Strip detection method based on image data processing

Abstract

The invention relates to the technical field of digital image processing and computer vision, and discloses a strip detection method based on image data processing, which comprises the following steps: acquiring an original image matrix of the surface of a carrier to be detected; dynamically adjusting diffusion flux according to a pixel neighborhood gradient modulus value, and processing an original image matrix to generate a reference matrix; applying spatial perturbation to the reference matrix and generating virtual residual distribution; performing differential operation on the original image matrix and the reference matrix to construct a feature matrix; and calculating an energy offset vector between the feature matrix and the virtual residual distribution, identifying a non-target noise cluster according to the linearity difference of the energy offset vector relative to a preset displacement amount, and reducing the pixel energy weight of the non-target noise cluster. According to the method, decoupling of physical entity features and transient noise features is realized, and weak damage features are reserved while non-uniform illumination and water mist noise are suppressed.

Description

Technical Field

[0001] This invention relates to a strip material detection method based on image data processing, belonging to the field of digital image processing and computer vision technology. Background Technology

[0002] Currently, in high-speed image acquisition and real-time information analysis based on machine vision, the imaging system acquires continuous images containing the target to be analyzed and uses digital image processing algorithms to analyze the pixel matrix to identify target features. Existing technologies mostly adopt linear filtering, fixed threshold segmentation, or background subtraction strategies to try to separate target defects from video streams. However, as rolling processes evolve towards higher speeds, with strip running speeds reaching over 20 m / s, the raw pixel information acquired by the imaging system contains complex dynamic components. The surface texture of the strip and the transient reflection noise generated by the cooling medium exhibit a high degree of spectral overlap in the pixel energy distribution, leading to increased difficulty in decoupling target features from the environmental background.

[0003] In real-world production scenarios, the light field distribution on the strip surface is spatially non-uniform due to the influence of equipment structure. The multiplicative noise generated masks minute defect features in low-contrast ranges. Traditional algorithms face an imbalance between detection sensitivity and false alarm rate when dealing with high-dynamic backgrounds. If the filtering intensity is reduced to preserve microcracks, random isolated noise points generated by cooling water mist are easily misjudged as physical damage. If the smoothing operator is strengthened to suppress background interference, the gradient features of defect edges undergo nonlinear degradation, resulting in the loss of critical information. Although the industry has attempted to introduce statistical models or multi-layer convolutional networks, these methods are computationally expensive and cannot provide real-time responses within 1ms of production cycles, creating a contradiction between algorithm complexity and engineering real-time requirements. In exploring ways to improve detection accuracy, industry research has gradually shifted from hardware performance compensation to the deep evolution of algorithm architecture. However, without the constraints of physical motion laws, simply stacking algorithms... This approach often leads to a technical dead end between processing efficiency and judgment accuracy. For example, Chinese invention patent CN117670858B discloses a method for detecting surface defects in metal strips based on the CFLOW-AD model. While it effectively alleviates the limitations of model training caused by the scarcity of defect samples and enhances stability against static background noise, the massive computational load of the multi-level decoder in high-speed rolling scenarios at speeds of 20 m / s can easily cause processing delays to exceed the 1 ms industrial real-time warning line. More importantly, this solution makes anomaly judgments based on semantic features of the image space and fails to take into account the spatial displacement vector constraints generated by the strip movement. As a result, when faced with random and transient high-frequency noise caused by cooling medium spraying, the system cannot identify the essential differences in motion patterns between noise clusters and physical defects from a physical causal perspective, and the risk of false alarms remains prominent.

[0004] Therefore, how to decouple the pixel-level background energy field from the defect feature energy, and retain the gradient features of weak damage while suppressing non-uniform illumination and water mist noise, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A strip detection method based on image data processing, comprising the following steps:

[0006] Step S1: Obtain the original image matrix of the surface of the carrier to be detected;

[0007] Step S2: Anisotropic diffusion processing is performed on the original image matrix using a diffusion flux adaptive adjustment mechanism based on local gradient divergence to generate a smoothed reference matrix. The diffusion flux adaptive adjustment mechanism dynamically sets the smoothing operator weights according to the gradient magnitude of the pixel neighborhood of the original image matrix, so as to retain low-contrast gray-level transition features while suppressing background texture energy.

[0008] Step S3: Apply a spatial perturbation with a preset displacement to the reference matrix, and invert to generate a virtual residual distribution that characterizes the positional offset of the reference matrix in the spatial coordinate system.

[0009] Step S4: Perform a difference operation on the original image matrix and the reference matrix to construct a feature matrix containing suspected target regions, where the suspected target regions are local connected regions in the feature matrix whose pixel energy values ​​exceed a preset saliency threshold.

