribution set, and superimposes them into a local brightness gradient vector. The brightness direction distribution is determined according to the pixel coordinates to obtain the pixel brightness gradient vector field.

[0071] Based on the generated pixel grayscale difference distribution set, which no longer contains abrupt values ​​caused by noise or other factors, the brightness gradient is then calculated. This process first reclassifies the differences in the distribution set as the brightness difference values ​​in the horizontal and vertical directions for each pixel location. For any pixel... Its horizontal brightness difference value It is obtained by subtracting the gray value of the pixel from the gray value of its right-hand neighbor. For example, for a pixel... Its grayscale value (Assuming the outliers detected in the previous step have been smoothed), its right-hand pixel grayscale value Then the horizontal difference at that point for Similarly, the pixels below it grayscale value Then vertical difference for Then, these two differences are used to construct a local brightness gradient vector for each pixel. In other words, the vector is represented as This vector describes the magnitude and direction of the brightness change at that point. Next, this process is repeated for every pixel in the image, calculating the brightness gradient vector based on the pixel's coordinates. Mapping these coordinates to corresponding locations determines the directional distribution of brightness across the entire image, for example, at the pixel level. The gradient vector is And pixels The gradient vector is By combining the gradient vectors of the pixels, a pixel brightness gradient vector field covering the entire detection area is obtained.

[0072] The chroma phase integration submodule calculates the phase offset angle between chroma channels based on the pixel luminance gradient vector field, performs multi-channel coupling based on the luminance gradient direction and chroma phase angle, maps and reassembles the channel data, and obtains the pixel difference tensor structure.

[0073] Based on the generated pixel luminance gradient vector field, chromaticity information is further integrated. This process assumes the input image is a color image and has been converted to a color space containing luminance and chromaticity components, such as YCbCr. The operation is first performed on the two chromaticity channels of the image (e.g., Cb and Cr channels), calculating the phase shift angle between the chromaticity channels. Specifically, for each pixel location, its local gradient in the Cb and Cr channels is calculated separately, using the same method as the luminance gradient, resulting in the chromaticity gradient vector.

[0074] and ;

[0075] The phase offset angle is obtained by calculating the angle between these two chromaticity gradient vectors. For example, the Cb gradient of a pixel is... The Cr gradient is The phase shift angle between the two is Next, based on the gradient direction of the pixel in the luminance channel and the previously calculated chrominance phase angle, a multi-channel coupling operation is performed. This operation reconstructs the channel data through a weighted mapping. The weight coefficients are set with reference to the differences in color and texture between the material being tested (such as electronic components) and the conveyor belt background. If color difference is the key to defect detection, the weight of chrominance correlation will be higher. For example, the weight of the luminance gradient direction can be set. for chromaticity phase angle weight for ,and For a brightness gradient direction of The phase angles of the chromaticity and chromaticity are: The difference feature value of a pixel with a degree of 1 is calculated as follows: Finally, the original data of each pixel, the luminance gradient vector, the chromaticity phase angle, and the calculated coupling difference feature value are restructured into a multidimensional data array, namely the pixel difference tensor structure, as shown in Table 1 below.

[0076] Table 1: Example Table of Pixel Difference Tensor Structure

[0077]

[0078] As shown in Table 1, this table lists the tensor structures of some pixels after multi-channel data coupling and recombination, which include position, brightness gradient, chromaticity phase and final coupling feature value, providing multi-dimensional data support for subsequent defect judgment.

[0079] Specifically, such as Figure 2 , 4 As shown, the spatiotemporal constraint construction module includes:

[0080] The temporal deformation extraction submodule obtains the pixel difference tensor structure, parses the node coordinate values ​​and strain components, arranges the deformation data of multiple frames in sequence according to the time series identifier, calculates the displacement change rate of adjacent time frames, and generates a time dimension change rate set according to the time step.

[0081] Obtain the pixel difference tensor structure, specifically by parsing out the timestamp. and Two frames of data, with a frame interval of And extract the coordinate values ​​of each pixel node and its corresponding coupling difference value as strain components. For example, for a pixel... ,exist The strain component value at time t is ,exist At any given moment, due to minute vibrations or material movement that may occur on the conveyor belt, its strain components become... Subsequently, based on and These two time series identifiers sequentially arrange the strain component data of these two frames, and then calculate the values ​​in this time series. The displacement rate of this pixel is calculated as follows: The strain component value at time minus The strain component value at time step 1, divided by the time step, is... Units per millisecond; this calculation is applied to every pixel in the image, for example, pixels. exist The strain component at time t is ,exist Time becomes Then its displacement rate is Units per millisecond. After performing this calculation on each pixel, the resulting rate of change values ​​for each pixel are organized according to their original coordinates in the image. The time step generates a time dimension change rate set that contains the changes of all pixels in the entire screen during this time period.

