Image feature extraction system for visual inspection

Abstract

The application relates to the technical field of visual detection, and discloses an image feature extraction system for visual detection, which comprises a spatial distribution quantization unit, an asymmetric feature routing unit and a graph reconstruction unit. The quantization unit divides a pixel matrix into pixel blocks and determines spatial information entropy. The routing unit routes each block to a first operator link or a second operator link according to the spatial information entropy and a Fourier transform domain phase consistency checking result, generates a high-dimensional semantic feature tensor or a low-dimensional structural feature tensor, and the reconstruction unit inlays the high-dimensional tensor into a low-dimensional tensor to output a reconstructed graph. According to the application, the asymmetric routing shunts the calculation resource distribution, the semantic density is used as the basis for the calculation resource distribution, the flow delay in processing high-resolution images is reduced, the phase consistency checking distinguishes physical features from noises, the anti-interference capability is enhanced, and the reverse coordinate conversion maintains the geometric topological correlation.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and in particular to an image feature extraction system for visual inspection. Background Technology

[0002] Current industrial vision inspection systems typically employ feature extraction operators or convolutional neural networks to extract features from the pixel matrix of the target surface. Existing processing architectures generally follow a spatial uniform processing approach, applying operations of equal scale and depth to the global field of view of the input image. This processing mode has good universality in routine inspection tasks.

[0003] However, in high-throughput industrial inspection scenarios such as semiconductor wafer inspection and continuous roll-to-roll thin film inspection, the information density on the workpiece surface exhibits an extremely asymmetrical distribution. Real defect features are sparse in physical space, and the homogeneous background occupies a large proportion of the global field of view. This spatially uniform processing method causes the system to consume most of its computational resources in background areas that lack semantic value, limiting the processing latency of the data link and making it difficult to meet the 10μs-level processing requirements. To improve detection accuracy, it is usually necessary to increase the absolute resolution of the input image, which leads to a surge in computational overhead. While reducing the sampling rate or using lightweight operators reduces the computational load, it can result in missed defects due to loss of detail. Software... When optimizing detection performance using component control algorithms, feature extraction schemes often fall into a dilemma of synchronizing feature dimension gain with computational load gain. For example, Chinese invention patent CN119540529B discloses an infrared small target detection method based on grayscale scale difference and information entropy. It uses local information entropy as a weighting factor to supplement the loss of gradient domain information and enhance the saliency of weak target features. The feature superposition logic is based on low sampling frequency and high tolerance for latency. In high-throughput scenarios with tens of thousands of frames per second, the entropy feature is only calculated using a global weighted operator across the entire image, which cannot break the bottleneck of uniform distribution of computing resources. Nonlinear logarithmic operations exacerbate the pressure of single-frame data flow, causing the feature extraction process to lose real-time response capability at ultra-high resolution.

[0004] Therefore, how to dynamically allocate feature routes based on the intrinsic properties of image pixels, and break the rigid coupling between processing latency and resolution while ensuring detection accuracy, is the technical problem to be solved by this invention. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an image feature extraction system for visual inspection, comprising: The spatial distribution quantization unit is used to obtain the original pixel matrix, divide it into several pixel blocks of the same scale, calculate the pixel gray level distribution probability within each pixel block, and determine the corresponding spatial information entropy based on the pixel gray level distribution probability. An asymmetric feature routing unit, connected to a spatial distribution quantization unit, is used to establish an asymmetric routing topology that includes a first operator link and a second operator link. The first operator link is used to extract high-resolution semantic features, and the second operator link is used to extract low-resolution structural features. The asymmetric feature routing unit routes each pixel block to the first operator link or the second operator link based on the comparison result of spatial information entropy and a preset decision threshold, and in combination with the phase consistency verification result of pixel blocks in the Fourier transform domain. The graph reconstruction unit is connected to the first operator link and the second operator link respectively. It is used to obtain the original spatial coordinates of the pixel blocks in the original pixel matrix, embed the high-dimensional semantic feature tensor output by the first operator link into the low-dimensional structural feature tensor output by the second operator link according to the original spatial coordinates, and smooth the embedding boundary to output the heterogeneous reconstructed feature graph.

[0006] Preferably, the spatial distribution quantization unit provides routing weight constraints for the asymmetric feature routing unit by running noise identification logic based on phase consistency, specifically including the following steps: Step S201, obtaining the local phase features of each pixel block in the Fourier transform domain; Step S202, determining the physical structure consistency index of the pixel block based on the local phase features; Step S203, comparing the physical structure consistency index with a preset noise discrimination standard. If the physical structure consistency index is lower than the preset noise discrimination standard, the spatial information entropy is determined to be an interference signal, and a forced routing instruction is sent to the asymmetric feature routing unit to guide the corresponding pixel block to the second operator link.

