An aluminum plate surface defect analysis method based on image multi-dimensional feature extraction

Abstract

The application provides an aluminum plate surface defect analysis method based on image multi-dimensional feature extraction, and relates to the technical field of computer vision and image data processing. The method acquires a multi-band image sequence set and a synchronous reference light environment parameter; extracts a basic scattering feature distribution graph, adopts a scattering segmentation threshold to intercept a local sequence and performs frequency domain mapping to construct a local physical deformation feature matrix; extracts an optical abnormality penalty weight based on topological coherence determination, performs weighted penalty aggregation calculation to output a surface damage probability measure; performs multi-level baseline verification and encapsulates a log to drive downstream execution state process shunting. The method realizes decoupling of physical geometric damage and optical interference and suppresses the false positive misjudgment rate.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image data processing technology, specifically to a method for analyzing surface defects of aluminum plates based on multidimensional image feature extraction. Background Technology

[0002] Industrial vision-based intelligent inspection technology has become a core component in ensuring the structural strength and ultimate fatigue life of aluminum plates, a critical substrate in aerospace and rail transportation industries. Under the highly complex conditions of modern production lines, automatic analysis systems based on multi-dimensional image feature extraction and state transitions undertake the fundamental and critical task of identifying quality risks from massive amounts of visual signals.

[0003] Existing visual inspection architectures still face challenging technical limitations when dealing with complex environments. Faced with complex boundary conditions such as sudden changes in localized lighting on the production line or residual protective oil film / lubricant dripping from equipment on aluminum plates, traditional systems relying on single-source two-dimensional grayscale feature mapping are prone to misinterpreting optical interference features of surface attachments (water stains, oil stains) as actual physical deformation damage, resulting in a persistently high false positive rate for the entire system. To address these limitations of single-modal perception, existing technologies (such as the defect detection method based on 3D reconstruction and transfer learning disclosed in patent document CN121095218B) attempt to introduce global-level 3D point cloud reconstruction and multi-band feature cascade networks. While this approach improves the comprehensiveness of recognition to some extent, global-level 3D reconstruction inevitably leads to high latency in cross-module communication of multimodal data, making it difficult to meet the core engineering requirements of real-time concurrent flow on the production line; furthermore, its over-reliance on manually annotated precise maps leaves room for improvement in adaptability to zero-sample conditions without observed variations. Summary of the Invention

[0004] The purpose of this invention is to provide a method for analyzing surface defects of aluminum plates based on multidimensional image feature extraction. This method introduces a dynamic masking mechanism driven by low-noise feedback and an optical anomaly penalty mechanism based on the spatial coherence of microscopic deformation to achieve high-fidelity damage feature extraction under optical noise interference conditions. Furthermore, it separates the true physical and geometric damage features of the surface from the optical interference features of surface deposits at the mathematical mapping level. This addresses the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for analyzing surface defects of aluminum plates based on multidimensional image feature extraction, comprising the following steps:

[0007] S1: Obtain a set of multi-band image sequences of the aluminum plate surface collected by the multi-source access node, and a reference illumination environment parameter synchronously associated with the set of multi-band image sequences;

[0008] S2: Based on the multi-band image sequence set and the reference illumination environment parameters, extract the basic scattering feature distribution map that characterizes the surface photometric dispersion.

[0009] In response to detecting that local values ​​in the basic scattering feature distribution map exceed a preset scattering segmentation threshold, a corresponding multidimensional image subset is extracted, and frequency domain mapping is performed on the multidimensional image subset to extract phase features, generating a local physical deformation feature matrix.

[0010] Based on the basic scattering feature distribution map, the local optical scattering interference weights corresponding to the coordinates are extracted. The local optical scattering interference weights and the local physical deformation feature matrix are then subjected to spatial mapping and weighted aggregation operations to output a surface damage probability measure.

[0011] S3: In response to the surface damage probability measure, the processor is triggered to perform the following control operation:

[0012] The surface damage probability measure is subjected to boundary verification with the preset multi-level baseline discrimination threshold. In response to the verification result crossing or residing in the corresponding state classification interval, a first state label representing the real physical deformation of the target area or a second state label representing optical interference is generated. The above state labels are packaged and encapsulated into classification log data for updating the production line status monitoring screen.

[0013] Based on the classification log data, a process diversion instruction is generated and sent to the downstream production line control node through the cross-module control signaling channel. The process diversion instruction is configured to drive the production line control node to perform corresponding status classification actions on the corresponding batch of aluminum plates.

[0014] Compared with the prior art, the beneficial effects of the present invention are:

[0015] This invention, after extracting the local physical deformation feature matrix, constructs a defect contour topological coherence feature matrix by extracting the cross-band surface texture distortion gradient between adjacent nodes. It then performs a nonlinear mapping and weighted penalty aggregation on the local optical scattering interference weights and this matrix. Based on spatial coherence logic, it distinguishes between discrete, randomly distributed optical artifacts and continuously oriented, real deformations, decoupling physical geometric damage features from the optical interference features of surface attachments. This addresses the pain point of false positives caused by water stains and oil films in complex production lines.

[0016] A dynamic adjustment mechanism based on environmental noise feedback is introduced. This mechanism maps the fundamental optical fluctuation variance to a global background noise factor and dynamically generates a scattering segmentation threshold by combining it with a dynamic penalty offset. When the intensity of macroscopic ambient light interference increases, this scattering segmentation threshold is adjusted upwards accordingly, thereby pre-shielding pseudo-random optical abrupt changes during logical preprocessing. This mechanism effectively intercepts redundant noise from entering the multi-band joint physical deformation feature extraction network while ensuring the sensitivity of deep feature mining.

[0017] This invention performs boundary verification between the output surface damage probability measure and a preset multi-level baseline discrimination threshold, automatically generating a first state marker representing real physical deformation or a second state marker representing optical interference, and further encapsulates this into classification log data to issue process diversion instructions. This control flow method establishes an instruction-driven closed loop from the front-end visual perception engine to the back-end production line control nodes, enabling downstream equipment to quickly respond to classification features and execute asynchronous diversion actions, achieving collaborative scheduling of the production line. Attached Figure Description

[0018] Figure 1 This is a diagram of the collaborative topology architecture of the present invention;

[0019] Figure 2 This is a flowchart illustrating the execution logic of steps S1 to S2 of the present invention.

[0020] Figure 3 This is a flowchart illustrating the execution logic of step S3 of the present invention.

[0021] Figure 4 A schematic diagram showing the results of a numerical verification experiment on the spatial coherence weighted penalty aggregation mechanism;

[0022] Figure 5 This is a comparison diagram of the heterogeneous physical response of the multi-band initial image data stream on the aluminum plate surface of the present invention.

[0023] Figure 6 This is a schematic diagram of the macroscopic architecture and data flow of a basic optical feature extraction network based on a multi-scale convolutional autoencoder. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0026] Example 1:

[0027] Please see Figures 1 to 6 The present invention provides a technical solution:

[0028] A method for analyzing surface defects of aluminum plates based on multidimensional image feature extraction, executed by a processor of a computing device, includes the following steps:

[0029] This embodiment constructs a "dynamic masking mechanism based on noise floor feedback" and a "weighted penalty aggregation mechanism based on spatial coherence." Specifically, a basic optical feature extraction network is pre-built and calibrated. The specific construction process is as follows: A historical benchmark dataset containing multiple sets of original multi-band reflection data of defect-free aluminum plates is obtained, along with their corresponding ideal flat state real business labels. Static blacklist filtering rules are enforced at the data access layer to completely remove any abnormal samples containing optical artifacts or substantial physical damage features from the input parameter queue in this initial optimization training phase. A lightweight multi-scale convolutional autoencoder (Multi-scaleCNN-Autoencoder) is configured as the initial model network, and during the training phase, multiple... The band input features (pre-processed with histogram equalization and outlier removal through data access nodes) are input into the encoding layer of the initial model network; the model layers are mapped through a multi-dimensional feature space to output a reconstruction prediction matrix; the joint loss function is used to maximize the log-likelihood distribution of normal specular reflectance and minimize the root mean square reconstruction error (RMSE); the convolutional kernel weights and bias parameters within the model are iteratively adjusted using gradient descent and backpropagation algorithms; the fluctuating gradient on the validation set is continuously monitored, and when the convergence descent gradient of the loss function is less than a preset stability threshold of 10 for 10 consecutive iterations... -5 At this time, an early stopping mechanism is triggered to prevent overfitting, and all parameters of the current network layer are fixed, thus outputting the basic optical feature extraction network deployed to the online environment for real-time inference. For the subsequently extracted multi-band local pixel reflection feature values, the absolute amplitude of the response value is extracted, and the topological direction of the gradient is evaluated by constructing a defect contour topological coherence feature matrix. Since real mechanical cutting marks have a continuous and oriented physical distribution pattern in space, while the optical interference phase caused by residual oil is random and discrete, the consistency state is determined by logical comparison based on this mechanism. A high-order penalty is triggered for random distribution, and a low-order penalty is triggered for oriented distribution to maintain high weight, thereby achieving the distinction between physical geometric damage features and optical interference features of surface attachments.