[0082] The tensor change calculation submodule, based on the time dimension change rate set, calls the spatial component information of the pixel difference tensor structure, performs superposition calculation on the time change rate and spatial component of multiple dimensions, analyzes the comprehensive change rate value in multiple directions, and establishes a multi-dimensional constraint matrix.

[0083] Based on the generated temporal rate of change set, which records the rate of change of each pixel within a specific time step, and by invoking the spatial component information from the previously established pixel difference tensor structure, the spatial component specifically refers to the brightness gradient magnitude of each pixel. For example, pixel points are obtained from the temporal rate of change set. The rate of change is The value is calculated in units per millisecond, and the brightness gradient magnitude at that point is retrieved from the pixel difference tensor structure. Next, the temporal rate of change and spatial components are superimposed and calculated. This calculation is performed using a weighted summation method. The weights are set based on prior analysis of image features during normal conveyor belt operation. Under normal motion conditions, spatial texture changes (gradients) are inherent features, while drastic temporal changes are more likely to indicate anomalies. Therefore, a higher weight is assigned to the temporal rate of change. for Spatial gradient magnitude weight for The sum of the two is Then the pixel The comprehensive rate of change, also known as its multidimensional constraint coefficient, is calculated as follows: |Rate of change over time|+ ×|Spatial gradient magnitude| For neighboring pixels The rate of change over time is the same when the same operation is performed. The spatial gradient magnitude is Then its multidimensional constraint coefficient is By performing this superposition calculation on all pixels within the field of view, the comprehensive rate of change value of each point is analyzed and obtained, and finally a multi-dimensional constraint matrix covering the entire screen is established.

[0084] The constraint clustering recognition submodule calls the multidimensional constraint matrix to calculate the constraint offset corresponding to the pixel, compares the offset with the response threshold, filters out the pixel coordinates whose offset exceeds the threshold, and performs clustering based on the similarity features of the spatial neighborhood to generate a set of constraint discontinuous regions.

[0085] The system calls the established multidimensional constraint matrix, which contains the comprehensive change rate of each pixel. Then, it calculates the constraint offset corresponding to each pixel. This offset is the absolute value of the difference between the constraint coefficient of a pixel and the average constraint coefficient of its surrounding neighboring pixels. Taking the neighborhood window as an example, calculate the pixel points The constraint offset is first obtained by acquiring its own and surrounding values. The multidimensional constraint coefficients of each pixel are shown in Table 2. The calculation of these... The arithmetic mean of the coefficients, Then the center point The constraint offset is Next, this offset is compared with an adaptively determined response threshold, which is determined by statistically analyzing the data. The average and standard deviation of the constrained offset of pixels within the neighborhood, assuming the calculated average offset for this region is... The standard deviation is The response threshold is then set to the average value plus a multiple. ( Set as The standard deviation of ), i.e. Then, filter out the pixel coordinates whose offset exceeds the threshold. offset Greater than the threshold Therefore, this point was filtered out as an outlier, while the remaining offsets were "lower" (i.e., lower than) Points that do not meet the criteria are considered normal. Finally, based on the similarity characteristics of their spatial neighborhoods, clustering is performed to group the selected points. Such abnormal pixels, if they are spatially adjacent to each other, are merged into a group, generating a set of constrained discontinuous regions.

[0086] Table 2: Example Table of Pixel Constraint Coefficient and Offset Calculation

[0087]

[0088] As shown in Table 2, the table displays the constraint coefficients of some pixels and their neighborhoods, the calculated average neighborhood coefficients, and the final constraint offsets. The constraint offset of pixel (3,3) is significantly higher than that of the other pixels.

[0089] Specifically, such as Figure 2, 5 As shown, the strain response determination module includes:

[0090] The strain data acquisition submodule acquires the nodal strain data of the constrained discontinuous region set, parses the nodal tensor dimension component values, pairs the strain components with the region index number, synchronously adjusts the components of adjacent regions and corrects missing values ​​according to the time series, and generates a regional strain dataset.