[0007] Preferably, the image feature extraction system for visual inspection further includes a recognition unit connected to the map reconstruction unit. The asymmetric feature routing unit further includes a feedback correction module for obtaining the recognition confidence score output by the recognition unit and dynamically correcting the preset decision threshold based on the recognition confidence score. If the recognition confidence score is lower than the preset confidence threshold, the preset decision threshold is reduced to increase the routing ratio of the first operator link.

[0008] Preferably, the first operator link includes multiple convolution operators with different local receptive fields, used to perform multi-scale mapping on pixel blocks to obtain texture parameters representing heterogeneous semantic details; the second operator link includes a global pooling operator, used to map pixel blocks into low-dimensional feature vectors representing background attributes.

[0009] Preferably, when determining the spatial information entropy, the spatial distribution quantization unit simultaneously calculates the boundary gradient change rate of the pixel block and uses the boundary gradient change rate as a weighted compensation factor for the spatial information entropy to correct the spatial information entropy.

[0010] Preferably, when dividing pixel blocks, the spatial distribution quantization unit sets the pixel step size to be equal to the boundary length of the pixel block so as to seamlessly cover the original pixel matrix; the numerical range of spatial information entropy is normalized to 0 to 1.

[0011] Preferably, when embedding a high-dimensional semantic feature tensor into a low-dimensional structural feature tensor, the graph reconstruction unit uses the low-dimensional structural feature tensor as a global spatial reference and performs pixel alignment of the high-dimensional semantic feature tensor according to the original spatial coordinates to maintain the geometric topological association between the heterogeneous semantic features and the background.

[0012] Preferably, the graph reconstruction unit performs Gaussian smoothing filtering at the splicing boundary between the high-dimensional semantic feature tensor and the low-dimensional structural feature tensor, and smooths out feature jumps at the splicing boundary by adjusting the standard deviation parameter of the Gaussian kernel function.

[0013] Preferably, the system also includes an adaptive compensation module, which monitors ambient light intensity fluctuations and applies compensation to a preset decision threshold based on the ambient light intensity fluctuations to offset the entropy shift of pixel blocks under unbalanced lighting conditions, so that the asymmetric feature routing unit can maintain consistency in recognizing semantic distribution.

[0014] The beneficial effects of this invention are: 1. In image feature extraction for visual inspection, this invention breaks the limitation of spatially uniform allocation of computing resources in traditional image processing by the synergistic effect of local topological quantization units and asymmetric feature routing units. The system guides low-entropy background regions to the first feature operator link and routes high-entropy potential defect regions to the second feature operator link according to the distribution of local information entropy of the image. This asymmetric resource scheduling mechanism makes the allocation of computing resources directly depend on the true semantic density inside the image, rather than the original resolution of the input. Therefore, while maintaining the accuracy of feature extraction of small physical defects, it reduces the turnaround time of the system when processing high-resolution visual images and solves the constraint between recognition accuracy and real-time performance in high-speed online detection.

[0015] 2. This invention introduces a discrimination logic based on phase consistency verification to solve the problem of pseudo-high-entropy signal interference caused by industrial environment interference. By extracting the local phase features of the block image in the Fourier transform domain before asymmetric routing decision, the system can distinguish physical structure features from random light source noise. This multi-dimensional discrimination mechanism ensures that only real physical defects with strong phase consistency enter the high-dimensional semantic extraction link, avoiding the ineffective occupation of computing resources by random noise. This mechanism enhances the system's anti-interference ability under complex lighting and different surface textures, and ensures the stability of the feature extraction process.

[0016] 3. By using the inverse coordinate mapping and Gaussian smoothing of the heterogeneous reconstruction unit, this invention solves the problem of spatial topological distortion after non-uniform feature extraction. The system uses the high-resolution semantic slices extracted by the second feature operator link as spatial anchors to embed into the low-dimensional feature tensor output by the first feature operator link. This heterogeneous feature fusion method not only preserves the fine-grained texture details of the defect area, but also maintains the geometric positional relationship between the defect and the global background of the workpiece. This maintenance of topological consistency provides a reliable data foundation for the downstream identification unit to accurately determine the defect type and spatial attributes. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a diagram showing the composition of the image feature extraction system for asymmetric routing topology according to the present invention. Figure 2 This is a flowchart illustrating the feature extraction and scheduling process of the present invention, which integrates adaptive light intensity compensation and feedback correction. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, an embodiment or embodiment referred to herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention. An embodiment appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] This invention is described in detail with reference to the schematic diagrams. When describing the embodiments of this invention, for ease of explanation, the cross-sectional views of the device structure will be partially enlarged without adhering to the general scale. Moreover, the schematic diagrams are only examples and should not limit the scope of protection of this invention. In addition, in actual manufacturing, the three-dimensional spatial dimensions of length, width and depth should be included.