[0010] Step S5: Calculate the energy offset vector between the feature matrix and the virtual residual distribution, and determine the linearity difference of the geometric mapping law of the energy offset vector relative to the preset displacement. The geometric mapping law is defined as the characteristic of the feature matrix shifting in the same direction and in the same amount in the spatial domain with the preset displacement. If the linearity difference exceeds the preset deviation threshold, the suspected target area is identified as a non-target noise cluster, and the pixel energy weight of the non-target noise cluster in the feature matrix is ​​reduced based on the preset weight factor, so as to achieve feature decoupling between physical entity features and transient noise moving with the carrier.

[0011] Preferably, during the anisotropic diffusion process, the smoothed background texture energy is extracted simultaneously and its information entropy in the spatial domain is calculated. The information entropy is obtained by logarithmically weighting the normalized grayscale distribution probability of the background texture energy and is used to characterize the imaging system's ability to perceive surface details of the carrier under test. When the value of the information entropy is continuously lower than the preset quality threshold, a system health warning signal is generated, which characterizes the performance degradation of the input link for acquiring the original image matrix, and the parameters of the image acquisition frequency or exposure time are corrected according to the warning signal.

[0012] Preferably, when performing anisotropic diffusion processing on the original image matrix using the diffusion flux adaptive adjustment mechanism based on local gradient divergence in step S2, the following steps are included: calculating the local gradient divergence of each pixel in the original image matrix; determining the diffusion coefficient based on the local gradient divergence, and controlling the energy exchange rate of the diffusion operator in different texture feature regions based on the diffusion coefficient; enhancing the smoothing intensity of the diffusion operator in texture flat regions, and suppressing the diffusion flux of the diffusion operator in gradient jump regions.

[0013] Preferably, when applying a preset displacement to the reference matrix in step S3, the following steps are included: obtaining the relative motion velocity of the surface of the carrier to be detected and the image acquisition frequency; determining the preset displacement based on the relative motion velocity and the image acquisition frequency; and moving the reference matrix in the spatial coordinate system by the preset displacement along the displacement direction corresponding to the relative motion velocity to generate a virtual residual distribution.

[0014] Preferably, when identifying a suspected target region as a non-target noise cluster in step S5, the following steps are included: calculating the average gradient vector direction of the suspected target region; if the consistency between the average gradient vector direction and the energy offset vector direction is lower than a preset ratio, then the suspected target region is determined to be a non-physical feature generated by the light field interference of the imaging environment.

[0015] Preferably, the pixel energy weights of non-target noise clusters in the feature matrix are reduced according to the following calculation rules: ,in, For the adjusted energy weights, As the initial weights, The magnitude of the energy offset vector. λ is the preset displacement mapping length, and λ is the preset response adjustment coefficient.

[0016] Preferably, after reducing the pixel energy weights of non-target noise clusters in the feature matrix, the method further includes the following steps: performing nonlinear saliency enhancement processing on the feature matrix; using geometric parameter extraction operators to calculate the geometric morphology parameters of physical entity features in the enhanced feature matrix; and determining the quality state of the surface of the carrier to be detected based on the geometric morphology parameters.

[0017] Preferably, when calculating the energy offset vector between the feature matrix and the virtual residual distribution in step S5, the following steps are included: constructing a local search window between the feature matrix and the virtual residual distribution; performing cross-correlation operation within the local search window to determine the displacement trajectory of the energy centroid; and decomposing the displacement trajectory in polar coordinates to obtain the amplitude and phase information of the energy offset vector.

[0018] Preferably, when identifying a suspected target region as a non-target noise cluster in step S5, the method further includes the following steps: obtaining the spatiotemporal consistency constraints between adjacent multi-frame images; determining the morphological stability parameters of the suspected target region during the temporal evolution process; if the morphological stability parameters are lower than a preset stability threshold, then the suspected target region is determined to be transient high-frequency noise caused by fluid medium interference.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In image data processing, by constructing a reference matrix based on anisotropic diffusion and performing pixel-level ratio calculations, the multiplicative noise in the discrete energy field of the image is converted into an additive response in the contrast space. This solves the interference of non-uniform illumination on weak feature extraction and achieves deep decoupling of defect signals and complex background energy without the need for additional lighting hardware.

[0021] 2. By introducing temporal displacement constraints of image sequences and combining them with the consistency verification of the average gradient vector direction of suspected regions, a physical causal relationship between spatial displacement and vector evolution is established. This enables the system to distinguish between physical defects that move with the strip and randomly distributed transient water mist noise, effectively solving the problem of false alarms caused by random overlap of noise clusters under high-speed operating conditions and ensuring the reliability of detection conclusions in harsh environments.