[0030] S1: Obtain a set of multi-band image sequences of the aluminum plate surface collected by the multi-source access node, and a reference illumination environment parameter synchronously associated with the set of multi-band image sequences;

[0031] Further implementation of step S1: The acquisition of a set of multi-band image sequences of the aluminum plate surface collected by the multi-source access node includes:

[0032] The multi-source access node, configured as an external camera, performs external data acquisition; acquires an initial image data stream containing multiple multi-band image acquisition channels; performs time-series alignment on the initial image data stream based on a preset timestamp alignment rule to eliminate time-series deviations caused by inconsistent sampling frequencies of the multi-band image acquisition channels, generating a time-series aligned sequence; and performs background noise filtering on the time-series aligned sequence to construct the multi-band image sequence set.

[0033] Further implementation of step S1: Obtain the reference illumination environment parameters that are synchronously associated with the multi-band image sequence set, including:

[0034] Obtain the operation status log of the lighting control unit associated with the multi-source access node; parse the operation status log to extract the light source radiation intensity parameters and light source spatial illumination attitude parameters corresponding one-to-one with the multi-band image acquisition channels; fuse the light source radiation intensity parameters and the light source spatial illumination attitude parameters to generate the reference illumination environment parameters, which serve as a reference baseline for evaluating the intensity of surface emissivity disorder.

[0035] For step S1, the multi-source data sensing access and time-series-environment baseline joint alignment steps are performed. Specifically, at the beginning of the online detection task, the processor of the computing device enters the multi-source data sensing stage. The initial image data stream, containing multiple multi-band image acquisition channels, is defined as the original data sequence with heterogeneous spectral characteristics generated by the multispectral array camera of the multi-source access node within the synchronization trigger period. "Multiple multi-bands" includes at least the following physical band channels:

[0036] Ultraviolet (UV) reflection channel (wavelength range preferably 365nm-400nm): Utilizing the high sensitivity of ultraviolet light to surface micro-undulations, it is specifically used to extract extremely shallow scratches and minute physical deformation features on the surface of aluminum plates.

[0037] Visible light (VIS) monochromatic channel (preferably in the 530nm-550nm green light band): serving as a reference geometric contour channel to obtain the macroscopic flatness and background texture of the aluminum plate surface.

[0038] Near-infrared (NIR) penetration channel (wavelength range preferably 850nm-940nm): Utilizing the low absorption and reflection characteristics of infrared light on liquid deposits, it is specifically used to penetrate residual oil films or water stains to obtain the true substrate characteristics under the cover.

[0039] Each initial image data stream is encapsulated into a standardized asynchronous sensing data packet at the transport protocol layer. The packet structure consists of a "packet header" and a "pixel payload": the packet header contains a unique channel identifier, a physical sampling frequency calibration value, and a crucial high-precision clock stamp. The clock stamp is generated by the local clock of the multi-source access node at the moment of exposure to support subsequent timing alignment. The pixel payload contains uncompressed raw photosensitive count values, preferably 12-bit or 16-bit grayscale response values ​​in this embodiment, to ensure that the image sequence set has sufficient resolution dynamic range to support subsequent calculations of local photometric distortion.

[0040] The multi-source access node of the external camera represents a heterogeneous sensing cluster composed of multiple high-speed industrial imaging units with microsecond-level synchronous shutter mechanisms and spectral filtering components. In this embodiment, the sensing cluster is configured above the target aluminum plate, with three independent optical sampling branches distributed at preset spatial angles (15° to 30°), corresponding to the ultraviolet (UV) reflection channel, the visible light (VIS) monochromatic channel, and the near-infrared (NIR) transmission channel, respectively. The microscopic settings of each sampling branch are as follows: the ultraviolet branch has a narrow-band bandpass filter with a center wavelength of 385nm (bandwidth ±10mm) mounted at the front end of the industrial lens to capture the microscopic phase transitions on the surface of the aluminum plate; the visible light branch has a bandpass filter with a center wavelength of 540nm to establish a reference image of the macroscopic geometric contour; and the near-infrared branch has a long-wavelength pass filter with a center wavelength of 940nm to utilize the long-wavelength penetration characteristics of infrared light to remove interference from the surface liquid. Each imaging branch performs unified pulse signal alignment with the illumination control unit through a common external hard trigger bus, ensuring multi-spectral coordinated exposure of the same physical area of ​​the target aluminum plate within a timing tolerance of ±5 microseconds, thereby constructing a high-confidence initial image data stream.

[0041] For image stream timing alignment, the initial image data stream, which includes multiple multi-band image acquisition channels, is acquired by multiple source access nodes. A preset timestamp alignment rule is triggered to execute a global timing synchronization operation. The specific data stream control steps are as follows: The protocol encapsulation of the initial image data stream of each channel is stripped, and the high-precision clock stamp identifier embedded in the data packet frame header is parsed and extracted; a timing tolerance monitoring window of a preset width (preferably ±5 microsecond jitter margin in this embodiment) is opened in a preset logical monitoring queue; the backbone frequency band channel is selected as the time reference axis, and the clock stamp sequences of other subordinate multi-band channels are input into the nearest neighbor time comparison link; in response to the timestamp of a subordinate frame falling within the tolerance monitoring window of the backbone reference frame, it is determined that they belong to the same physical sampling period and spatial tensor concatenation is performed; in response to the detection of timing holes exceeding the tolerance window, features of two adjacent frames are extracted and linear weighted time domain interpolation compensation is performed to forcibly eliminate timing deviations, and finally, a timing alignment sequence with absolute synchronization in all dimensions is output. The processor invokes a local Gaussian smoothing filter algorithm to perform background noise removal on the time-aligned sequence. During this filtering process, the spatial convolution kernel size used to control the denoising dimension is preferably configured as a 3×3 pixel array, and its standard deviation constant... The configuration is set to 1.2. This ensures that while filtering out high-frequency hardware-acquired noise pulses, the macroscopic geometric gradient response of the weak physical deformation on the aluminum plate surface can be preserved; thereby constructing a multi-band image sequence set with time synchronization.

[0042] While acquiring the image stream, the system also acquires the operating status logs of the lighting control unit associated with the multi-source access node, as well as the situational fluctuation signals collected in real time by environmental photoelectric sensors pre-deployed around the target detection station.

[0043] The lighting control unit represents the core microcontroller module that performs dynamic modulation of multi-band light source energy and hardware-level timing synchronization. The lighting control unit integrates a high-speed microprocessor, multiple constant current drive circuits, and a communication interface, receiving control signals from the computing device via a communication bus. The lighting control unit is responsible for performing pulse-width modulation (PWM) current control on heterogeneous wavelength light sources (UV, VIS, and NIR) to regulate transient radiant flux; simultaneously, it receives trigger pulses from position sensors via a common external hard trigger bus, ensuring absolute step alignment between the light source's excitation period and the external camera's shutter exposure window within microsecond tolerances. The lighting control unit polls in real-time for data from the spatial encoder connected to the light source gimbal and current feedback samples from the drive circuit, encapsulating these characteristic sequences reflecting the physical properties of the light source into an operational status log containing high-precision clock stamps. This provides a traceable hardware feature base for subsequent analysis of "light source radiant intensity parameters" and "light source spatial illumination attitude parameters." The operational status log contains the underlying hardware drive level sequence bound to the high-precision clock stamps and the spatial gimbal encoder array data. The system analyzes the running status logs to extract the corresponding light source radiation intensity parameters and light source spatial illumination attitude parameters for each channel of the multi-band image acquisition.

[0044] During this extraction process, the light source radiation intensity parameter (denoted as...) It is defined as the absolute radiative flux of the corresponding acquisition channel at the moment of triggering exposure microseconds. Its specific extraction logic is as follows: parse the pulse width modulation (PWM) duty cycle reading and the real-time drive current negative feedback characteristic value in the operation status log, and linearly project them into a normalized scalar value that objectively represents the transient energy intensity based on a pre-fixed photoelectric conversion mapping table.

[0045] Light source spatial illumination attitude parameters (denoted as) The is defined as representing the geometric spatial topological relationship between the emitting end face of the light source and the physical detection plane of the target aluminum plate. Its specific extraction logic is as follows: parse the spatial gimbal encoder array readings in the operation status log, extract data containing three dimensions including the incident zenith angle of the light source, the spatial azimuth angle, and the relative focal projection distance, and combine and encapsulate them into a three-dimensional direction vector feature.

[0046] Light source radiation intensity parameters Spatial illumination attitude parameters of the light source The data is input into a pre-defined physical photometric projection engine for fusion and construction. The specific logic is as follows: Extract the incident zenith angle component of the light source from the 3D direction vector features, calculate the cosine eigenvalue of this zenith angle component; and normalize the scalar values ​​of the light source radiance parameters. A scalar multiplication operation is performed with the calculated cosine eigenvalue to extract the attenuated effective vertical surface illuminance projection component. The effective vertical surface illuminance projection component is then divided by the square of the relative focal projection distance to perform spatial dissipation compensation correction. This extracts the transient theoretical illuminance absolute value with absolute physical dimensions. A normalization mapping is performed on the transient theoretical illuminance absolute value: a historically calibrated illuminance limit compensation upper limit constant (representing the theoretical maximum illuminance projection value that the current device can generate at its closest distance under full load and no attenuation conditions) is retrieved from the configuration node beforehand. The currently calculated transient theoretical illuminance absolute value is divided and converted by the pre-stored illuminance limit compensation upper limit constant to obtain a relative intensity scalar with percentage attributes. An extreme value truncation operation is performed on the relative intensity scalar: in response to a relative intensity scalar greater than the unit value 1, it is forcibly truncated to 1; in response to a relative intensity scalar normally falling within the [0,1] interval, it is retained. The closed-loop numerical sequence of [0,1] after the above proportional conversion and boundary truncation is used as the reference lighting environment parameter to characterize the intensity of surface reflectivity disorder.