[0022] Furthermore, in the description of this invention, it should be noted that the terms such as "upper," "lower," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or component referred to has a specific orientation, or is constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Unless otherwise explicitly specified and limited, the terms installation, connection, and linking in this invention should be interpreted broadly. For example, they can refer to fixed connection, detachable connection, or integrated connection; similarly, they can refer to mechanical connection, electrical connection, or direct connection, or indirect connection through an intermediate medium, or internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] An image feature extraction system for visual inspection includes: The spatial distribution quantization unit is used to obtain the original pixel matrix, divide it into several pixel blocks of the same scale, calculate the pixel gray level distribution probability within each pixel block, and determine the corresponding spatial information entropy based on the pixel gray level distribution probability. An asymmetric feature routing unit, connected to a spatial distribution quantization unit, is used to establish an asymmetric routing topology that includes a first operator link and a second operator link. The first operator link is used to extract high-resolution semantic features, and the second operator link is used to extract low-resolution structural features. The asymmetric feature routing unit routes each pixel block to the first operator link or the second operator link based on the comparison result of spatial information entropy and a preset decision threshold, and in combination with the phase consistency verification result of pixel blocks in the Fourier transform domain. The graph reconstruction unit is connected to the first operator link and the second operator link respectively. It is used to obtain the original spatial coordinates of the pixel blocks in the original pixel matrix, embed the high-dimensional semantic feature tensor output by the first operator link into the low-dimensional structural feature tensor output by the second operator link according to the original spatial coordinates, and smooth the embedding boundary to output the heterogeneous reconstructed feature graph.

[0025] Preferably, the spatial distribution quantization unit provides routing weight constraints for the asymmetric feature routing unit by running noise identification logic based on phase consistency, specifically including the following steps: Step S201, obtaining the local phase features of each pixel block in the Fourier transform domain; Step S202, determining the physical structure consistency index of the pixel block based on the local phase features; Step S203, comparing the physical structure consistency index with a preset noise discrimination standard. If the physical structure consistency index is lower than the preset noise discrimination standard, the spatial information entropy is determined to be an interference signal, and a forced routing instruction is sent to the asymmetric feature routing unit to guide the corresponding pixel block to the second operator link.

[0026] Preferably, the image feature extraction system for visual inspection further includes a recognition unit connected to the map reconstruction unit. The asymmetric feature routing unit further includes a feedback correction module for obtaining the recognition confidence score output by the recognition unit and dynamically correcting the preset decision threshold based on the recognition confidence score. If the recognition confidence score is lower than the preset confidence threshold, the preset decision threshold is reduced to increase the routing ratio of the first operator link.

[0027] Preferably, the first operator link includes multiple convolution operators with different local receptive fields, used to perform multi-scale mapping on pixel blocks to obtain texture parameters representing heterogeneous semantic details; the second operator link includes a global pooling operator, used to map pixel blocks into low-dimensional feature vectors representing background attributes.

[0028] Preferably, when determining the spatial information entropy, the spatial distribution quantization unit simultaneously calculates the boundary gradient change rate of the pixel block and uses the boundary gradient change rate as a weighted compensation factor for the spatial information entropy to correct the spatial information entropy.

[0029] Preferably, when dividing pixel blocks, the spatial distribution quantization unit sets the pixel step size to be equal to the boundary length of the pixel block so as to seamlessly cover the original pixel matrix; the numerical range of spatial information entropy is normalized to 0 to 1.

[0030] Preferably, when embedding a high-dimensional semantic feature tensor into a low-dimensional structural feature tensor, the graph reconstruction unit uses the low-dimensional structural feature tensor as a global spatial reference and performs pixel alignment of the high-dimensional semantic feature tensor according to the original spatial coordinates to maintain the geometric topological association between the heterogeneous semantic features and the background.