[0022] 3. An adaptive adjustment mechanism for diffusion flux based on local gradient divergence is adopted, which enables the anisotropic diffusion operator to dynamically adjust the smoothing intensity according to the texture complexity of the pixel neighborhood. While suppressing high-energy background textures, it prevents low-contrast defect edges from being over-smoothed, achieving a dynamic balance between background suppression accuracy and feature fidelity, and improving the system's ability to generalize to different material surface states. Attached Figure Description

[0023] Figure 1 This is a flowchart of the detection logic for spatial perturbation and feature decoupling in this invention;

[0024] Figure 2 This is a schematic diagram of the hardware architecture and data interaction principle of the strip inspection system of the present invention. Detailed Implementation

[0025] To make the technical problems, technical solutions and beneficial effects of the present invention clearer and more understandable, the present invention will be described in detail below with reference to specific embodiments. It should be noted that the following embodiments are intended to explain the present invention and not to limit the scope of protection of the present invention.

[0026] This invention provides a strip inspection method based on image data processing. It acquires the original image matrix of the surface of the carrier to be inspected, and uses local gradient divergence to adjust the diffusion flux to generate a background reference matrix. By applying spatial displacement perturbation to the reference matrix and calculating the energy offset vector of the feature matrix, combined with temporal displacement constraints of the image sequence and gradient vector direction consistency judgment, the decoupling and identification of surface defect features of the strip are achieved. Due to the transient reflection noise generated by the cooling medium spray and the non-uniform background texture caused by rolling line vibration on the strip surface under high-speed rolling conditions, the system acquires the original image matrix I of the carrier surface to be inspected and treats it as a two-dimensional discrete energy field. To reconstruct a smooth background layer, the processor calls a 3×3 pixel sliding window to traverse the original image matrix. The gradient magnitude in the neighborhood of each pixel is calculated, and the smoothing weight of the anisotropic diffusion operator is dynamically set according to the gradient magnitude. When the local gradient divergence in the sliding window is lower than the preset stability threshold, the region is determined to be a stable background texture. At this time, the diffusion coefficient along the strip running direction is increased to 0.85 to generate the reference matrix B. By adjusting the diffusion flux in the horizontal and vertical directions, the method suppresses the texture energy in the strip running direction and retains the abrupt edges perpendicular to the running direction, providing a background reference for the subsequent extraction of defect signals. Since the light field distribution in the rolling field is non-uniform due to equipment obstruction and light source attenuation, the multiplicative noise generated often masks the small defect features in the low contrast range. After generating the reference matrix B, the system performs residual feature decoupling operation. The processing unit calculates the ratio between the original image matrix I and the reference matrix B at the pixel level and constructs the residual feature distribution model through the following formula: Where R is the residual energy distribution, I is the original image matrix, B is the reference matrix, and α is a smoothing factor used to prevent the denominator from being zero, with a value of 0.01. Under this operation path, the system maps the image data from the original grayscale space to the normalized contrast space, eliminates the background bias caused by non-uniform illumination, and performs amplitude stretching of the defect signal energy in the shadow area relative to the local background, thereby achieving deep decoupling between the defect signal and the complex background energy field.

[0027] To address the highly directional visual pseudo-edges formed in the pixel domain by direct sunlight or shadows from mechanical structures, the method applies a displacement perturbation to the reference matrix B to perform interference recognition; the strip running speed v is obtained as... m / s and the image acquisition frequency f is The Hz value is used to determine the preset displacement Δs as 2 pixels. The reference matrix B is then translated by a length Δs along the displacement direction in the spatial coordinate system to generate a virtual residual distribution matrix. The system calculates the characteristic matrix and the virtual residual distribution matrix. The energy offset vector between the two regions is used to determine the linearity difference of the offset vector relative to the preset displacement Δs. The specific calculation logic is as follows: extract the spatial coordinates of all pixels in the suspected target area and construct a two-dimensional energy centroid coordinate set; calculate the discrete point covariance matrix of the set, and extract the principal component eigenvalues ​​with index 1 and index 2 in the matrix; divide the principal component eigenvalue with index 1 by the sum of the eigenvalues ​​with index 1 and index 2, and the result is the linearity value. If the value is lower than 0.7, the region is determined to be a non-target noise cluster generated by light field interference, and the pixel energy weight is reduced by calling the weight adjustment function shown below. The processing unit performs covariance matrix decomposition on the energy offset vector set in the suspected target area to obtain the eigenvalues ​​representing the displacement concentration, and defines the ratio of the principal direction eigenvalue to the total eigenvalue as the linearity L, where the linearity L satisfies the value range of 0 to 1; if the calculated linearity L is lower than the preset deviation threshold of 0.7, the region is determined to be a non-target noise cluster generated by light field interference, and the pixel energy weight is reduced by calling the weight adjustment function shown below. ,in, For the adjusted energy weights, As the initial weights, The magnitude of the energy offset vector. The mapping length is the preset displacement, and λ is the preset response adjustment coefficient. This process utilizes the physical causal relationship between spatial displacement and vector evolution to eliminate false alarm interference caused by the discontinuity of the illumination field.