[0047] Table 0: Examples of photoelectric conversion mapping representations:

[0048]

[0049] In the real-time visual inspection task flow of aluminum plates, after receiving the running status log of the current trigger frame, the processor extracts the real-time PWM duty cycle reading (denoted as ). ) and the real-time drive current negative feedback characteristic value (denoted as ).by and To jointly retrieve coordinates, the four nearest reference calibration nodes are located within the discrete grid of the photoelectric conversion mapping table. A bilinear interpolation algorithm is then used to perform weighted continuous analysis on the reference scalar values ​​of these four adjacent nodes. This mechanism can detect and eliminate subtle luminance drift caused by thermal decay of the light source device, power supply ripple, or physical aging, ensuring the final output light source radiance parameters are accurate. It is a scalar for characterizing transient energy.

[0050] S2: Based on the multi-band image sequence set and the reference illumination environment parameters, extract the basic scattering feature distribution map that characterizes the surface photometric dispersion.

[0051] In response to detecting that local values ​​in the basic scattering feature distribution map exceed a preset scattering segmentation threshold, a corresponding multidimensional image subset is extracted, and frequency domain mapping is performed on the multidimensional image subset to extract phase features, generating a local physical deformation feature matrix.

[0052] Based on the basic scattering feature distribution map, the local optical scattering interference weights corresponding to the coordinates are extracted. The local optical scattering interference weights and the local physical deformation feature matrix are then subjected to spatial mapping and weighted aggregation operations to output a surface damage probability measure.

[0053] Step S2 is further explained as follows: Based on the multi-band image sequence set and the reference illumination environment parameters, a basic scattering feature distribution map characterizing the surface photometric dispersion is extracted, and a local physical deformation feature matrix is ​​generated, including:

[0054] The multi-band image sequence set and the reference illumination environment parameters are input into a pre-configured basic optical feature extraction network to calculate the local photometric disorder degree that deviates from the ideal flat state in the global space and generate the basic scattering feature distribution map.

[0055] A dynamically generated scattering segmentation threshold is used to perform connected component segmentation matching on the basic scattering feature distribution map to extract the coordinate set of regions with abrupt changes in luminance; the coordinate set of regions with abrupt changes in luminance is converted into a region of interest extraction mask, and a multidimensional image subset is extracted from the multi-band image sequence set using the region of interest extraction mask;

[0056] The multidimensional image subset is input into a pre-configured multi-band joint physical deformation feature extraction network to extract surface physical deformation gradient features across band boundaries and construct the local physical deformation feature matrix.

[0057] Step S2 is further explained as follows: Based on the basic scattering feature distribution map, local optical scattering interference weights corresponding to the coordinates are extracted. Spatial mapping and weighted aggregation operations are performed on the local optical scattering interference weights and the local physical deformation feature matrix to output a surface damage probability measure, including:

[0058] Based on a unified spatial reference feature grid, projection alignment resampling is performed on the basic scattering feature distribution map with different feature scales and the local physical deformation feature matrix; the local normalized photometric disorder is extracted from the aligned basic scattering feature distribution map as the local optical scattering interference weight;

[0059] The local optical scattering interference weights are subjected to element-wise nonlinear mapping and weighted penalty aggregation calculation with the multi-band local pixel reflection feature values ​​in the aligned local physical deformation feature matrix; based on the calculation results, a surface damage probability measure characterizing the true physical damage degree of the surface is extracted.

[0060] Step S2 is further explained as follows: Connectivity segmentation matching is performed on the basic scattering feature distribution map using a dynamically generated scattering segmentation threshold, including:

[0061] Obtain the basic optical fluctuation variance of the surface of the current batch of aluminum plates within a preset background area, and map the basic optical fluctuation variance into a global background noise factor that characterizes the intensity of the current ambient light interference.

[0062] Based on a preset threshold adaptive adjustment curve, the global background noise factor is converted into a dynamic penalty offset.

[0063] The preset baseline sensitivity threshold and the dynamic penalty offset are aggregated to dynamically generate the scattering segmentation threshold;

[0064] In response to the condition that the value of a local grid in the basic scattering feature distribution map is greater than the scattering segmentation threshold, the local grid is determined to be a suspected defective pixel connected region, and the continuous suspected defective pixel connected regions are aggregated to generate the coordinate set of the photometric abrupt change region; wherein, when the global background noise factor, which represents the ambient light interference intensity, increases, the scattering segmentation threshold is synchronously increased to shield pseudo-random optical abrupt change signals from entering the subsequent physical deformation feature extraction network.

[0065] Step S2 is further explained as follows: The local optical scattering interference weights and the multi-band local pixel reflection feature values ​​in the aligned local physical deformation feature matrix are subjected to element-wise nonlinear mapping and weighted penalty aggregation calculation, including:

[0066] Traverse each feature node in the aligned local physical deformation feature matrix, extract the cross-band surface texture distortion gradient between adjacent feature nodes, and construct a defect contour topological coherence feature matrix to characterize the spatial coherence of micro-deformation.

[0067] Based on the preset surface deformation topological coherence comparison logic, the topological coherence feature matrix of the defect contour is logically compared: in response to the condition that the topological coherence feature matrix of the defect contour in the target area is discretely randomized, a hard degradation strategy is triggered, the original calculated value of the local optical scattering interference weight under the corresponding coordinates is discarded, and its state pointer is forcibly replaced and configured with the preset random optical noise strong attenuation coefficient; in response to the condition that the topological coherence feature matrix of the defect contour in the target area is continuously oriented, the confidence of the true deformation is maintained, and the parameter flow node under the corresponding coordinates is directly assigned the preset true deformation feature weak penalty coefficient.

[0068] The random optical noise strong attenuation coefficient or the real deformation feature weak penalty coefficient is used to perform element-wise scalar multiplication attenuation calculation on the multi-band local pixel reflection feature value to obtain the pure damage feature set after penalty correction.

[0069] The set of pure damage features within the target area is subjected to nonlinear feature activation and boundary truncation based on the hyperbolic tangent function. Subsequently, a spatial weighted aggregation summation operation is performed by fusing locally preset spatial distribution weight factors to generate the surface damage probability measure.

[0070] Step S2 involves basic scattering feature extraction and dynamic mask adaptive truncation based on noise floor feedback, specifically including:

[0071] The preliminary basic scattering feature mapping step involves synchronously inputting the multi-band image sequence set generated in step S1 and the reference illumination environment parameters into the basic optical feature extraction network that has already undergone offline calibration. Based on the mapping relationship, the basic optical feature extraction network calculates the local photometric disorder degree deviating from the ideal flat state in the global space, generating a global basic scattering feature distribution map. The specific implementation steps and logical definitions are as follows:

[0072] like Figure 6 As shown, the multi-band image sequence set (denoted as...) The reference lighting environment parameters generated in the aforementioned steps (denoted as...) The dimensions are concatenated to construct a joint input tensor. The baseline illumination environment parameter constraint network determines the expected reflectivity of the aluminum plate surface under different radiation intensities and illumination attitudes. The basic optical feature extraction network receives the joint input tensor and extracts cross-scale global spatial features through its encoding layer. The decoding layer outputs a reconstruction prediction matrix (denoted as ) with the same dimensions as the input. The output reconstruction prediction matrix. Defined as: parameters under the current reference lighting environment Below is a theoretical reflectivity distribution map that the aluminum plate surface in this region should exhibit when it is in a "perfectly defect-free and ideally flat state". The original multi-band image sequence set was obtained. With the reconstruction prediction matrix It performs pixel-by-pixel nonlinear difference operations to extract the residual distribution representing the deviation of actual observed values ​​from theoretically predicted values, and defines the operation result as the local photometric disorder (denoted as ). The specific calculation logic of nonlinear difference operation is as follows: traverse each pixel node in the global coordinate system of the image, and extract the corresponding pixel in the multi-band image sequence set. The original reflection vector in the reconstructed prediction matrix and The corresponding predicted reflection vector is used; a vector subtraction operation is performed between the original reflection vector and the predicted reflection vector to extract the difference vector; the L2 norm (square of the modulus) of the difference vector is calculated and used as the quantization response value of the pixel's deviation from the ideal flat state. The local photometric disorder obtained from the above calculations is then used. Mapped to the global spatial coordinate system, a basic scattering feature distribution map is generated (denoted as...). ).

[0073] The basic scattering characteristic distribution map is defined as: a signal-to-noise ratio enhanced characteristic map that removes ambient light fluctuation interference and retains only transient anomalies in surface reflectivity. This is achieved by introducing a reference illumination environment parameter. As a reconstruction constraint, a deep decoupling of "lighting and texture" is achieved. If the aluminum plate surface only has uneven lighting but is physically flat, then the multi-band image sequence set... With reconstructed prediction matrix The difference will approach zero, resulting in local luminosity disorder. Maintaining a low baseline avoids misjudgments caused by sudden changes in ambient light. If there are actual scratches or attachments on the aluminum plate surface, its scattering characteristics will prevent the network from accurately reconstructing the ideal reflectivity at that location, thus affecting the basic scattering characteristic distribution map. This creates a high-contrast local luminous disorder.

[0074] Global background noise factor, denoted as It characterizes the basic optical fluctuation variance of the surface of the current batch of aluminum plates within a preset defect-free background area. In this embodiment, it serves as an adaptive adjustment feedback source to objectively quantify the interference intensity of transient ambient light (sudden changes in external workshop lighting) on ​​the visual perception link.