[0031] Preferably, the graph reconstruction unit performs Gaussian smoothing filtering at the splicing boundary between the high-dimensional semantic feature tensor and the low-dimensional structural feature tensor, and smooths out feature jumps at the splicing boundary by adjusting the standard deviation parameter of the Gaussian kernel function.

[0032] Preferably, the system also includes an adaptive compensation module, which monitors ambient light intensity fluctuations and applies compensation to a preset decision threshold based on the ambient light intensity fluctuations to offset the entropy shift of pixel blocks under unbalanced lighting conditions, so that the asymmetric feature routing unit can maintain consistency in recognizing semantic distribution.

[0033] Example 1: In a high-throughput industrial continuous manufacturing vision inspection system for wafer surface defects, the input raw image pixel scale is... The system requires a spatial resolution of less than 10 μs for feature extraction from a single frame of image. It must balance the computational overhead of high spatial resolution sampling for capturing micrometer-level physical defects with the computational resource overhead of full-image convolution. The spatial distribution quantization unit acquires the original pixel matrix, divides it into several pixel blocks of the same scale, and extracts the pixel grayscale distribution probability within any pixel block. To determine spatial information entropy ,in, For spatial information entropy, The probability of pixel grayscale distribution. For gray levels within a pixel block, spatial information entropy The calculation formula is as follows: Asymmetric feature routing units are based on spatial information entropy The route decision is executed based on the comparison result with the preset decision threshold, when the spatial information entropy is determined. When the value is less than a preset decision threshold, the asymmetric feature routing unit generates a first routing instruction, directing the data pointer of the corresponding pixel block to the second operator link used to extract low-resolution structural features. Meanwhile, when determining the spatial information entropy... When the threshold value is greater than or equal to a preset decision threshold, the spatial distribution quantization unit runs noise identification logic based on phase consistency. Relying on the local energy physics model, it calls the Log-Gabor filter bank to perform a two-dimensional frequency domain mapping on the pixel blocks, extracting odd-symmetric and even-symmetric filter response components at each frequency scale. The spatial distribution quantization unit calculates the local amplitude and local energy at the corresponding scale based on the obtained filter response components. The ratio of the sum of local energy to the sum of local amplitude at all frequency scales is output as the quantization value. The phase consistency index of the pixel blocks in the Fourier transform domain is calculated. To distinguish physical feature information from light source speckle signals, among which The phase consistency index is used to characterize the degree of phase superposition of signals in the frequency domain. When the value exceeds the preset phase verification threshold of 0.5, the asymmetric feature routing unit directs the corresponding original pixel blocks to the first operator link for extracting high-resolution semantic features. In the specific computational flow of constructing this extraction logic, the system selects a Log-Gabor filter array with a center frequency spanning four logarithmic distribution scales and a bandwidth octave constant of 1.5. The core execution steps of the algorithm are as follows: the odd-symmetric response signal and the even-symmetric response signal within a specific scale frequency band are squared and summed. The positive square root of the summation result is defined as the local amplitude. At the same time, a fixed noise floor constant is subtracted from the even-symmetric signal to reduce low-frequency interference. The positive difference output is defined as the local energy at the corresponding scale. The total local energy accumulated at each scale is divided by the total local amplitude, and a small constant is superimposed on the denominator to avoid division by zero error. Thus, a dimensionless stable phase consistency value distributed between 0 and 1 is calculated.

[0034] The graph reconstruction unit obtains the original spatial coordinates of each pixel block in the original pixel matrix, and uses the high-dimensional semantic feature tensor output by the first operator link as spatial anchor points to embed into the low-dimensional structural feature tensor output by the second operator link based on the original spatial coordinates. Simultaneously, Gaussian smoothing filtering is used to process the pixel gradients at the embedding boundaries. When performing Gaussian smoothing filtering, the graph reconstruction unit extracts the mean difference in width grayscale between adjacent pixels on both sides of the embedding boundary based on the linear constraint formula. Real-time calculation of Gaussian kernel function standard deviation parameter ,in, The gray-scale mean difference ratio coefficient of the boundary region With the basic smoothing constant Based on the solidified results of gray-scale step calibration experiments during the system initialization phase, a physically smooth mapping relationship is established for different degrees of mosaic feature transitions, generating a heterogeneous reconstructed feature map. This reconstruction process concentrates the computational resources for feature extraction on spatial information entropy. The higher pixel density allows the feature extraction latency of the visual detection system to remain within a 10μs operating window, achieving alignment between feature extraction accuracy and system real-time processing. It's important to emphasize that, since the high-dimensional semantic feature tensor is a multi-channel abstract data matrix and lacks a single grayscale attribute, the map reconstruction unit establishes a coordinate inversion pointer to inversely map back to the original input pixel matrix before performing smoothing. This extracts the real physical pixel clusters on both sides of the region with the same topological coordinates as the aforementioned stitching boundary at the physical image level, and uses the real pixel array to calculate the aforementioned single-dimensional boundary region grayscale mean difference. After obtaining the standard deviation parameter of the corresponding region Subsequently, the map reconstruction unit synchronously broadcasts this scalar parameter calculated based on the underlying physical optics step and applies it equivalently to all independent channel planes of the high-dimensional semantic feature tensor for parallel Gaussian filtering, thereby opening up a feasible physical conversion path from physical grayscale information to cross-domain quantization abstraction of the smoothness threshold of the high-dimensional feature tensor.