[0091] Obtain the generated set of constrained discontinuous regions. This set contains pixel clusters that are initially identified as anomalous, such as a region labeled "Region 1" containing pixel coordinates. and Next, the parsing node is located at the corresponding timestamp. The dimensional component values ​​in the pixel difference tensor structure are extracted, specifically the coupling difference value of each node is extracted as its strain component, assuming the pixel The strain components are , pixels The strain components are Then, the strain components are paired with the index number "Region 1" to form a key-value pair structure, that is, the data corresponding to "Region 1". At the same time, based on the time series, the timestamp is retrieved. Data from the same region, assuming "Region 1" is... The strain component at time t is During processing, it is necessary to synchronize and adjust the data and correct any missing values. For example, if in At that moment, a new anomaly was added to "Region 1". But At that moment, the point is a normal point, causing it to... If there is no corresponding value in the dataset, interpolation will be used to correct the missing value. Specifically, the value at that point will be taken as the reference value. Moment The average value of the strain components of normal pixels within the spatial neighborhood, assuming its neighborhood mean is... Then As this point is The correction value at time is obtained by applying the correction value to the discontinuous region over two consecutive time frames. and After acquiring, arranging, and correcting the data, a structured regional strain dataset containing multiple time slices and spatial regional strain information is generated.

[0092] The strain ratio calculation submodule extracts the principal and secondary direction components based on the regional strain dataset, calculates the principal-secondary ratio and compares it with the average ratio benchmark, filters out regions where the difference exceeds the benchmark, and generates a tensor dimension strain ratio difference set.

[0093] Based on the generated regional strain dataset, which contains multi-dimensional quantitative information of anomalous regions at continuous time points, the timestamps of each region are extracted. The term "major directional component" is defined as the average value of the "coupling difference value" of pixels within a region, while the "minor directional component" is defined as the average value of the "brightness gradient magnitude" of pixels within a region. Taking "Region 1" as an example, its major directional component value is... Secondly, the directional component values, assuming the corresponding brightness gradient magnitudes extracted from the pixel difference tensor structure are respectively and Then its mean is Next, calculate the ratio of the major and minor components, i.e. Then, the calculated ratio is compared with a preset average ratio benchmark, which is determined by... After statistical analysis of random regions in a set of standard defect-free sample images, the average ratio obtained from the statistics is assumed to be... The standard deviation is The criterion for judging the magnitude of the difference is set as the mean plus three standard deviations, i.e. Any ratio exceeding The situation was judged as having a "relatively high" difference, due to the ratio in "Region 1". Greater than Therefore, this region was selected based on the difference between its ratio and the benchmark average. The data will be recorded, and by performing the same extraction, calculation, and comparison filtering process on the regions, a tensor dimension strain ratio difference set containing only regions whose difference magnitude exceeds the baseline and their difference values ​​will be generated.

[0094] The Hidden Markov Recognition Submodule calls the tensor dimension strain ratio difference set, calculates the temporal difference transition probability to fit the observed sequence, judges the sequence change based on the Hidden Markov state transition characteristics, identifies persistent and transient offset regions, and generates an abnormal response region type set.

[0095] The generated tensor dimension strain ratio difference set is invoked. This set constitutes a time-series sequence of observations. For example, for "Region 1", the differences occur over four consecutive time steps. The observed differences within the range were respectively To determine the state of the sequence, two implicit states need to be defined first: state A (normal and stable) and state B (abnormal fluctuations), and a difference threshold (such as the difference magnitude judgment criterion used in the previous step) needs to be used. The threshold (the difference between the ratio and the benchmark mean) discretizes continuous observations into two observation symbols: symbol O1 (the difference is less than the benchmark mean). ) and symbol O2 (difference value greater than or equal to Therefore, the observation sequence The sequence is converted to [O1, O2, O2, O2]. Next, based on the original large-scale data statistics, the transition probabilities between states and the emission probability of a specific observation symbol in a specific state are calculated. Specific values ​​are shown in Table 3. Then, for the observation sequence [O1, O2, O2, O2], its corresponding implicit state sequence is calculated. This process is achieved by iteratively calculating the maximum probability path to reach state A or state B and generate the corresponding observation symbol at each time step. For example, in... When the observation is O1, the probability of reaching state A is higher. When the observation is O2, calculate the probability of transitioning from state A to state A and then emitting O2 again, and the probability of transitioning from state A to state B and then emitting O2 again. Select the one with the larger probability as the current optimal path. This process is repeated throughout the entire sequence to obtain the most probable state transition path, assuming it to be [state A, state B, state B, state B]. Finally, based on this state sequence, determine the regional changes. If a region remains in state B for more than a preset persistence threshold (e.g., ...), the change is considered. If the state is a continuous offset region (e.g., [state A, state B, state A]), it is identified as a "continuous offset region". Conversely, if a pattern like [state A, state B, state A] appears, it is identified as a "transient offset region". In the example, "region 1" is continuous. The current state is B, exceeding the threshold. Therefore, it is classified as "persistent", and the classification results of the regions are finally integrated to generate a set of abnormal response region types.