[0075] Specifically, based on a pre-configured morphological isolation template, a preset background region without feature texture is defined from the basic scattering feature distribution map of the current batch. A structuring element matrix of a preset size (preferably 15×15 pixels) is retrieved as the morphological isolation template, and morphological opening operations of erosion and dilation are sequentially performed on the basic scattering feature distribution map to filter out high-frequency scattered patch noise. Based on a preset connected component retrieval algorithm, the largest closed-loop connected components in the map whose local gradient mean is lower than the minimum texture threshold (preferably 0.01) are extracted, and their coordinate system is mapped back to the original matrix, thereby objectively establishing the preset background region without feature texture. A feature sampling grid of a preset size (preferably 5×5 pixels) is called, and the scattering values ​​in the background region are continuously traversed with a preset stride. Discreteness statistics are performed to extract the local standard deviation features in each local receptive field. Global mean aggregation calculation is performed on all extracted local standard deviation features to obtain the basic optical fluctuation variance.

[0076] The basic optical fluctuation variance is input into the normalization mapping module, and extreme value normalization is performed on it based on a preset linear mapping rule. The specific mapping calculation logic is shown in the following equation:

[0077]

[0078] in The absolute physical value representing the variance of the underlying optical fluctuations currently extracted from the preset background region; The minimum value of the underlying optical fluctuation, pre-calibrated from a historical benchmark defect-free dataset, is preferably 0.02 in this embodiment; The maximum value of the basic optical fluctuation under the tolerance boundary of the system is preferably 1.85 in this embodiment. The fluctuation variance value with the original hardware sensing device dimension is linearly scaled proportionally to the standard dimensionless interval [0,1], thereby extracting the global background noise factor. The preset threshold adaptive adjustment curve is retrieved to perform nonlinear feature transformation. The logic of the threshold adaptive adjustment curve is as follows: extract the global background noise factor. As a fundamental parameter, it is multiplied by the environmental sensitivity control constant (preferably 1.5) to obtain the activation feature number; a power function operation is performed with the natural constant e as the base and the activation feature number as the exponent to generate a nonlinear curvature mapping value characterizing the intensity of the interference response; the nonlinear curvature mapping value is weighted and summed with the system environmental background lower limit constant to generate an initial mapping reference value, denoted as . The initial mapping baseline value represents the absolute physical fluctuation offset that has not been constrained by engineering boundaries.

[0079] For the initial mapping reference value Perform linear truncation and proportional scaling operations; the specific logic is as follows: First, extract and lock the theoretical minimum and maximum values ​​of the initial mapping baseline value derived from logical deduction; then, perform a subtraction operation between the currently calculated initial mapping baseline value and the theoretical minimum value to obtain the first difference response value; simultaneously, perform a subtraction operation between the theoretical maximum and the theoretical minimum value to obtain the second difference response value; divide the first difference response value by the second difference response value and perform proportional scaling to obtain the normalization coefficient; extract the preset target safety mapping upper and lower limits, and project the normalization coefficient proportionally into the closed-loop interval formed by the target safety mapping upper and lower limits to obtain the scaled final output value, which is used as the dynamic penalty offset, denoted as... The system environment background lower limit constant characterizes the sensor's inherent thermal noise floor under conditions without external sudden strong light interference. Based on historical benchmark operating conditions, the preferred value of the system environment background lower limit constant in this embodiment is 0.05. In practical applications, it is dynamically calibrated within the dimensionless range of [0.02, 0.08] according to the equipment accuracy. Dynamic penalty offset. The global threshold compensation scale applied to suppress pseudo-random optical noise is characterized. In this embodiment, the upper and lower bounds of the target security mapping in the aforementioned steps of extreme value normalization mapping are defined as a dimensionless interval of [0.15, 0.45].

[0080] The baseline sensitivity threshold is retrieved. This threshold characterizes the photometric anomaly cutoff line under ideal experimental conditions (no abnormally strong direct light). Based on historical statistics, its optimal quantized value is set to an objective constant of 0.55. Subsequently, the baseline sensitivity threshold is used as a fixed benchmark to apply the dynamically calculated penalty offset. As a dynamic compensation term, both are input into the differential compensation engine to perform a dimensionless scalar addition operation. This is based on the dynamic penalty offset. Having been normalized and constrained to the interval [0.15, 0.45], the output of the above scalar addition operation is naturally and absolutely anchored within the normalized closed-loop interval [0.70, 1]. This summation value residing within the safety boundary is used as the real-time updated scattering segmentation threshold, denoted as... .

[0081] A connected component segmentation matching logic comparison is performed on the basic scattering feature distribution map using a scattering segmentation threshold. This occurs in response to the detection of a local grid value exceeding the scattering segmentation threshold. The condition is used to determine if the local region deviates from the tolerance boundary of the ideal flat state, triggering the extraction task of a multi-dimensional image subset. The specific extraction logic is as follows:

[0082] The coordinates of all pixels whose values ​​exceed the scattering segmentation threshold are located and aggregated to generate a set of coordinates for photometric abrupt change regions. This set is defined as the spatial index carrying the candidate features of defects. A preset mask generation operator is invoked to initialize a background matrix with all zeros that is absolutely consistent with the spatial size of the original image. The set of coordinates for photometric abrupt change regions is traversed, and the feature nodes at the index positions corresponding to the coordinates in the all-zero background matrix are assigned logical retention flags (binarized scalar 1), while the remaining unselected background regions are maintained as logical blocking flags (binarized scalar 0). The mapping of the two-dimensional matrix is ​​completed through the above discrete assignment rules to generate the region of interest extraction mask, which represents the hard geometric constraint boundary of the system on the suspected defect candidate regions in the spatial dimension. Subsequently, the region of interest extraction mask is subjected to a pixel-by-pixel bitwise logical AND operation with the multi-band image sequence set in step S1. During this logical filtering process, the ineffective regions (background regions) carrying logical blocking flags in the mask are physically shielded due to the multiplication by zero property, and only the local high-dimensional feature data under the mask coverage is retained proportionally.

[0083] Finally, only local data blocks highly aligned with the mutation regions in space, band, and time are extracted from the original sequence, outputting a multidimensional image subset. This is done when the global background noise factor... Characterizing the scattering segmentation threshold as ambient light interference intensity increases Synchronous upregulation occurs, thereby shielding pseudo-random optical mutation signals from entering subsequent steps in the early stages.

[0084] In this embodiment, before performing online inference of the multi-band joint physical deformation feature extraction network, the network instance is pre-built and calibrated in an offline environment. The specific calibration mechanism is as follows: the initial joint physical deformation feature extraction network is configured using a dual-stream feature pyramid architecture (Dual-stream-FPN) to ensure the smooth cascading of heterogeneous band features in multi-scale space. A multi-band baseline image dataset is acquired, which includes a subset of surface defect images simultaneously acquired in the infrared and visible light bands, and real physical deformation topology labels generated by a high-precision 3D profilometer. This subset of surface defect images is then input into an initial joint physical deformation feature extraction network to perform cross-modal feature fusion. A joint loss function is configured in the backpropagation stage, representing the geometric deviation constraint between the predicted phase space output by the model and the real physical micro-topology. The joint loss function is defined as the mean square error loss between the multi-band reconstructed phase features and the real physical deformation topology labels, combined with a total variational regularization term representing gradient continuity. Based on this joint loss function, the spatial convolution kernel parameters within the network are iteratively adjusted. When the joint loss gradient on the validation set is continuously lower than the extreme convergence constant (10^6), the network continues to operate. A predetermined number of training epochs (preferably 15 epochs) is then established. -4 When the model reaches a global fit state, the current feature space mapping weights are solidified and deployed as a multi-band joint physical deformation feature extraction network to the online production environment.

[0085] Step S2 further executes the spatial coherence penalty aggregation and surface damage probability extraction steps: after completing the dynamic mask truncation, it enters the deep feature decoupling stage aimed at removing optical noise interference.

[0086] The extracted multidimensional image subset is input into a pre-configured multi-band joint physical deformation feature extraction network. This network extracts surface physical deformation gradient features across illumination boundaries of different bands based on a frequency domain mapping algorithm. Frequency domain mapping and gradient extraction are performed on the multidimensional image subset to construct a local physical deformation feature matrix (denoted as...). The specific execution steps are as follows:

[0087] A subset of multidimensional images is acquired, and a two-dimensional fast Fourier transform (2D-FFT) algorithm is used to map the local pixel spatial distribution of each band to the frequency domain to extract the corresponding complex spectrum matrix. Based on a preset bandpass filter window, effective frequency domain components containing microscopic physical deformation carrier characteristics are extracted from the complex spectrum matrix to filter out low-frequency ambient light and high-frequency white noise. A two-dimensional inverse fast Fourier transform (2D-IFFT) algorithm is then used to reverse-map and reconstruct the effective frequency domain components to the spatial domain to generate a spatial domain complex matrix. Phase calculation is performed on the spatial domain complex matrix to obtain an initial set of phase layers representing the absolute mapping of the surface micro-topology in the spatial coordinate system. Cross-band surface physical deformation gradient features are introduced, and their extraction logic is as follows: a reference band phase map is selected from the multi-band array, and a pixel-by-pixel phase difference operation is performed between it and the phase maps of the remaining subordinate bands to obtain a phase difference response matrix reflecting local topological depth transitions. By performing spatial first-order partial derivative operations on the phase difference response matrix, the directional gradient vector characterizing the severity of physical deformation is extracted, thus completing the quantitative extraction of surface physical deformation gradient features across waveband boundaries. A local physical deformation feature matrix is ​​then constructed based on the extracted surface physical deformation gradient features. Its internal calculation logic follows these steps: perform magnitude aggregation calculation on gradient vectors in each direction to generate a gradient energy map; call the hyperbolic tangent function tanh to perform nonlinear activation and truncation compression on the scalar values ​​in the gradient energy map. Since the gradient magnitude is always non-negative, this mapping operation projects its absolute value into a unified physical deformation confidence interval [0,1); use the compressed confidence values ​​as matrix elements, arrange and encapsulate them according to the original pixel coordinate index, and finally output the local physical deformation feature matrix.