[0035] Example 2: In a vision inspection platform simulating a high-throughput continuous manufacturing environment, an industrial area scan camera is selected to acquire the original pixel matrix, and the resolution of a single frame image is set to [value missing]. The frame rate was set to 100Hz. The test dataset contained 5000 original images with micron-level physical defects. To examine the system's stability under non-ideal industrial conditions, Gaussian white noise with a signal-to-noise ratio of 25dB was superimposed on the original image data. The scale parameter of the spatial distribution quantization unit for pixel block division was also specified. The scale is set to 64×64. The determination of this scale depends on the balance between the resolution requirements of feature extraction and the computational management overhead. When the size is below 32×32, the memory bandwidth usage for maintaining the block index pointers exceeds 80%, resulting in a scheduling delay of more than 5μs when routing instructions are issued. Furthermore, when the scale parameter... When the value is higher than 128×128, the spatial information entropy within the block is... The inability to characterize the gray-level abrupt changes of local minute defects leads to a decrease in the accuracy of feature sparsity recognition. Therefore, 64×64 is selected as the engineering example value under standard working conditions.

[0036] The experiment included the present invention sample group, a control sample group A, and a control sample group B. Control sample group A used a global equal-scale convolution method, while control sample group B removed the phase consistency check step in asymmetric feature routing. When processing an image where the background accounted for 95%, control sample group A had a single-frame processing latency of 86.4 ms, while the present invention sample group used spatially distributed quantization units to determine the gray-level distribution probability of pixel blocks. And calculate spatial information entropy The background area Values ​​in the range of 0.82 to 1.18 indicate defective areas. With values ​​ranging from 3.55 to 4.15, the system guides the low-entropy background to the second operator link for low-resolution feature extraction based on a preset decision threshold of 1.5, stabilizing the overall processing latency at 7.8 μs. This data confirms the ability of the asymmetric routing architecture to suppress homogeneous redundant computation.

[0037] To determine the response boundary of the decision parameters, the preset decision threshold was adjusted from 1.0 to 5.0. It was observed that when the preset decision threshold was below 1.5, the slight fluctuations in pixel grayscale caused by ambient light scattering affected some background blocks. When the value reaches 1.45, the task load of the first operator link increases from 5% to 35%, and the system processing latency increases non-linearly from 8.2μs to 1.2ms. When the preset decision threshold is increased to above 4.0, the local entropy value of edge fuzziness defects drops to about 3.8, causing the first routing command to be missed. The missed detection rate suddenly increases from 0.45% to 12.1%. The appearance of this performance inflection point confirms that 1.5 to 3.0 is the optimal working window of the present invention under the current optical contrast.

[0038] In the anti-interference verification of actively injected 25dB Gaussian white noise, the spatial information entropy of the comparison sample group B, due to the lack of phase consistency verification, was lower. Due to the randomness of noise phase, the phase shifts to around 2.5, causing misjudgments in the routing logic and degrading the single-frame processing latency to 45.6μs. The sample group of this invention determines the phase consistency index in the Fourier transform domain through spatially distributed quantization units. Used to characterize the degree of phase superposition of signals in the frequency domain, random noise. The measured value remained below 0.15, while the physical defect characteristics, due to the high superposition of phases in the frequency domain, The measured value stabilized above 0.75. The system, based on a preset phase verification threshold of 0.5, shielded against high-entropy noise interference, maintaining a detection accuracy of 98.6% with only a slight increase in processing latency to 9.1 μs. This gradient comparison confirms the synergistic value of phase consistency verification and asymmetric routing features in ensuring feature extraction accuracy under strong interference environments. In fact, the preset decision criterion of 1.5 and the phase truncation threshold of 0.5 used in implementing this scheme are precisely the results extracted from a sample library of different illumination fluctuation contrasts through forward and backward reasoning. Experimental observations show that strong speckle reflection... The peak limit of the pseudo-high entropy of the background that can be excited by the radiation is about 1.45. In order to prevent redundant blocks from entering the deep operator under the maximum noise interference, the decision boundary is conservatively rounded to 1.5 to establish a reliable defense line. Similarly, the absolute upper limit of phase consistency based on random white noise interference has never exceeded 0.15, while the lower limit of phase aggregation of real material scratches has never fallen below the 0.75 baseline. The midline coordinate of the wide safety interval between the two is directly selected as 0.5 as the standard for solidification and control, eliminating the blindness of constant selection from the perspective of engineering experience and experimental measurement.