[0028] Under extreme conditions of cooling water mist or oil film coverage, transient discrete pixel clusters appearing in images exhibit morphological properties similar to physical damage features in a single frame. A spatiotemporal gradient vector direction consistency verification mechanism is introduced to perform logical arbitration; the average gradient vector direction θ of suspected defect regions in the feature matrix is ​​extracted, and a direction matching constraint relationship is established between adjacent frames based on the displacement Δs; the processor determines the gradient vector direction of the suspected target in the current frame. Gradient vector direction corresponding to the position in the preceding frame The difference is checked against a preset angle threshold of 15 degrees. When the gradient vector direction maintains statistical consistency between adjacent frames and its evolution trajectory matches the physical displacement generated by the running speed v, the suspected defect area is determined to be a real defect; otherwise, it is classified as environmental noise and its energy is zeroed. This mechanism utilizes geometric stability as an information filter, cutting off the transmission path of random disturbance signals from a physical dimension. To ensure the long-term stability of the system and monitor imaging quality, the method simultaneously extracts the background high-frequency energy term δ smoothed by the diffusion operator and calculates its spatial domain information entropy H during the anisotropic diffusion process. The unit normalizes the smoothed background texture energy to obtain the grayscale distribution probability, and uses a discrete logarithmic weighted summation algorithm to generate a quality index characterizing the imaging performance. When the information entropy H value is continuously lower than the preset quality threshold of 0.35, it is determined that the current input data has progressive blurring, and a system health warning signal is output to trigger parameter correction of camera exposure time or image acquisition frequency. The entire process uses matrix-based elementary algebra operations and local convolution operators for data flow, avoiding dependence on large-scale computing resources, thereby ensuring sub-millisecond real-time response on the edge computing unit.

[0029] Example 1: In a high-speed cold rolling production line for strip with a running speed v of 20 m / s, the strip surface is covered with transient water mist and oil film due to the sprayed cooling medium. Under an imaging environment with an image acquisition frequency f of 1000 Hz, this generates grayscale jump signals that highly overlap with defect features. Due to the non-uniform light field energy bias in the image edge areas caused by the ambient lighting structure, weak scratch features in shadow or bright light areas are often masked by noise. Furthermore, traditional algorithms amplify random pixel cluster interference generated by water mist when improving detection sensitivity. To address this situation, the system acquires the original image matrix I and uses an anisotropic diffusion operator to adjust the diffusion flux based on the gradient magnitude of the pixel neighborhood. A smoothing operation with a coefficient of 0.85 is performed in the strip running direction to generate a reference matrix B. Based on this, the processor performs a nonlinear decoupling operation according to the ratio relationship shown in the formula below: Where R is the residual energy distribution, I is the original image matrix, B is the reference matrix, and α is the smoothing factor, with a value of 0.01. In this computational path, the smoothed background provided by the reference matrix B serves as a dynamic reference for the extraction of residual features, enabling the system to transform multiplicative noise into additive feature responses in the contrast space. This achieves effective stripping of defect signal energy against a non-uniform background. To eliminate visual false edges caused by mechanical structure shadows, the processor applies a spatial perturbation to the reference matrix B based on a preset displacement Δs of 2 pixels determined by v and f, and generates a virtual residual distribution matrix. By calculating the characteristic matrix and the virtual residual distribution matrix The system determines the linearity difference between the energy offset vector and the preset displacement Δs. If the linearity L is lower than the preset deviation threshold of 0.7, the system calls the weight adjustment function shown below to reduce the pixel energy weight of that region. ,in, For the adjusted energy weights, As the initial weights, The magnitude of the energy offset vector. λ is the mapping length of the preset displacement, and λ is the response adjustment coefficient. Here, the physical causal constraints of spatial displacement and vector evolution are used to enable the system to identify false edges based on the dynamic stability of features and eliminate noise.

[0030] After weight adjustment, the system extracts the gradient vector direction of the suspected target region. And the gradient vector direction at the corresponding position in the previous frame. The system performs direction matching verification. It constructs a statistical distribution histogram of the gradient direction of the suspected target region in the current frame and calculates the Barthel distance between this histogram and the corresponding histogram in the previous frame. When the difference in gradient vector direction between adjacent frames is within a 15-degree angle threshold and the Barthel distance is within a preset convergence interval, the suspected target is determined to be a real defect with continuous physical motion, and the energy of randomly distributed water mist particles with disordered gradient directions is zeroed. During this process, the background reference generated by anisotropic diffusion and the geometric consistency verification in the spatiotemporal domain form a closed loop. The known motion characteristics of the strip are used as an information filter, solving the problem of high-dynamic background... The technical dilemma between target detection sensitivity and false alarm suppression is addressed by establishing a collaborative path of local diffusion background reconstruction and spatial displacement causal verification, thus transforming from single-pixel feature detection to physical logic mapping. The combination of the reference matrix B and the residual energy distribution R decouples the contrast deviation caused by non-uniform illumination, while the spatial perturbation Δs and gradient vector direction consistency verification perform multiple logical locking on suspected targets from a kinematic dimension. This multi-dimensional feature collaborative architecture enables the system to solve the traditional signal-to-noise ratio limitation problem under the physical displacement constraint framework by changing the judgment boundary of the problem when dealing with complex reflection noise under high-speed cold rolling conditions.