[0096] Table 3: State Transition and Emission Probability Table

[0097]

[0098] As shown in Table 3, this table lists the state transition probability and emission probability used for sequence change determination. The preset probability values ​​are the basis for identifying the dynamic characteristics of the region.

[0099] Specifically, such as Figure 2 , 6 As shown, the coupling focus mapping module includes:

[0100] The neighborhood data parsing submodule obtains the abnormal response region type set, parses the pixel response sequence, extracts the neighborhood tensor block based on the spatial index and establishes gradient reference data, normalizes according to the response weight and calculates the response difference between pixels, summarizes the coupling balance parameters, and generates the initial coupling balance value.

[0101] Obtain the generated set of abnormal response region types, and parse out a pixel response sequence identified as a "persistent offset region" with index "region 1". This sequence contains pixel coordinates. and Then, based on these two spatial indices, extract... A neighborhood tensor block is generated, and the brightness gradient magnitudes of pixels within the neighborhood are retrieved from the original pixel difference tensor structure to establish gradient reference data. For example, for a pixel... The neighborhood of , whose gradient reference data is a region containing gradient magnitude The matrix is ​​then used to assign response weights to each pixel based on the region type, for pixels within the "persistent offset region". and Assign weights Normal pixels in its neighborhood that do not belong to any abnormal region are assigned weights. Then, the gradient reference data is normalized. Specifically, the sum of the weighted gradient values ​​of the pixels in the neighborhood is calculated, assuming this value is... Then pixel gradient magnitude Normalized to Next, calculate the center pixel and its... The normalized response difference between neighboring pixels, i.e., the normalized value is calculated. The absolute value of the difference from the normalized value of each neighbor, for example, the difference with a certain neighbor is... Then, put this The differences are summed to obtain a coupling balance parameter. Let the summation result be... This process is applied to pixels within “Region 1”, and the calculated coupling balance parameters are finally summarized to generate the initial coupling balance value.

[0102] The neighborhood gradient comparison submodule calls the initial coupling balance value, calculates the gradient change rate for the pixel neighborhood tensor block, compares the pixel gradient change rate with the abrupt threshold, filters the set of pixels whose change rate exceeds the threshold and calculates the local difference response value to obtain the local difference response coefficient.

[0103] Call the generated initial coupling balance value, for example, pixel. The initial coupling equilibrium value is And for the pixel located The gradient rate of change of the neighborhood tensor block is calculated by first obtaining the average value of the initial coupling equilibrium value of the pixels in the neighborhood, assuming it is . Then calculate the difference between the balance value of the center pixel and the average value, and divide it by the unit spatial distance between pixels. ,Right now This is a pixel. The gradient change rate is then calculated, and subsequently compared with a preset mutation threshold. Setting this threshold requires first calculating the gradient change rate of pixels in the image, forming a global gradient change rate set, and then calculating the arithmetic mean and standard deviation of this set. (Assuming this is done by analyzing the gradient change rate of the feeder conveyor belt...) Analysis of normally running frame images yields the global average rate of change. The standard deviation is The mutation threshold is then set to the mean plus twice the standard deviation, i.e. The gradient rate of change is Those 1 and below are considered "normal," while those exceeding 1 are considered "normal." The one that is considered "protruding" is due to the pixel. rate of change Greater than the threshold Therefore, this pixel is selected, forming a set of pixels whose rate of change exceeds a threshold. Then, the local difference response value of this pixel is calculated, which is equal to the difference between its gradient rate of change and the abrupt change threshold. This difference response value is used as the final output to obtain the local difference response coefficient of the pixel.