[0039] Example 3: In the continuous visual monitoring task of the aluminum-plastic film coating production line, the production line speed is maintained at 60m / min. Due to the fluctuation of the external power supply load, the ambient light source experiences a brightness drift of ±15%, causing the grayscale reference of the original pixel matrix acquired by the spatial distribution quantization unit to change in different time batches. Because the inherent static threshold faces the risk of missing weak scratches under low contrast conditions or causing computing power congestion in the first operator link due to routing errors under bright field noise conditions, the spatial distribution quantization unit receives the original pixel matrix with grayscale shift characteristics and determines the spatial information entropy. During the process, the ambient light compensation factor is determined by statistically analyzing the global pixel mean of the pixel matrix. Beforehand, under a standard constant light source reference, an unloaded calibration procedure is run to collect the global pixel mean of a defect-free reference substrate and store it in a register as the basic illumination reference value. When the spatial distribution quantization unit collects frame sequences, it simultaneously obtains the real-time global pixel mean of the current original pixel matrix. The quotient of the division between the basic illumination reference value and the real-time global pixel mean is directly set as the ambient light compensation factor. To filter out the interference of external environmental drift on the background grayscale baseline, and to increase the spatial information entropy With ambient light compensation factor The calibration spatial information entropy is obtained by performing a product operation, where For spatial information entropy, Spatial information entropy is a compensation coefficient used to correct illumination deviations. The calculation logic follows the following formula: ,in, This represents the probability distribution of pixel gray levels within a pixel block. This refers to the grayscale level within a pixel block.

[0040] The asymmetric feature routing unit establishes an asymmetric routing topology. The first operator link includes 12 residual convolutional layers for extracting high-resolution semantic features, and the second operator link includes 3 linear sampling layers for extracting low-resolution structural features. This is achieved when the calibration spatial information entropy is greater than a preset decision threshold of 1.5 and the phase consistency index... When the value exceeds the preset phase verification threshold of 0.5, the asymmetric feature routing unit directs the corresponding pixel block to the first operator link; otherwise, it directs the pixel block to the second operator link. The system then obtains the recognition confidence score fed back by the downstream recognition unit. When the recognition confidence level of 10 consecutive frames of images When the arithmetic mean is lower than the preset safety benchmark of 0.85, the asymmetric feature routing unit initiates a threshold adjustment mode, reducing the preset decision threshold in steps of 0.05 until the recognition confidence level is reached. The arithmetic mean rebounded to above 0.90. In this dynamic correction stage, to resolve the physical timing conflict between the long-time discrimination feedback of the downstream recognition model and the requirement for the front-end routing to complete the judgment within 10μs, the feedback adjustment mechanism is configured at the hardware level as a dual-process asynchronous scheduling architecture. The main routing thread responsible for high-throughput distribution of front-end images continuously reads the current decision threshold fixed in the local register at a pipeline rate. The confidence collection and long-term mean calculation logic are isolated and run in an independent background coprocessing process. After the background thread completes the data accumulation across frame rates and calculates the new threshold, it only overwrites the register data of the main routing thread through an instantaneous atomic lock-free operation, thereby blocking the drag and interference of the long delay of the back-end feedback on the high-speed scheduling efficiency of the front end at the physical link level. The map reconstruction unit uses the original spatial coordinates of each pixel block to heterogeneously stitch the feature tensors output by the first operator link and the second operator link. Under the condition of a 10% decrease in light intensity, the ambient light compensation factor The parameter was revised from 1.0 to 1.12, which keeps the calibration spatial information entropy of the weak scratch area above the routing threshold of 1.5 and stabilizes the computing power utilization of the first operator link below the hardware load warning line. This feedback-based parameter evolution method eliminates the impact of external environmental disturbances on feature extraction performance, enabling the visual inspection system to maintain a match between feature extraction quality and real-time processing capability in the aluminum-plastic film coating scenario.