[0031] Example 2: In a physical testing platform for strip surface quality equipped with a 2048-pixel linear array camera, the strip running speed v was set to 25 m / s. The original test data came from the real-time acquisition of the stainless steel strip surface by the linear array camera at a sampling line frequency of 100 kHz. The optical resolution of the camera is 0.05 mm / pixel and the dynamic range is not less than 12 bits. This experiment was used to verify the feature extraction accuracy of the method under extreme reflection environments with superimposed Gaussian white noise with a signal-to-noise ratio of 15 dB and a light field uniformity deviation of 25%. The sampling line frequency f was set considering the matching relationship between the strip running speed and the minimum feature size of the physical target of 0.1 mm. In order to meet the signal sampling criteria and balance the computational load of the back-end processing unit, the sampling line frequency tended to its lower limit of calculation to maintain a longitudinal sampling interval of 0.25 mm, that is, f was set to 100 kHz.

[0032] To construct a multi-dimensional control system, the experiment included the present invention sample group, a missing diffusion reconstruction control group, a missing spatial perturbation control group, and an out-of-range parameter control group. The present invention sample group adopted the method described in the specific implementation. The missing diffusion reconstruction control group used a median filter operator instead of the diffusion flux adaptive adjustment mechanism based on local gradient divergence. The missing spatial perturbation control group performed threshold segmentation after generating the feature matrix without performing offset vector calculation of the virtual residual distribution. The out-of-range parameter control group set the diffusion coefficient to 0.98, which deviates from the upper limit of the preferred working window of 0.85. The experimental environment simulated water mist noise generated by spraying by setting a semi-transparent scattering sheet on the light source path and introduced a gradient shift in the spatial distribution of light intensity using an asymmetric reflecting mirror. After the experiment started, the processing unit received the original image matrix I containing noise perturbation. In the missing diffusion reconstruction control group, the... The median filter operator's non-selective smoothing of high-frequency details caused the gray level of the crack feature with a width of 0.12 mm to degrade from 185 to 162, resulting in a signal-to-noise ratio improvement of 3.2 dB. In the sample of this invention, the diffusion flux adaptive adjustment mechanism identified the average gradient divergence of the texture region as 12.5. At this time, the diffusion coefficient along the strip running direction was increased to 0.85, and the generated reference matrix B filtered out the high-frequency reflection component. The calculated residual energy distribution R at the defect location had a response value of 0.14, and the signal-to-noise ratio improved to 18.5 dB. When the diffusion coefficient of the out-of-range parameter control group increased from 0.85 to 0.98, the feature energy between adjacent pixels underwent nonlinear diffusion due to excessive diffusion flux saturation, resulting in a 42.5% decrease in the contrast of the defect edge. This indicates that there is a performance inflection point near 0.85 that balances background smoothness and feature fidelity.

[0033] During the interference removal phase, the misclassification rate of the missing spatial perturbation control group for suspected target regions was 15.6%, mainly due to the inability to distinguish visual pseudo-edges formed by light source shadows. The present invention's sample group calculates the virtual residual distribution by applying a preset displacement Δs of 2 pixels to the reference matrix B. The measured linearity of the energy offset vector in the suspected target area was 0.42, which is lower than the preset deviation threshold of 0.7. The processor determined that the signal was a non-target noise cluster and reduced the energy weight. The weight adjustment function output... The initial value was reduced from 1.0 to 0.12, and the system extracted the gradient vector direction. Gradient vector direction corresponding to the position in the preceding frame The direction matching verification was performed, and the gradient direction consistency deviation of the actual defects was less than 8.5 degrees, while the gradient direction of water mist noise showed random fluctuations of more than 45 degrees. Finally, the defect recognition rate of the sample group of this invention was 98.2%, and the false alarm rate was less than 0.5%. The experimental results proved that the strip detection method based on image data processing solved the technical contradiction between multiplicative noise suppression and weak signal preservation under high dynamic imaging background through the synergy of anisotropic diffusion background reconstruction and spatial displacement causal constraints.