[0104] The energy distribution generation submodule performs weighted superposition calculation on pixel energy factors based on local difference response coefficients, constructs energy transfer data and integrates accumulated energy, calculates the superposition intensity of energy intervals, and obtains the energy distribution map of the defect area.

[0105] Based on the calculated local difference response coefficients, such as pixel points The coefficient is The coupling difference values ​​(here, the pixel energy factor) in the original pixel difference tensor structure are weighted and superimposed for calculation, assuming that the pixel The coupling difference value is Its weighted energy value This calculation is applied to the selected pixels with locally different response coefficients to construct energy transfer data. This data is a mapping table that records the coordinates of each anomalous pixel and its corresponding weighted energy value. Then, the cumulative energy of each anomalous region is calculated. Specifically, the weighted energy values ​​of the pixels within a region (e.g., "Region 1") are summed. Assuming that "Region 1" also contains pixels... Its weighted energy is Then the cumulative energy of "Region 1" is Then, the superposition intensity of energy intervals is calculated. First, energy intervals are defined, for example, a "low energy" interval is greater than or equal to 0 and less than 100, a "medium energy" interval is greater than or equal to 100 and less than 200, and a "high energy" interval is greater than or equal to 200. The cumulative energy value of the region is then determined. The superposition intensity of a region is determined by its assigned interval. In this example, the energy of "Region 1" is... As shown in Table 4, the process is repeated for the identified abnormal areas, which fall within the "medium energy" range. The energy characteristics of each area are quantified and classified, and finally a defect area energy distribution map is obtained, which characterizes the energy intensity and distribution state of different defect areas.

[0106] Table 4: Energy Distribution in Defect Region

[0107]

[0108] As shown in Table 4, this table displays the cumulative energy calculated after energy weighting and integration for different anomalous regions, and classifies them according to the preset energy range to form the final energy distribution result.

[0109] Specifically, such as Figure 2 , 7 As shown, the boundary stability assessment module includes:

[0110] The energy data receiving submodule acquires the energy distribution map of the defect area, performs grayscale channel separation on the image pixels and extracts the energy amplitude, calculates and normalizes the energy difference between adjacent pixels, scans the energy distribution layer by layer according to the multi-layer energy threshold interval and counts the response density, and generates a multi-layer energy response dataset.

[0111] Obtain the generated energy distribution map of the defect region, which includes "Region 1" identified as "medium energy" with a cumulative energy of [value missing]. And composed of pixels and The process begins by separating the grayscale channels of these two pixels. Since the energy distribution map is single-channel data, this step directly extracts the energy value of each pixel as its energy amplitude, i.e., the pixel's energy level. The energy range is , pixels The energy range is Next, the energy difference between these two adjacent pixels is calculated, and its value is... Then, this difference is normalized. The normalization reference is the maximum possible inter-pixel energy difference determined by analyzing a large amount of image data collected under normal operating conditions of the conveyor belt system. This value is set as... The normalized energy difference is Subsequently, based on preset multi-layer energy threshold intervals—namely, "low energy" greater than or equal to 0 and less than 100, "medium energy" greater than or equal to 100 and less than 200, and "high energy" greater than or equal to 200—the entire energy distribution map is scanned in layers. When scanning the "medium energy" layer, the response density of pixels in that layer is statistically analyzed. Specifically, the image is divided into layers. The grid is calculated, and the proportion of pixels belonging to the "medium energy" range within each grid is calculated out of the total number of pixels in the grid. For example, in a grid containing "region 1", if there are... If a pixel belongs to the "medium energy" range, then the response density of the grid is: This statistic is performed on the grid and energy layers, ultimately generating a multi-layer energy response dataset that records the response density of multiple energy layers at different spatial locations.

[0112] The energy gradient detection submodule, based on a multi-layer energy response dataset, calculates pixel energy differences and quantizes gradient directions, extracts time-series energy changes and calculates gradient direction offsets, determines the range of directional continuous regions based on gradient direction stability conditions, and obtains a set of spatial energy continuity parameters.