[0041] Example 4: In the continuous manufacturing scenario of high-reflectivity ultrathin copper foil, the visual inspection system adapts to the texture features of the metal surface and the distribution of stray light in the environment by determining the inspection benchmark. Before the recognition program starts, the spatial distribution quantization unit acquires 500 frames of the original pixel matrix of the defect-free surface. The system traverses each pixel block in each original pixel matrix and determines the spatial information entropy. ,in For spatial information entropy, This represents the probability distribution of pixel gray levels within a pixel block. For gray levels within a pixel block, spatial information entropy The calculation logic follows the following formula: Statistical spatial information entropy The distribution state in the original pixel matrix is ​​used to determine the initial value of the preset decision threshold. ,in The initial value is the preset decision threshold, and the initial value is... Set as the spatial information entropy in the original pixel matrix The 99.0 percentile value thus limits the homogeneous background region to the routing reception range of the second operator link.

[0042] When the visual inspection system experiences changes due to camera mounting height adjustments or lens focal length fine-tuning, the system runs a phase calibration program. It uses standard scale samples to determine the phase response characteristics of the current optical system, and the spatial distribution quantization unit calculates the phase offset of the standard scale edge at different frequency components to obtain a reference phase consistency index. and the benchmark phase consistency index Set as the zero-point reference value for phase verification logic, where As a reference phase consistency index, the asymmetric characteristic routing unit uses the zero-point reference value and the measured phase consistency index. The deviation rate is used to correct the phase distribution disorder caused by optical dispersion, where As a phase consistency index, it is used to characterize the degree of phase superposition of signals in the frequency domain, and image feature extraction systems in different deployment environments maintain a unified routing decision benchmark.

[0043] Example 5: When a vision inspection system deals with transient emergencies in a continuous production line, the asymmetric feature routing unit monitors the real-time occupancy rate of hardware resources. And run the feature overflow truncation program, when the spatial information entropy is reduced due to large-area loss on the surface of the workpiece being measured. When the proportion of pixel blocks exceeding the preset decision threshold of 1.5 exceeds 40% of the total area of ​​the original pixel matrix, the real-time utilization rate of hardware resources extracted by the asymmetric feature routing unit is [not specified]. And determine whether it has reached the 85% safety warning line, if so, the real-time occupancy rate If the safety threshold is exceeded, the asymmetric feature routing unit will then rely on spatial information entropy. The pixel blocks in the routing waiting state are sorted from low to high values, and the calculation requests of the high-entropy pixel blocks at the top of the sort are forcibly redirected from the first operator link to the second operator link. The asymmetric feature routing unit sends a speed-down blocking signal to the external main control PLC bus, pushing the local high-entropy pixel blocks that have encountered the redirection operation into an independently configured asynchronous high-speed buffer to perform a delayed suspension operation. The system continuously reads the real-time occupancy rate in a loop. Current value pending detection of real-time occupancy rate When the load falls below the safety threshold, the suspended high-entropy pixel blocks are retrieved from the asynchronous cache and redirected to the first operator link. This program ensures that the system processing latency remains within a 10μs closed-loop control cycle by dynamically reducing the proportion of high-resolution feature extraction regions, eliminating the risk of frame drops caused by instantaneous saturation of computing power load. The aforementioned 85% hardware monitoring protection threshold setting decision is derived from data characterization of the bottleneck test of memory bandwidth contention in the underlying computing device. Stress load testing confirms that when the throughput capacity of the multi-threaded computing pool is forced to approach the 88% critical point, the scheduling time required for cache flow addressing increases exponentially, and the resulting data flow congestion will cause the system processing latency to exceed the preset threshold. Lowering the blocking protection threshold to 85% not only ensures the chip's high-efficiency concurrency benefits of over 80%, but also retains a 3% to 5% memory exchange hard isolation buffer layer at the end, preventing the severe industrial consequences of the entire detection platform crashing due to memory deadlock caused by a local high-density defect flood.