[0088] Based on a unified spatial reference feature grid processor, projection alignment and resampling operations are performed on the distribution maps of basic scattering features with different feature scales and the local physical deformation feature matrix to ensure the absolute mapping correlation of heterogeneous features in the spatial coordinate system. The locally normalized photometric disorder is extracted from the aligned scrambled matrix as the local optical scattering interference weight (denoted as...). The derivation logic of this feature normalization flow mechanism is as follows: based on the node-by-node traversal rule, scan the feature mesh of the target local region; statistically analyze and extract the local photometric disorder generated in the previous steps within the global view of the entire connected domain. The parameter at the maximum extreme point in the current region (denoted as ) ) and the parameters of the minimum and extreme points (denoted as ); Locate the current target coordinate node and extract the absolute value of the local photometric disorder corresponding to that single node (denoted as ). This represents the absolute physical dimension of the deviation of the micro-pixel pointed to by the current transient traversal pointer from the reference flat state; after extracting the above reference parameters, the absolute value of the local photometric disorder is... With minimum and extreme point parameters The absolute difference is extracted by subtraction and then divided by the range span of the local region (the difference between the parameters at the maximum and minimum points) to perform linear mapping scaling. Finally, the absolute photometric values ​​with uncertain physical dimensions are projected as continuous probability distribution coefficients within the [0,1] closed-loop interval, thereby generating local optical scattering interference weights with standard attenuation properties. .

[0089] Further retrieve the topological coherence feature matrix of the defect contour (denoted as...). ), and perform a comparison to prevent false positives:

[0090] The defect contour topological coherence feature matrix characterizes the cross-band surface texture distortion gradient of micro-deformation between adjacent spatial feature grids. This multidimensional feature set constitutes a three-dimensional tensor with direction vectors, and the mapping system is currently in the "recognition" stage. The process involves a transition between "real physical destruction with continuous topology" and "identification of random, discrete optical artifacts." Specifically, a topological feature extraction operator with eight neighborhood sampling directions is pre-loaded; the processor, using the topological feature extraction operator as its core component, traverses the aligned local physical deformation feature matrix based on a pixel-by-pixel sliding step rule. Each feature node in the model is analyzed; at each sliding anchor point, the multi-band frequency domain phase difference response values ​​of the central feature node and its eight adjacent feature nodes in the corresponding direction are extracted to quantize and extract the cross-band surface texture distortion gradient between adjacent feature nodes; based on the vector direction and magnitude of the distortion gradient, feature splicing and dimensional expansion are performed to converge and reorganize the scalar gradient group of the local plane into a feature tensor set with three-dimensional direction vector features, and finally construct the defect contour topological coherence feature matrix for transmitting state commands downstream. .

[0091] The strong attenuation coefficient of random optical noise is denoted as: The characteristic of this is a strong blocking factor for random isolated noise points. In this embodiment, its preferred value is set to the range of [0.01, 0.05]. The principal component extraction operation of the feature direction is performed on the constructed defect contour topological coherence feature matrix to obtain the maximum and second largest eigenvalues ​​in its three-dimensional direction vector, and the distribution ratio coefficients of the maximum and second largest eigenvalues ​​are calculated. The distribution ratio coefficients are compared with a preset orientation consistency threshold (preferably 3.5): In response to the condition that the distribution ratio coefficient is less than or equal to the orientation consistency threshold, it is determined that the defect contour topological coherence feature matrix in the target area exhibits a discrete random distribution. At this time, a hard degradation strategy is triggered, and the original calculated value of the local optical scattering interference weight under the coordinates corresponding to the target feature node is directly discarded at the physical level. The target attenuation control factor is configured to the preset strong attenuation coefficient of random optical noise. .

[0092] In response to the condition that the distribution proportion coefficient is greater than the orientation consistency threshold, it is determined that it exhibits a continuous orientation distribution (example representing a real mechanical scratch). The target attenuation control factor is mapped to a weak penalty coefficient for the real deformation characteristics, denoted as . The preferred value is [0.85, 1.0], which aims to retain the high weight of real defects.

[0093] Strong attenuation coefficient of random optical noise extracted using dynamic mapping Or a weak penalty coefficient for true deformation characteristics (In this embodiment, it is collectively referred to as the target attenuation control factor, denoted as...) The numerical processing logic is applied to the multi-band local pixel reflection feature values. The derivation mechanism for feature dimension compression and spatial weighted aggregation calculation is shown in the following equation:

[0094] Here, variable N represents the total number of physical pixels of effective spatial feature grid nodes contained within the extraction mask (i.e., the locally connected region of optical abrupt change) of the region of interest extracted in the current batch. N is a discrete positive integer that fluctuates in real time with the area captured by the dynamic mask. The multi-band local pixel reflection feature values ​​(denoted as ) corresponding to multiple effective spatial feature grid nodes i within the target region are obtained. ), and compare it with the target attenuation control factor Perform element-wise scalar multiplication attenuation; spatially aggregate and encapsulate the discrete feature value sequence output after the above scalar multiplication attenuation calculation to construct a pure damage feature set after penalty correction; its representation, after topological consistency degradation filtering, has physically removed pseudo-random optical interference amplitude distortion and is ready to enter the downstream nonlinear activation intermediate state damage representation quantity with high physical confidence. Then, for each element in the pure damage feature set in the target region, perform nonlinear feature activation and boundary truncation based on the hyperbolic tangent function in sequence; for the truncated feature sequence, introduce a spatial Gaussian distribution weight factor (denoted as i) based on two-dimensional coordinates preset at the effective spatial feature grid node i. Perform a spatial weighted aggregation summation operation, and finally output the surface damage probability measure after penalty correction. The spatial Gaussian distribution weighting factor in this embodiment. The confidence gain distribution of the central node relative to the edge node in a connected domain is characterized by a normalized distribution of [0,1] that smoothly decreases from the center to the edge.

[0095] S3: In response to the surface damage probability measure, the processor is triggered to perform the following control operation:

[0096] The surface damage probability measure is subjected to boundary verification with the preset multi-level baseline discrimination threshold. In response to the verification result crossing or residing in the corresponding state classification interval, a first state label representing the real physical deformation of the target area or a second state label representing optical interference is generated. The above state labels are packaged and encapsulated into classification log data for updating the production line status monitoring screen.

[0097] Based on the classification log data, a process diversion instruction is generated and sent to the downstream production line control node through the cross-module control signaling channel. The process diversion instruction is configured to drive the production line control node to perform corresponding status classification actions on the corresponding batch of aluminum plates.

[0098] Step S3 further implementation instructions: Generate a first state marker characterizing the actual physical deformation of the target area or a second state marker characterizing optical interference, and package the above state markers into classified log data for updating the production line status monitoring screen, including:

[0099] Retrieve the preset baseline discrimination threshold library in the current business processing queue; the baseline discrimination threshold library is pre-configured and calibrated with a set safety critical upper limit characterizing physical deformation damage, and an environmental floor noise lower limit characterizing optical fluctuations;

[0100] The surface damage probability measure is input into the baseline discrimination threshold library for logical verification; in response to the condition that the surface damage probability measure reaches or exceeds the set safety critical upper limit, it is determined that there is physical deformation damage of the material surface in the local area, and a first state label characterizing the physical geometric damage characteristics of the surface is generated.

[0101] In response to the condition that the surface damage probability measure is lower than the set safety threshold and the local photometric disorder at the corresponding location is greater than the environmental floor noise threshold, it is determined that there is surface attachment interference in the target coordinate region, and a second state label characterizing the optical interference characteristics of the surface attachment is generated.

[0102] The first status marker or the second status marker is packaged and encapsulated into the classified log data and sent to the production line status monitoring screen.

[0103] For step S3, the multi-branch linkage status determination, instruction issuance, and fallback steps are performed as follows:

[0104] A multi-condition comprehensive classification dataset is obtained, which contains multiple sets of multi-band image features of aluminum plate surfaces collected under different industrial lighting conditions (backlighting, diffuse reflection, water stains, and oil film coverage), and their corresponding manually determined real business labels (including absolute flatness, optical artifacts, and substantial physical damage). The input features are fed into an initial parameter optimization network, using the minimum false positive recall rate and the maximum log-likelihood of structural coherence as the joint loss function. The parameter grid is iteratively adjusted using a backpropagation algorithm. When the convergence gradient of the loss function on the validation set is less than a preset stability threshold of 10... -4 At that time, the parameter system of the current network output is solidified, thereby outputting the set safety critical upper limit and environmental base noise bottom line for deployment online.