[0113] Based on the generated multi-layer energy response dataset and the original energy distribution map, firstly targeting The energy distribution map at time step calculates the energy difference of pixels, in pixels. For example, its energy value is Its right pixel Energy value The pixels below it Energy value The energy difference in the x-direction is divided into The energy difference in the y-direction is divided into This forms the gradient vector. And calculate its gradient direction, the angle of which is . The result is Next, extract the pixel in the range of degrees. Energy data at time t, assuming its energy is His neighbors and The energy is respectively and Calculations yielded The gradient direction at time step is Then calculate the gradient direction offset between these two time points, i.e. Then, based on the gradient direction stability condition, the range of the directional continuous region is determined. This stability condition is set as follows: if a pixel and its... At least in the neighborhood The gradient direction offset of each pixel is less than a preset threshold. If the pixel is oriented continuously, then this threshold is considered to be continuous in direction. According to Statistical analysis of normally functioning frames, taking the offset distribution. quantile, set as Degree, due to pixels offset Degree less than The degree is determined by the fact that, assuming its neighborhood satisfies the above conditions, the point is marked as directionally continuous. Pixels marked as directionally continuous and spatially connected are grouped together to form a directionally continuous region, and the average offset of the pixels within this region is calculated as its continuity parameter. For example... The final result is a set of spatial energy continuity parameters that includes multiple directional continuous regions and their range and continuity parameters.

[0114] The noise suppression integration submodule calls the spatial energy continuity parameter set, performs sliding smoothing on the temporal energy points and calculates the prediction error, estimates the difference between the predicted and observed values ​​based on the Kalman filter update equation, adjusts the outlier weights to correct the aggregated energy data, and generates the defect feature integration result.

[0115] The generated set of spatial energy continuity parameters is invoked, and for one of the temporal energy points identified as having noise, such as a pixel... Five consecutive time steps arrive The observed energy value sequence is ,in For a spike, first perform sliding smoothing on the sequence using a size of The window, in Time, based on the first two values The predicted current value is Then the prediction error is the difference between the observed value and the predicted value, i.e. Next, an estimation is performed based on a predictive update mechanism. The difference between the predicted and observed values ​​at time [time], At time t, the state estimate is ,predict The value at time The observed value at this time is The difference between the observed and predicted values ​​is Then, the weight of the outlier is adjusted to correct the data. The weight adjustment is achieved by calculating a gain coefficient. To achieve, The value is determined by the prediction uncertainty. With observation uncertainty The ratio determines, that is Assuming that the uncertainty is predicted based on system stability. Set as However, because this point is a sudden spike, the observation uncertainty is... Set to a larger value ,but The corrected energy value is the old estimate plus the product of the gain and the difference, i.e. As shown in Table 5, the original observation values Replace with correction value By performing this weight adjustment and data correction on the identified outliers, a smoother and more accurate defect feature integration result with noise suppression is finally generated.

[0116] Table 5: Examples of Pixel Energy Temporal Filtering

[0117]

[0118] As shown in Table 5, this table records the energy value changes of a single pixel at continuous time steps, as well as the process of correcting it through prediction and update mechanisms. The abnormal observation value of 145.2 at time t-1 was corrected to 111.28.

[0119] Please see Figure 8 The image defect testing equipment control method is executed based on the aforementioned image defect testing equipment control system and includes the following steps:

[0120] S1: Perform multi-dimensional feature analysis on the input screen, extract pixel unit grayscale difference, brightness gradient vector and chromaticity phase angle, perform spatiotemporal correlation and tensor organization, and construct pixel difference tensor structure;

[0121] S2: Call the pixel difference tensor structure, calculate the tensor change rate of the time dimension, superimpose the spatiotemporal change rate to form a multidimensional constraint matrix, compare the constraint offset with the response threshold, cluster the pixel positions that exceed the threshold, and generate a set of constraint discontinuous regions.

[0122] S3: Receive the set of constrained discontinuous regions, calculate the tensor dimension strain ratio and difference magnitude, input the hidden Markov model to identify persistent and transient anomaly regions, and generate a set of anomaly response region types.

[0123] S4: Call the abnormal response region type set, extract the neighborhood tensor block for gradient comparison. When the change rate of the tensor coupling balance value exceeds the mutation threshold, the weighted local difference response value is injected into the energy coefficient of the center pixel to form the energy distribution map of the defect region.

[0124] S5: Call the energy distribution map of the defect area, identify high response areas through multi-layer threshold scanning, detect the continuity of spatial energy gradient direction and temporal stability, mark the effective defect area, use the Kalman filter algorithm to suppress noise, and output the integrated defect feature result.