[0034] Example 3: This example combines Figures 1 to 2 The method for strip inspection based on image data processing is described, such as... Figure 1 As shown, step S1 is executed to obtain the original image matrix of the surface of the carrier to be detected. Step S2 is executed to perform anisotropic diffusion processing on the original image matrix using a diffusion flux adaptive adjustment mechanism based on local gradient divergence to generate a smoothed reference matrix. Then, step S3 is executed to apply a spatial perturbation with a preset displacement to the reference matrix and invert to generate a virtual residual distribution that represents the positional offset of the reference matrix in the spatial coordinate system. On this basis, step S4 is executed to perform a difference operation on the original image matrix and the reference matrix to construct a feature matrix containing the suspected target region. Finally, step S5 is executed to calculate the energy offset vector between the feature matrix and the virtual residual distribution, determine the linearity difference to identify non-target noise clusters, and reduce their pixel energy weights to achieve feature decoupling.

[0035] like Figure 2As shown, the above detection process operates within a physical interaction hardware architecture. The environment of the carrier to be detected includes the carrier, cooling medium, water mist, and illumination field. The surface reflection signal generated by this environment enters an imaging system composed of a line scan camera and an industrial camera lens. After acquiring data, the imaging system outputs the raw image matrix to the processor in the edge computing unit. The processor internally stores and processes the raw image matrix, the reference matrix, and the feature matrix, and is constrained by the limit of cached data memory bits. At the same time, the processor generates a system health warning signal based on the data analysis results and performs parameter correction on the imaging system through the feedback path, specifically including dynamic adjustment of exposure time and acquisition frequency.

[0036] Example 4: During the continuous operation of the strip surface quality inspection system, the industrial camera lens experiences progressive blurring due to oil mist condensation caused by the evaporation of the emulsion. This results in the global loss of edge components in the original image matrix I. The anisotropic diffusion operator suffers computational degradation during background reconstruction due to the decrease in the original contrast of the input data, thereby reducing the signal-to-noise ratio of the feature matrix. To monitor the performance degradation of this type of equipment, the processor simultaneously extracts the high-frequency energy residual term δ, which was smoothed out by the diffusion operator, during the execution of anisotropic diffusion and the generation of the reference matrix B. The high-frequency energy residual term δ satisfies the following formula: Where δ is the residual matrix representing high-frequency texture information, I is the original image matrix, and B is the smoothed baseline matrix. To obtain quantitative imaging quality evaluation indicators, the system divides the stable background region of the strip corresponding to the residual matrix δ into 256 equally spaced gray-level energy intervals, and counts the pixel distribution frequency in each energy interval to obtain the normalized gray-level distribution probability. The information entropy H in the spatial domain is calculated using a discretized logarithmic weighted summation algorithm, where the formula for calculating the information entropy H is as follows: Where H is the information entropy and g is the grayscale energy level index. Let g be the probability distribution of the g-th gray level energy within the sampling area.

[0037] To determine the logical trigger point for image quality warnings, the processor executes a standardized parameter calibration procedure. During the system self-test phase, a clear background at a standard focal length is acquired as raw reference data. By recording the entropy evolution of this reference data at different simulated blur levels, it is determined that when the contrast loss of a suspected target edge reaches 30%, the corresponding information entropy H decays to 65% of its initial value, thus calibrating the quality threshold value to 0.35. During the system self-test phase, an image of a defect-free strip surface is acquired at a standard focal length as a reference benchmark, and the calculated initial information entropy value is 0.54; manual adjustment... The focal length of the industrial camera lens causes progressive blurring of the image until the contrast loss at the suspected target edge reaches an experimental critical value of 30%. At this point, the corresponding spatial domain information entropy value is recorded as 0.35, and this value is fixed as the logical judgment threshold for subsequent online monitoring. During actual monitoring, if the entropy value of 50 consecutive frames in the current image sequence is lower than 0.35, a system health warning signal is output. After receiving the system health warning signal, the system performs closed-loop feedback correction based on the physical mapping relationship between information entropy H and target contrast, adjusting the exposure parameters according to the linear relationship shown in the formula below: ,in, The adjusted exposure time. The initial exposure time is β, which is a preset linear compensation coefficient with a value of 1.2, and H is the information entropy calculated for the current frame. If the adjusted exposure time exceeds the maximum pulse width limit of the camera, the processor reduces the image acquisition frequency f to extend the photon accumulation time. The internal quality evaluation index is used to compensate for the imaging degradation caused by environmental interference, ensuring that the detection system maintains the consistency of detection conclusions under fluctuating equipment performance.