[0105] Set a safety threshold upper limit, denoted as Its characterization serves as a confidence threshold for determining the irreversible damage to physical materials. A safety criticality upper limit is set. This is achieved by constructing an offline data playback environment and continuously injecting a feature-driven data stream containing historical extreme edge conditions (e.g., large-area residual water stains combined with minor mechanical damage); dynamically adjusting and monitoring the state transition of the judgment link under different trial parameters, and extracting the time-series fluctuation curves of the global false positive interception rate and the actual damage recall rate; based on the set tolerance extreme value extrapolation, under the premise of ensuring that 99.9% of pure optical interference components are accurately released, the interception rate curve of the identification of substantial micro-scratches reaches convergence and no longer increases as the critical extreme value downward rule, locking the probability measure corresponding to the extreme value inflection point, and solidifying it as the set safety critical upper limit. The baseline value.

[0106] The environmental floor noise floor is denoted as It represents the extreme value of normal optical reflection fluctuations allowed by the current environment. In this embodiment, the preferred value of this dimensionless threshold is defined within the numerical range of [0.08, 0.15]. If the environmental floor noise baseline is set to deviate from the lower limit of 0.08, weak natural vibrations of equipment or normal material gloss will be misjudged as interference from attachments, resulting in invalid cleaning process instructions being issued; if the environmental floor noise baseline exceeds the upper limit of 0.15, the true shallow oil film optical artifact will be regarded as normal background, resulting in false negative omissions. In actual production line deployment, under conditions of drastic changes in factory ambient light or high-frequency micro-vibrations caused by the start-up and shutdown of surrounding heavy equipment, the environmental floor noise baseline is dynamically adjusted upward by receiving the situational fluctuation signaling from the aforementioned synchronously accessed environmental photoelectric sensor. Approaching the upper limit ensures the throughput and anti-interference stability of the entire inspection link; under silent or constant temperature and light conditions at night, the environmental floor noise floor is smoothly lowered. Approaching the lower limit to improve the ability to capture weak optical interference features.

[0107] Setting a safety threshold upper limit If the setting deviates from the lower limit, it will misjudge minor oil film that only requires simple cleaning as physical damage, leading to false positive scrap rate and material waste; if it exceeds the upper limit, it will be less sensitive to shallow physical scratches, allowing potentially hazardous parts to flow into the downstream assembly process.

[0108] Surface damage probability measure Input baseline discrimination threshold library for boundary verification: The baseline discrimination threshold library is configured as a structured calibration dictionary, which internally contains a series of feature parameters obtained by calibration through offline engineering sample sets.

[0109] The baseline discrimination threshold library contains a set safety critical upper limit characterizing physical deformation failure. This threshold is defined as a rigid lower bound for the surface damage probability measure to determine substantial damage. It represents the ultimate mechanical scratch response that the aluminum plate surface can withstand. The range of this safety threshold upper bound is specified as [0.75, 0.92]. In embodiments targeting aerospace-grade aluminum alloys (7075 series), this safety threshold upper bound is preferably configured as 0.85. When the surface damage probability measure... When this value is reached or exceeded, the first state flag is immediately triggered to ensure a safe redundancy in the structure's fatigue life.

[0110] Decision Branch 1: Response to Surface Damage Probability Measure Reaching or exceeding the set safety threshold Under the given conditions, it is determined that there is actual physical deformation and damage on the material surface in the local area, and a first state label representing the physical geometric damage characteristics is generated.

[0111] Decision Branch Two: Response to Surface Damage Probability Measure Below the set safety threshold Furthermore, the local luminous disorder at the corresponding location is still greater than the environmental floor noise threshold. Under the condition that only surface attachment interference exists within the target coordinate region, a second state marker characterizing the optical interference feature is generated.

[0112] Decision Branch 3: Response to Surface Damage Probability Measure Below the set safety threshold Furthermore, the local luminous disorder at the corresponding location is less than or equal to the environmental floor noise floor. Under the absolute normal conditions, the current target area is determined to be in an ideal flat state, and a third state mark representing both physical and optical compliance is generated. The third state mark is synchronously packaged and encapsulated into classified log data, and a default release command is issued through the cross-module control signaling channel to drive the production line to maintain the highest cycle time of normal processing operation, thereby realizing the absolute closed loop of the global state machine control flow.

[0113] The first or second state flag is packaged and encapsulated into categorized log data, and then sent to the production line status monitoring screen via a high-speed data bus, driving the monitoring screen to highlight abnormal blocks in real time. Simultaneously, corresponding process diversion instructions are generated based on the categorized log data and sent to downstream production line control nodes via cross-module control signaling channels, driving these nodes to execute corresponding state classification actions based on a state machine-driven paradigm. The first and second state flags are configured as enumerated polymorphic control identifiers during communication. After the system issues classification log data, the downstream production line control nodes read the logs in a polling manner: when the configuration attribute of the target object in the current workflow is determined to be in the first state flag of being effective, the control flow of the control node will drive the corresponding batch of aluminum plates to perform the corresponding state classification action; in response to the classification log data, the system triggers the state machine engine to transition from the normal processing operation state to the abnormal isolation and rejection state, and issues a physical intervention command through the cross-module control signaling channel to push the aluminum plate into the scrap transfer channel; when the configuration attribute is determined to be in the second state flag of being effective, the processing operation state is maintained, but the bypass diversion mechanism is triggered simultaneously to guide the aluminum plate to the cleaning section.

[0114] Furthermore, surface damage probability measure value range mapped to Within the standard probability interval.

[0115] Response to surface damage probability measure The condition that the output value infinitely approaches the theoretical minimum of 0 indicates that the spatial feature gradient of the current target area is discrete and random in a multi-directional structural topology. In actual physical monitoring operations, this state objectively corresponds to purely optical interference phenomena (such as discontinuous water stains on aluminum plate surfaces, residual "protective oil film / equipment dripping lubricant," or high-frequency reflections from environmental background noise). If the current feature is determined to lack any physical continuity that could cause mechanical damage, it is blocked from entering the deep detection model at the front end, allowing it to proceed directly or routing the work order to the cleaning and retry branch.

[0116] Response to surface damage probability measure The condition that the output value infinitely approaches the theoretical maximum value of 1 indicates that the current target region exhibits a highly concentrated directional texture distortion gradient in the multi-band frequency domain. This state objectively points to definite physical geometric damage (such as deep mechanical cutting marks or metal fatigue cracks). Based on this, it is determined that irreversible damage has occurred to the material surface structure, and the state machine engine is immediately triggered to generate the highest priority asynchronous anomaly isolation command and physical shunt signaling.

[0117] Global background noise factor With dynamic penalty offset and the subsequently generated scattering segmentation threshold A positively correlated nonlinear exponential mapping relationship is established between the environmental sensitivity control constant and the system environmental background lower limit constant. In concurrent industrial vision scenarios, when the external workshop is subjected to sudden strong direct sunlight or high-frequency reflection from heavy equipment, the variance of the basic optical fluctuation will increase sharply. If static feature checkpoint interception is used, the photometric mutation caused by reflection will be fully diverted to the subsequent heavy physical deformation analysis network. The global background noise factor will be used to... Designed as a scattering segmentation threshold The positive driving parameters conform to the objective law of negative feedback closed-loop regulation in automatic control theory.

[0118] Distribution ratio coefficient and target attenuation control factor The values ​​of exhibit a step-like conditional mapping relationship. In response to the condition that the distribution proportion coefficient crosses the orientation consistency threshold, the extracted control variable transitions from the strong attenuation coefficient of random optical noise to the weak penalty coefficient of true deformation characteristics.

[0119] Furthermore, this embodiment is configured in the following digital twin verification scenario: taking over the data node of the aluminum plate production line in a concurrent flow state. This visual node is dynamically affected by the movement of the large overhead crane and the high-frequency arcing and flickering of the surrounding heavy welding equipment, causing the ambient light intensity to fluctuate over time; at the same time, the surface of the aluminum plate target material is covered with different types of characterization states left over from the upstream process, including local water stains, high-frequency reflective oil films, and coexisting hidden weak mechanical scratches.

[0120] Minimum of basic optical fluctuation within a predefined featureless texture background The fixed calibration is 0.02, which is the maximum value of the basic optical fluctuation. The fixed calibration is 1.85; the system environment background lower limit constant is set to 0.05, and the environmental sensitivity control constant is configured to 1.5; the orientation consistency threshold baseline used for comparison is fixed at 3.5. After receiving the multidimensional image sequence data stream, the logical decision link triggers the synchronous update of features and the probability measure extraction engine based on the above input parameters.

[0121] This embodiment introduces "feature purification fidelity"; it is achieved by obtaining the target attenuation control factor. The final output surface damage probability measure after penalty correction The numerator of the ratio is used as the numerator of the ratio operation. Simultaneously, the original spatial probability expectation scalar value generated by benchmark accumulation aggregation within the same feature grid area before the introduction of the topological attenuation penalty step is extracted and used as the denominator of the ratio operation. A direct division operation is performed between the numerator and denominator to obtain the dimensionless result, which is defined as the feature purification fidelity. This index characterizes the ability to faithfully retain the effective physical deformation features attached to the original multidimensional input signal after performing extreme attenuation truncation flow targeting optical noise. The closer this value is to a constant 1, the more completely the true physical damage representation is inherited; the closer this value is to 0, the more thoroughly invalid random optical background noise is cleaned and isolated.