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A control system for an image defect testing equipment, characterized in that, The system includes: The difference tensor construction module performs multi-dimensional feature analysis on the input image, extracts pixel unit grayscale difference, brightness gradient vector and chromaticity phase angle, performs spatiotemporal correlation and tensor organization, constructs pixel difference tensor structure, and passes it to the spatiotemporal constraint construction module. The spatiotemporal constraint construction module receives the pixel difference tensor structure, calculates the change rate of the time dimension tensor, superimposes the spatiotemporal change rate to form a multidimensional constraint matrix, compares the constraint offset with the preset response threshold, clusters the pixel positions that exceed the threshold, generates a set of constraint discontinuous regions, and transmits it to the strain response determination module. The strain response determination module receives the set of constrained discontinuous regions, calculates the tensor dimension strain ratio and difference magnitude, inputs it into the hidden Markov model to identify persistent and transient abnormal regions, generates a set of abnormal response region types, and passes it to the coupling focus mapping module. The coupling focus mapping module receives the abnormal response region type set, extracts the neighboring tensor blocks for gradient comparison, and when the change rate of the tensor coupling balance value exceeds the preset mutation threshold, the weighted local difference response value is injected into the energy coefficient of the center pixel to form the energy distribution map of the defect region, which is then transmitted to the boundary stability assessment module.

2. The control system for the image defect testing equipment according to claim 1, characterized in that, The pixel difference tensor structure includes a gray-level difference feature domain, a brightness gradient distribution layer, and a chromaticity angle distribution map. The multidimensional constraint matrix includes a time rate of change component, a spatial clustering component, and a threshold limit group. The abnormal response region type set includes persistent abnormal region categories, transient abnormal region categories, and response intensity levels. The defect region energy distribution map includes a high-energy concentration region, an energy attenuation band, and an energy transition layer.

3. The control system for the image defect testing equipment according to claim 1, characterized in that, The difference tensor construction module includes: The grayscale difference extraction submodule obtains the grayscale value of the pixel unit in the input image, calculates the grayscale difference amplitude between adjacent pixels according to the pixel sequence position, marks the grayscale change relationship according to the row and column index and removes abnormal pixel differences, and generates a pixel grayscale difference distribution set. The brightness gradient calculation submodule calculates the brightness difference values ​​of adjacent pixels in the horizontal and vertical directions based on the pixel grayscale difference distribution set, superimposes them into a local brightness gradient vector, determines the brightness direction distribution according to the pixel coordinates, and obtains the pixel brightness gradient vector field. The chroma phase integration submodule calculates the phase offset angle between chroma channels based on the pixel luminance gradient vector field, performs multi-channel coupling based on the luminance gradient direction and chroma phase angle, maps and reassembles the channel data, and obtains the pixel difference tensor structure.

4. The image defect testing equipment control system according to claim 3, characterized in that, The spatiotemporal constraint construction module includes: The temporal deformation extraction submodule obtains the pixel difference tensor structure, parses the node coordinate values ​​and strain components, arranges the deformation data of multiple frames in sequence according to the time series identifier, calculates the displacement change rate of adjacent time frames, and generates a time dimension change rate set according to the time step. The tensor change calculation submodule, based on the time dimension change rate set, calls the spatial component information of the pixel difference tensor structure, performs superposition calculation on the time change rate and spatial component of multiple dimensions, analyzes the comprehensive change rate value in multiple directions, and establishes a multidimensional constraint matrix. The constraint clustering identification submodule calls the multidimensional constraint matrix to calculate the constraint offset corresponding to the pixel, compares the offset with the response threshold, filters out the pixel coordinates whose offset exceeds the threshold, and performs clustering based on the similarity features of the spatial neighborhood to generate a set of constraint discontinuous regions.

5. The image defect testing equipment control system according to claim 4, characterized in that, The strain response determination module includes: The strain data acquisition submodule acquires the node strain data of the constrained discontinuous region set, parses the node tensor dimension component values, pairs the strain components with the region index number, adjusts the components of adjacent regions synchronously according to the time series and corrects missing values, and generates a regional strain dataset. The strain ratio calculation submodule extracts the principal and secondary direction components based on the regional strain dataset, calculates the principal-secondary ratio and compares it with the average ratio benchmark, filters out regions where the difference exceeds the benchmark, and generates a tensor dimension strain ratio difference set. The average ratio benchmark is obtained by... After statistical analysis of random regions in a standard, defect-free sample image, the following criteria were set: The Hidden Markov Recognition Submodule calls the tensor dimension strain ratio difference set, calculates the temporal difference transition probability to fit the observation sequence, determines the sequence change based on the Hidden Markov state transition characteristics, identifies persistent and transient offset regions, and generates an abnormal response region type set.