[0038] Example 5: In the debugging scenario of a newly deployed strip inspection system, the processor executes a calibration program for pixel scale and stability threshold. It calls a calibration plate with standard grid spacing that is in a static state and acquires calibration images through a linear scan camera to calculate the conversion factor between physical space units and pixel coordinate units. For the stability threshold required in the anisotropic diffusion process, the gray-level gradient distribution of the smooth background area in the calibration plate is extracted, the variance mean of the gradient of the 3×3 pixel neighborhood in this area is calculated, and 1.5 times of it is determined as the stability threshold in the diffusion flux adaptive adjustment mechanism. The uncertainty of input data caused by differences in camera installation height or environmental noise is eliminated, so that the construction process of the reference matrix B remains consistent under different hardware layouts.

[0039] When the system is applied to aluminum strip with high reflectivity, the processor performs on-site calibration for non-target noise cluster judgment logic. The system acquires background images when the production line is not in operation with strip, records visual pseudo-edge features caused by the discontinuity of the illumination light field, and records the energy offset vector distribution after applying a preset displacement of 2 pixels Δs. By statistically analyzing the nonlinear discrete characteristics of the pseudo-edge features during the displacement process, the processor calculates the cumulative distribution function of its linearity and determines the value corresponding to the 95% confidence level in this function as the preset deviation threshold under the current light field environment. This process associates the physical displacement causal constraint mechanism with the specific industrial imaging environment, making the reduction of pixel energy weights for interference signals statistically based, and realizing the identification of environmental noise and physical entity features.

[0040] Example 6: In the initial parameter calibration scenario of a cross-batch strip inspection system, the processor executes a benchmark calibration program with anisotropic diffusion smoothing weights to adapt to metal carriers with different reflectivity. It acquires grayscale image streams of the strip surface under no-load operation as calibration samples, uses a sliding window to traverse the sample images, calculates the statistical distribution characteristics of local gradient magnitudes, and then determines a stationary threshold based on the cumulative probability curve of the gradient distribution. The processor statistically analyzes the frequency distribution of gradient values ​​in the background region within the range of 0 to 255, and selects the gradient magnitude with a cumulative probability of 90% as the logical cut-off point for the adaptive adjustment mechanism of diffusion flux. When the mean variance of the background gradient of the calibration sample is measured to be 15.6, the stationarity threshold is determined by proportional mapping calculation. The value is 18.72, thus establishing the generation boundary of the reference matrix B under a specific material texture; when the system faces the condition of imaging position displacement caused by the micro-vibration of the mechanical support, a sub-pixel-level calibration procedure for the spatial perturbation displacement Δs is performed to maintain the discrimination stability of the energy offset vector, and the virtual residual distribution under different perturbation step sizes is recorded. The overlap index with the feature matrix is ​​used, and the value of the response adjustment coefficient λ is obtained by fitting with the least squares method. If the sum of squares of the residuals of the displacement linearity measured by the experiment reaches a minimum value when the response adjustment coefficient λ is equal to 0.55, then this value is solidified as the weight adjustment parameter in the current deployment environment. The whole process eliminates the feature coupling interference caused by installation error or environmental vibration through quantitative physical mapping experiments, and realizes the geometric stability locking of suspected defect signals.

[0041] In the deployment phase of the detection method, the processor performs background suppression parameter calibration for specific materials, selects a sequence of defect-free background images as reference samples, calculates the gradient magnitude distribution of each pixel's 3×3 neighborhood, and statistically analyzes the cumulative probability function under different exposure gains. The 90% quantile value with a cumulative probability is determined as the stabilization threshold for the diffusion flux adaptive adjustment mechanism. By adjusting the diffusion coefficient in steps and comparing it with the background residual variance after generating the baseline matrix B, the local minimum value of the background residual variance with a change rate of less than 5% is selected as the diffusion coefficient to perform background reconstruction on different reflectivity surfaces. Visual pseudo-edges generated by complex lighting environments are identified. The test environment simulates direct light from different intensities and mechanical structure occlusion. The original image matrix I containing non-uniform light field distribution features is collected. Spatial perturbation with a step range of 0.5 pixels to 3 pixels is applied to the baseline matrix B. The energy offset vector of the suspected target area of ​​the feature matrix under different perturbation step sizes is calculated. First-order linear regression is performed on it to calculate the root mean square error of the fitting residual. According to the mapping relationship between the root mean square error and the corresponding physical displacement, the deviation from the linear evolution law in the mapping reaching a statistical limit of 30% is determined as the preset deviation threshold for judging non-target noise clusters. By utilizing the physical causal stability of displacement and evolution, transient reflection noise is separated from physical entity features.