[0122] Under the condition of continuously inputting feature sets of different categories of environmental disturbances, the state response and calculation output logs generated in this embodiment are shown in Table 1:

[0123] Table 1: Examples of system response calculations under different environmental disturbances and defect modes

[0124]

[0125] Based on the comparison of the data flow mapping between Scenario 2 and Scenario 4, it can be directly deduced that when the sensed external signaling source transitions from a stable constant light input to a strong backlight direct illumination condition, the extracted front-end basic optical fluctuation variance increases abruptly from 0.08 to 1.35. This embodiment immediately invokes a nonlinear activation rule to push the noise floor factor parameter value up to 0.726; and sets the scattering segmentation threshold as the first line of defense. The value is adaptively and leapfrogged from a baseline of 0.65 to 1.88. This embodiment can actively tighten the entry acceptance tolerance of the feature extraction grid and truncate the injection of dirty data caused by large-area high-brightness flooding oil film at the forefront of the temporal capture stage.

[0126] Comparing Scenario 4 (photometric flooding with a large oil film) and Scenario 5 (photometric flooding with minor, hidden flaws), both types of targets in existing technologies trigger the same high-priority false positive mechanism due to high-frequency reflective absolute amplitudes exceeding the static safety limit of 0.85. This invention introduces a differentiated dimensionality reduction mechanism:

[0127] For Scenario 4, the distribution ratio coefficient of the oil film matrix representing the target area, determined through traversal matching, is only 2.1, which does not cross the comparison baseline setting. Therefore, a strong attenuation coefficient of 0.01 for random optical noise is issued and bound. Under this suppression, the surface damage probability measure output for Scenario 4 is forcibly attenuated to the silent range of 0.038. The upper limit of the preset trigger threshold of the traditional industrial baseline (0.85 probability judgment line) is used as a reference point for calculation: the safe reference point is obtained, and the value of the reference point is subtracted from the extracted actual surface damage probability measure of 0.038 to extract the net attenuation difference. Then, the net attenuation difference is divided by the aforementioned safe reference point to calculate the ratio. The ratio result confirms that this method has the beneficial effect of reducing the frequency of invalid false alarms caused by optical artifacts.

[0128] For scenario five, the mesh feature of the shallow, hidden solid damage has a distribution ratio coefficient of 4.2, successfully triggering a weak penalty coefficient of 0.88 for true deformation. It exhibits high fidelity while preserving this weak deformation feature, still stably outputting a surface damage probability measure of 0.825.

[0129] Figure 1 In this context, EXC represents a multi-source access node configured as an external camera. It is deployed at the front end of the physical production line to perform external data acquisition actions and obtain an initial image data stream containing multiple multi-band image acquisition channels.

[0130] LCU stands for Lighting Control Unit, which provides an operating status log to extract light source radiant intensity parameters and spatial illumination attitude parameters corresponding to the multi-band image acquisition channels.

[0131] MIS stands for Multiband Image Sequence Collection, which is a core data subset constructed after eliminating temporal deviations caused by inconsistent sampling frequencies based on timestamp alignment rules and performing background noise filtering.

[0132] RLE stands for Reference Illumination Parameter, which is generated by fusing light source parameters and serves as a reference baseline for evaluating the intensity of reflectivity disorder on the aluminum plate surface.

[0133] BSF stands for Basic Scattering Feature Distribution Map, which is a distribution map of local luminous disorder that deviates from the ideal flat state in the global space, calculated by a pre-configured network.

[0134] DST stands for Dynamically Generated Scattering Segmentation Threshold, which is calculated based on the aggregation of global background noise factor and baseline sensitivity threshold. It is used for connected component segmentation and matching to dynamically shield pseudo-random optical abrupt signals.

[0135] PFM stands for Local Physical Deformation Feature Matrix. It is a matrix constructed by extracting a subset of the image using a region of interest extraction mask and then inputting it into a pre-configured multi-band joint physical deformation feature extraction network to extract phase features.

[0136] OPW stands for Local Optical Scattering Disturbance Weight, which is a normalized weighting coefficient extracted from photometric disorder based on unified spatial reference projection resampling to characterize local optical scattering disturbance.

[0137] DPM stands for Surface Damage Probability Measure. It extracts an evaluation index that characterizes the degree of substantial physical damage by performing nonlinear mapping and weighted penalty aggregation calculation on optical scattering interference weights and reflection characteristic values.

[0138] BLC stands for Baseline Discrimination Threshold Validation. It retrieves a pre-configured discrimination threshold library with a set safety threshold upper limit and an environmental base noise lower limit, and performs cross- or dwell logic validation on probability measures.

[0139] FPD stands for First State Marker, which is a marker generated in response to a measurement reaching or exceeding the safety threshold, characterizing the presence of actual physical deformation and damage to the material surface in the target area.

[0140] FON stands for Second State Marker, which is generated in response to a metric below the upper limit and a luminance above the lower noise floor. It characterizes the optical interference caused by surface deposits (oil film, water stains) in the target area.

[0141] STM stands for Production Line Status Monitoring Dashboard. It receives and parses the tagging information packaged and encapsulated through categorized log data, and performs real-time production line macro status update rendering.

[0142] PLD represents the downstream production line control node, which receives process diversion instructions from the cross-module control signaling channel and drives the equipment to perform corresponding status classification actions on the corresponding batch of aluminum plates.

[0143] Figure 1The core control logic and physical flow nodes involved form a causal and collaborative link. A set of multi-band image sequences of the aluminum plate surface acquired by multi-source access nodes, along with synchronously associated reference illumination environment parameters, serve as the underlying sensing input of the system bus and are jointly guided to the basic optical feature extraction network. Based on the aforementioned multi-band image sequence set and reference illumination environment parameters, a basic scattering feature distribution map characterizing the surface photometric dispersion is extracted. In response to local values ​​exceeding a preset scattering segmentation threshold, phase features are extracted through frequency domain mapping to generate a local physical deformation feature matrix, and simultaneously, the corresponding local optical scattering interference weights are extracted from the distribution map. This penalty weight and the local physical deformation feature matrix undergo spatial mapping and weighted aggregation operations to output a surface damage probability measure after removing optical artifacts. During the control flow delivery phase, the processor responds to the surface damage probability measure and performs boundary verification with the preset multi-level baseline discrimination threshold. This generates a first state label representing the real physical deformation or a second state label representing optical interference. Finally, the state label is encapsulated and delivered to drive the update of the production line status monitoring screen and trigger the production line control node to perform physical process diversion actions based on state classification on the aluminum plate.

[0144] Figure 4 Numerical validation aims to test the theoretical response characteristics of this method under preset standardized boundary conditions (constant temperature and light, strong backlighting, high-frequency flashing light, and composite feature morphology test vectors). In the figure, the horizontal axis represents the preset test scenario set vector, and the vertical axis represents the calculated probability measure and feature fidelity values ​​obtained based on the aggregation logic of this method. Different filled bars represent the surface damage probability measure and feature purification fidelity obtained from the corresponding flow calculations, respectively.

[0145] Figure 4 Numerical calculation results clearly reveal the high-fidelity extraction of effective deformation features and the suppression of extreme value cutoff of pseudo-random optical interference by the proposed method. This demonstrates that the core methods proposed in this paper, including the distribution ratio coefficient and the target attenuation control factor, can achieve feature decoupling under conditions of concurrent heterogeneous data streams. In the test interval characterizing optical interference (such as the area of ​​strong backlighting superimposed with a photometric flooding oil film), the calculated surface damage probability measure exhibits an attenuation characteristic that is forcibly suppressed to a minimum (the calculated baseline value approaches 0.038). In contrast, in the area characterizing physical damage (such as the area of ​​strong backlighting superimposed with hidden dark damage), the surface damage probability measure and feature purification fidelity exhibit a high confidence characteristic that remains above the upper limit of the safety judgment (the calculated baseline value is above 0.825). This difference in quantitative distribution directly confirms that the introduction of the target attenuation control factor can reduce the ineffective fracturing intervention rate caused by optical artifacts and immunize against pseudo-random interference from high-brightness water stains.

[0146] Appendix Figure 5This visually presents the three heterogeneous band response channel data streams simultaneously captured by the multi-source access nodes at the front end within the same transient sampling period for a physical target material in a flowing state with composite contaminant sources on its surface. The left window in the figure represents the ultraviolet (UV) reflection channel data. Utilizing the high surface scattering characteristics and low transmittance of short-wave ultraviolet light, the microscopic scratches in the target area exhibit extremely high-contrast, high-brightness scattering characteristics; simultaneously, due to the physical effects of production line flow and mechanical rolling, surface deposits form discrete trails and longitudinal stretch marks distributed along the rolling direction on the substrate surface. Because the fluid medium strongly absorbs ultraviolet light, the aforementioned dynamic residues appear as extremely dark, scattered black stripes.

[0147] Figure 5 The central window represents the visible light (VIS) monochromatic channel data. Under this baseline geometric profile band, the residual liquid from the aforementioned longitudinal smear causes optical interference and high-frequency whitening reflections, masking the true physical damage. This window reveals the physical root cause of the high false positive rate in traditional single-source two-dimensional grayscale mapping when facing dynamic heterogeneous conditions;

[0148] The right-hand window in the figure represents near-infrared (NIR) penetration channel data. Leveraging the strong penetrating power of long-wave infrared light into liquids, this channel data achieves optical "visual penetration" of the aforementioned stretched water stains and coated oil films (the dynamic liquid interference medium exhibits a substantially transparent state), thereby exposing the continuous, true mechanical geometric fracture morphology of the substrate surface.