6. The image defect testing equipment control system according to claim 5, characterized in that, The coupling focus mapping module includes: The neighborhood data parsing submodule obtains the abnormal response region type set, parses the pixel response sequence, extracts the neighborhood tensor block according to the spatial index and establishes gradient reference data, normalizes according to the response weight and calculates the response difference between pixels, summarizes the coupling balance parameters, and generates the initial coupling balance value. The neighborhood gradient comparison submodule calls the initial coupling balance value, calculates the gradient change rate for the pixel neighborhood tensor block, compares the pixel gradient change rate with the abrupt threshold, filters the set of pixels whose change rate exceeds the threshold and calculates the local difference response value to obtain the local difference response coefficient. The energy distribution generation submodule performs weighted superposition calculation on pixel energy factors based on the local difference response coefficient, constructs energy transfer data and integrates accumulated energy, calculates the superposition intensity of energy intervals, and obtains the energy distribution map of the defect area.

7. The image defect testing equipment control system according to claim 6, characterized in that, The initial coupling balance value refers to the initial coupling amount of the pixel neighborhood calculated based on the pixel response difference and normalized weights in the neighborhood tensor block. The local difference response coefficient refers to the quantitative parameter of difference calculated based on the difference in gradient change rate between the target pixel and its neighboring pixels, combined with the initial coupling equilibrium value. The mutation threshold refers to a threshold parameter determined based on the statistical characteristics of the global gradient change rate.

8. The control system for the image defect testing equipment according to claim 1, characterized in that, The boundary stability assessment module receives the energy distribution map of the defect region, identifies high-response areas through multi-layer threshold scanning, detects the continuity of spatial energy gradient direction and temporal stability, marks effective defect regions, uses Kalman filtering algorithm to suppress noise, and outputs the defect feature integration result. The integrated defect feature results include an effective defect boundary set, a stability evaluation parameter set, and a noise suppression coefficient set.

9. The control system for the image defect testing equipment according to claim 8, characterized in that, The boundary stability assessment module includes: The energy data receiving submodule acquires the energy distribution map of the defect area, performs grayscale channel separation on the image pixels and extracts the energy amplitude, calculates and normalizes the energy difference between adjacent pixels, and scans the energy distribution layer by layer according to the multi-layer energy threshold interval and counts the response density to generate a multi-layer energy response dataset. The energy gradient detection submodule, based on the multi-layer energy response dataset, calculates the pixel energy difference and quantizes the gradient direction, extracts the time series energy change and calculates the gradient direction offset, determines the range of the directional continuous region according to the gradient direction stability condition, and obtains the spatial energy continuity parameter set. The noise suppression integration submodule calls the spatial energy continuity parameter set, performs sliding smoothing on the time-series energy points and calculates the prediction error, estimates the difference between the predicted and observed values ​​based on the Kalman filter update equation, adjusts the outlier weights to correct the aggregated energy data, and generates the defect feature integration result.

10. A control method for a screen defect testing device, characterized in that, The control system of the image defect testing equipment according to any one of claims 1-9 includes the following steps: S1: Perform multi-dimensional feature analysis on the input screen, extract pixel unit grayscale difference, brightness gradient vector and chromaticity phase angle, perform spatiotemporal correlation and tensor organization, and construct pixel difference tensor structure; S2: Call the pixel difference tensor structure, calculate the change rate of the time dimension tensor, superimpose the spatiotemporal change rate to form a multidimensional constraint matrix, compare the constraint offset with the response threshold, cluster the positions of pixels exceeding the threshold, and generate a set of constraint discontinuous regions. S3: Receive the set of constrained discontinuous regions, calculate the tensor dimension strain ratio and difference magnitude, input the hidden Markov model to identify persistent and transient abnormal regions, and generate a set of abnormal response region types. S4: Call the abnormal response region type set, extract the neighborhood tensor block for gradient comparison, and when the change rate of the tensor coupling balance value exceeds the mutation threshold, inject the weighted local difference response value into the energy coefficient of the center pixel to form the energy distribution map of the defect region. S5: Call the energy distribution map of the defect area, identify high response areas through multi-layer threshold scanning, detect the continuity of spatial energy gradient direction and temporal stability, mark the effective defect area, use the Kalman filter algorithm to suppress noise, and output the integrated result of defect features.

Images