[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A strip material detection method based on image data processing, characterized in that, Includes the following steps: Step S1: Obtain the original image matrix of the surface of the carrier to be detected; Step S2: Anisotropic diffusion processing is performed on the original image matrix using a diffusion flux adaptive adjustment mechanism based on local gradient divergence to generate a smoothed reference matrix. The diffusion flux adaptive adjustment mechanism dynamically sets the smoothing operator weights according to the gradient magnitude of the pixel neighborhood of the original image matrix, so as to retain low-contrast gray-level transition features while suppressing background texture energy. Step S3: Apply a spatial perturbation with a preset displacement to the reference matrix, and invert to generate a virtual residual distribution that characterizes the positional offset of the reference matrix in the spatial coordinate system. Step S4: Perform a difference operation on the original image matrix and the reference matrix to construct a feature matrix containing suspected target regions, where the suspected target regions are local connected regions in the feature matrix whose pixel energy values ​​exceed a preset saliency threshold. Step S5: Calculate the energy offset vector between the feature matrix and the virtual residual distribution, and determine the linearity difference of the geometric mapping law of the energy offset vector relative to the preset displacement. The geometric mapping law is defined as the characteristic of the feature matrix shifting in the same direction and in the same amount in the spatial domain with the preset displacement. If the linearity difference exceeds the preset deviation threshold, the suspected target area is identified as a non-target noise cluster, and the pixel energy weight of the non-target noise cluster in the feature matrix is ​​reduced based on the preset weight factor, so as to achieve feature decoupling between physical entity features and transient noise moving with the carrier.

2. The strip detection method based on image data processing according to claim 1, characterized in that, During the anisotropic diffusion process, the smoothed background texture energy is extracted simultaneously and its information entropy in the spatial domain is calculated. The information entropy is obtained by logarithmically weighting the normalized grayscale distribution probability of the background texture energy and is used to characterize the imaging system's ability to perceive the surface details of the carrier under test. When the value of the information entropy is continuously lower than the preset quality threshold, a system health warning signal is generated, which characterizes the performance degradation of the input link for acquiring the original image matrix. Based on the warning signal, parameter corrections are performed on the image acquisition frequency or exposure time.

3. The strip detection method based on image data processing according to claim 1, characterized in that, In step S2, when performing anisotropic diffusion processing on the original image matrix using the diffusion flux adaptive adjustment mechanism based on local gradient divergence, the following steps are included: calculating the local gradient divergence of each pixel in the original image matrix; determining the diffusion coefficient based on the local gradient divergence, and controlling the energy exchange rate of the diffusion operator in different texture feature regions based on the diffusion coefficient; enhancing the smoothing intensity of the diffusion operator in texture flat regions, and suppressing the diffusion flux of the diffusion operator in gradient jump regions.

4. The strip detection method based on image data processing according to claim 1, characterized in that, When applying a preset displacement to the reference matrix in step S3, the following steps are included: obtaining the relative motion velocity of the surface of the carrier to be detected and the image acquisition frequency; determining the preset displacement based on the relative motion velocity and the image acquisition frequency; and moving the reference matrix in the spatial coordinate system by the preset displacement along the displacement direction corresponding to the relative motion velocity to generate a virtual residual distribution.

5. The strip detection method based on image data processing according to claim 1, characterized in that, When identifying a suspected target region as a non-target noise cluster in step S5, the following steps are included: calculating the average gradient vector direction of the suspected target region; if the consistency between the average gradient vector direction and the energy offset vector direction is lower than a preset ratio, the suspected target region is determined to be a non-physical feature generated by the light field interference of the imaging environment.

6. The strip detection method based on image data processing according to claim 1, characterized in that, The pixel energy weights of non-target noise clusters in the feature matrix are reduced according to the following calculation rules: ,in, For the adjusted energy weights, As the initial weights, The magnitude of the energy offset vector. λ is the preset mapping length for the displacement, and λ is the preset response adjustment coefficient.

7. The strip detection method based on image data processing according to claim 1, characterized in that, After reducing the pixel energy weights of non-target noise clusters in the feature matrix, the following steps are also included: performing nonlinear saliency enhancement processing on the feature matrix; using geometric parameter extraction operators to calculate the geometric morphology parameters of physical entity features in the enhanced feature matrix; and determining the quality state of the surface of the carrier to be detected based on the geometric morphology parameters.

8. The strip inspection method based on image data processing according to claim 1, characterized in that, In step S5, when calculating the energy offset vector between the characteristic matrix and the virtual residual distribution, the following steps are included: constructing a local search window between the characteristic matrix and the virtual residual distribution; performing cross-correlation operations within the local search window to determine the displacement trajectory of the energy centroid; and decomposing the displacement trajectory in polar coordinates to obtain the amplitude and phase information of the energy offset vector.

9. The strip detection method based on image data processing according to claim 1, characterized in that, When identifying a suspected target region as a non-target noise cluster in step S5, the following steps are also included: obtaining the spatiotemporal consistency constraints between adjacent multi-frame images; determining the morphological stability parameters of the suspected target region during the temporal evolution process; if the morphological stability parameters are lower than a preset stability threshold, then the suspected target region is determined to be transient high-frequency noise caused by fluid medium interference.

Images