[0149] The computational logic involved in this application can be constructed using algorithms such as regression analysis in machine learning, establishing a mathematical model by analyzing the inherent trends and interrelationships of the collected parameters. This process can be implemented using specialized computational tools (such as Python's Scikit-learn library or the R language environment). Throughout all calculations, to eliminate the influence of different physical dimensions and ensure that data is compared and analyzed on the same scale, the input parameters in each formula are dimensionless. The dimensionless techniques used include, but are not limited to, max-min normalization or Z-score standardization.

[0150] To decouple the core algorithm from specific application strategies and ensure the configurability and ease of debugging of the technical solution, all configurable operating parameters in the specific implementation path of this invention are read through a standardized "configuration interface". The data source of this configuration interface is a "data storage module" (e.g., a non-transitory computer-readable storage medium, such as a configuration file, database entry, or cloud configuration service), which is configured to store configuration data in key-value pair format.

[0151] It should be emphasized that the foregoing embodiments are merely illustrative of preferred implementations of the present invention and are not intended to limit the scope of protection of the present invention. This application also provides a computer-readable storage medium having computer program instructions stored thereon.

Claims

1. A method for analyzing surface defects of aluminum plates based on multidimensional image feature extraction, characterized in that, The specific steps include: S1: Obtain a set of multi-band image sequences of the aluminum plate surface collected by the multi-source access node, and a reference illumination environment parameter synchronously associated with the set of multi-band image sequences; S2: Based on the multi-band image sequence set and the reference illumination environment parameters, extract the basic scattering feature distribution map that characterizes the surface photometric dispersion. In response to detecting that local values ​​in the basic scattering feature distribution map exceed a preset scattering segmentation threshold, a corresponding multidimensional image subset is extracted, and frequency domain mapping is performed on the multidimensional image subset to extract phase features, generating a local physical deformation feature matrix. Based on the basic scattering feature distribution map, the local optical scattering interference weights corresponding to the coordinates are extracted. The local optical scattering interference weights and the local physical deformation feature matrix are then subjected to spatial mapping and weighted aggregation operations to output a surface damage probability measure. S3: In response to the surface damage probability measure, the processor is triggered to perform the following control operation: The surface damage probability measure is subjected to boundary verification with the preset multi-level baseline discrimination threshold. In response to the verification result crossing or residing in the corresponding state classification interval, a first state label representing the real physical deformation of the target area or a second state label representing optical interference is generated. The above state labels are packaged and encapsulated into classification log data for updating the production line status monitoring screen. Based on the classification log data, a process diversion instruction is generated and sent to the downstream production line control node. The process diversion instruction is configured to drive the production line control node to perform corresponding status classification actions on the corresponding batch of aluminum plates.

2. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 1, characterized in that: The acquisition of a set of multi-band image sequences of the aluminum plate surface collected by the multi-source access nodes includes: The multi-source access node, configured as an external camera, performs external data acquisition; acquires an initial image data stream containing multiple multi-band image acquisition channels; performs time-series alignment on the initial image data stream based on a preset timestamp alignment rule to eliminate time-series deviations caused by inconsistent sampling frequencies of the multi-band image acquisition channels, generating a time-series aligned sequence; and performs background noise filtering on the time-series aligned sequence to construct the multi-band image sequence set.

3. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 2, characterized in that: Obtain reference illumination environment parameters that are synchronously associated with the multi-band image sequence set, including: Obtain the operation status log of the lighting control unit associated with the multi-source access node; parse the operation status log to extract the light source radiation intensity parameters and light source spatial illumination attitude parameters corresponding one-to-one with the multi-band image acquisition channels; fuse the light source radiation intensity parameters and the light source spatial illumination attitude parameters to generate the reference illumination environment parameters, which serve as a reference baseline for evaluating the intensity of surface emissivity disorder.

4. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 3, characterized in that: Based on the multi-band image sequence set and the reference illumination environment parameters, a basic scattering feature distribution map characterizing the surface photometric dispersion is extracted, and a local physical deformation feature matrix is ​​generated, including: The multi-band image sequence set and the reference illumination environment parameters are input into a pre-configured basic optical feature extraction network to calculate the local photometric disorder degree that deviates from the ideal flat state in the global space and generate the basic scattering feature distribution map. A dynamically generated scattering segmentation threshold is used to perform connected component segmentation matching on the basic scattering feature distribution map to extract the coordinate set of regions with abrupt changes in luminance; the coordinate set of regions with abrupt changes in luminance is converted into a region of interest extraction mask, and a multidimensional image subset is extracted from the multi-band image sequence set using the region of interest extraction mask; The multidimensional image subset is input into a pre-configured multi-band joint physical deformation feature extraction network to extract surface physical deformation gradient features across band boundaries and construct the local physical deformation feature matrix.

5. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 4, characterized in that: Based on the basic scattering feature distribution map, local optical scattering interference weights corresponding to the coordinates are extracted. Spatial mapping and weighted aggregation operations are then performed on the local optical scattering interference weights and the local physical deformation feature matrix to output a surface damage probability measure, including: Based on a unified spatial reference feature grid, projection alignment resampling is performed on the basic scattering feature distribution map with different feature scales and the local physical deformation feature matrix; the local normalized photometric disorder is extracted from the aligned basic scattering feature distribution map as the local optical scattering interference weight; The local optical scattering interference weights are subjected to element-wise nonlinear mapping and weighted penalty aggregation calculation with the multi-band local pixel reflection feature values ​​in the aligned local physical deformation feature matrix; based on the calculation results, a surface damage probability measure characterizing the true physical damage degree of the surface is extracted.

6. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 5, characterized in that: Generate a first state marker representing the actual physical deformation of the target area or a second state marker representing optical interference, and package these state markers into categorized log data for updating the production line status monitoring dashboard, including: Retrieve the preset baseline discrimination threshold library in the current business processing queue; the baseline discrimination threshold library is pre-configured and calibrated with a set safety critical upper limit characterizing physical deformation damage, and an environmental floor noise lower limit characterizing optical fluctuations; The surface damage probability measure is input into the baseline discrimination threshold library for logical verification; in response to the condition that the surface damage probability measure reaches or exceeds the set safety critical upper limit, it is determined that there is physical deformation damage of the material surface in the local area, and a first state label characterizing the physical geometric damage characteristics of the surface is generated. In response to the condition that the surface damage probability measure is lower than the set safety threshold and the local photometric disorder at the corresponding location is greater than the environmental floor noise threshold, it is determined that there is surface attachment interference in the target coordinate region, and a second state label characterizing the optical interference characteristics of the surface attachment is generated. The first status marker or the second status marker is packaged and encapsulated into the classified log data and sent to the production line status monitoring screen.

7. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 6, characterized in that: The basic scattering feature distribution map is subjected to connected component segmentation matching using a dynamically generated scattering segmentation threshold, including: Obtain the basic optical fluctuation variance of the surface of the current batch of aluminum plates within a preset background area, and map the basic optical fluctuation variance into a global background noise factor that characterizes the intensity of the current ambient light interference. Based on a preset threshold adaptive adjustment curve, the global background noise factor is converted into a dynamic penalty offset. The preset baseline sensitivity threshold and the dynamic penalty offset are aggregated to dynamically generate the scattering segmentation threshold; In response to the condition that the value of a local grid in the basic scattering feature distribution map is greater than the scattering segmentation threshold, the local grid is determined to be a suspected defective pixel connected region, and the continuous suspected defective pixel connected regions are aggregated to generate the coordinate set of the photometric abrupt change region; wherein, when the global background noise factor, which represents the ambient light interference intensity, increases, the scattering segmentation threshold is synchronously increased to shield pseudo-random optical abrupt change signals from entering the subsequent physical deformation feature extraction network.

8. The method for analyzing aluminum plate surface defects based on multidimensional image feature extraction according to claim 7, characterized in that: The local optical scattering interference weights are subjected to element-wise nonlinear mapping and weighted penalty aggregation calculations with the multi-band local pixel reflection feature values ​​in the aligned local physical deformation feature matrix, including: Traverse each feature node in the aligned local physical deformation feature matrix, extract the cross-band surface texture distortion gradient between adjacent feature nodes, and construct a defect contour topological coherence feature matrix to characterize the spatial coherence of micro-deformation. Based on the preset surface deformation topological coherence comparison logic, a logical comparison is performed on the topological coherence feature matrix of the defect contour: In response to the condition that the defect contour topological coherence feature matrix in the target area is discretely random, a hard degradation strategy is triggered, discarding the original calculated value of the local optical scattering interference weight under the corresponding coordinates, and forcibly replacing its state pointer with a preset random optical noise strong attenuation coefficient. In response to the condition that the defect contour topological coherence feature matrix in the target area is continuously oriented, the confidence of the true deformation is maintained, and the parameter flow node under the corresponding coordinate is directly assigned the preset weak penalty coefficient of the true deformation feature. The random optical noise strong attenuation coefficient or the real deformation feature weak penalty coefficient is used to perform element-wise scalar multiplication attenuation calculation on the multi-band local pixel reflection feature value to obtain the pure damage feature set after penalty correction. The set of pure damage features within the target area is subjected to nonlinear feature activation and boundary truncation based on the hyperbolic tangent function. Subsequently, a spatial weighted aggregation summation operation is performed by fusing locally preset spatial distribution weight factors to generate the surface damage probability measure.

Images