[0044] Before performing heterogeneous feature tensor splicing, the graph reconstruction unit obtains the sampling step size ratio between the first operator link and the second operator link. A spatial index mapping model is established. For each feature point in the high-dimensional semantic feature tensor output by the first operator link, the graph reconstruction unit extracts its corresponding original spatial coordinates and calculates the linear offset of the original spatial coordinates in the low-dimensional structural feature tensor. Based on the linear offset, the embedding anchor coordinates of the high-dimensional semantic feature tensor on the structural feature base map are determined. The method for determining the embedding anchor coordinates follows the coordinate mapping criterion of spatially aligning high-frequency semantic components with global structural components. This ensures that local subtle defect features and global background features maintain geometric topological consistency when reconstructing heterogeneous feature maps. This addresses the high-dimensional features that inevitably arise after spatial alignment. To address the physical heterogeneity problem of the mismatch between high-dimensional semantic features and low-dimensional features in terms of channel dimension, the graph reconstruction unit immediately initiates a dimension normalization procedure after determining the embedding anchor points. It calls a bottleneck layer convolution operator with a kernel scale of 1×1 to perform linear reduction of the channel dimension of the high-dimensional semantic feature tensor, compressing the number of deep feature channels to be exactly equal to the number of channels in the low-dimensional structural feature base map. After the dimensionality reduction is completed, the graph reconstruction unit directly performs numerical superposition on the processed high-resolution semantic slices in the form of matrix point-to-point addition and writes them into the grid area selected by the corresponding anchor points of the low-resolution background feature base map. This completes the structural reorganization step of heterogeneous mosaicking at the tensor operation architecture level.

[0045] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. An image feature extraction system for visual inspection, characterized in that, include: The spatial distribution quantization unit is used to obtain the original pixel matrix, divide it into several pixel blocks of the same scale, calculate the pixel gray level distribution probability within each pixel block, and determine the corresponding spatial information entropy based on the pixel gray level distribution probability. An asymmetric feature routing unit, connected to a spatial distribution quantization unit, is used to establish an asymmetric routing topology containing a first operator link and a second operator link. The first operator link is used to extract high-resolution semantic features, and the second operator link is used to extract low-resolution structural features. The asymmetric feature routing unit routes each pixel block to the first operator link or the second operator link based on the comparison result of spatial information entropy and preset decision threshold, and in combination with the phase consistency verification result of pixel blocks in the Fourier transform domain. The graph reconstruction unit is connected to the first operator link and the second operator link respectively. It is used to obtain the original spatial coordinates of the pixel blocks in the original pixel matrix, embed the high-dimensional semantic feature tensor output by the first operator link into the low-dimensional structural feature tensor output by the second operator link according to the original spatial coordinates, and smooth the embedding boundary to output the heterogeneous reconstructed feature graph.

2. The image feature extraction system for visual inspection according to claim 1, characterized in that, The spatial distribution quantization unit provides routing weight constraints for the asymmetric feature routing unit by running noise identification logic based on phase consistency. Specifically, it includes the following steps: Step S201, obtaining the local phase features of each pixel block in the Fourier transform domain; Step S202, determining the physical structure consistency index of the pixel block based on the local phase features; Step S203, comparing the physical structure consistency index with the preset noise discrimination standard. If the physical structure consistency index is lower than the preset noise discrimination standard, the spatial information entropy is determined to be an interference signal, and a forced routing instruction is sent to the asymmetric feature routing unit to guide the corresponding pixel block to the second operator link.

3. The image feature extraction system for visual inspection according to claim 1, characterized in that, The image feature extraction system for visual inspection also includes a recognition unit connected to the map reconstruction unit. The asymmetric feature routing unit also includes a feedback correction module, which is used to obtain the recognition confidence score output by the recognition unit and dynamically correct the preset decision threshold based on the recognition confidence score. If the recognition confidence score is lower than the preset confidence threshold, the preset decision threshold is reduced to increase the routing ratio of the first operator link.

4. The image feature extraction system for visual inspection according to claim 1, characterized in that, The first operator link includes multiple convolution operators with different local receptive fields, used to perform multi-scale mapping on pixel blocks to obtain texture parameters that represent heterogeneous semantic details; the second operator link includes a global pooling operator, used to map pixel blocks into low-dimensional feature vectors that represent background attributes.

5. The image feature extraction system for visual inspection according to claim 1, characterized in that, When determining the spatial information entropy, the spatial distribution quantization unit simultaneously calculates the boundary gradient change rate of pixel blocks and uses the boundary gradient change rate as a weighted compensation factor for the spatial information entropy to correct the spatial information entropy.

6. The image feature extraction system for visual inspection according to claim 1, characterized in that, When dividing pixels into blocks, the spatial distribution quantization unit sets the pixel step size to be equal to the boundary length of the pixel block to seamlessly cover the original pixel matrix; the numerical range of spatial information entropy is normalized to 0 to 1.

7. The image feature extraction system for visual inspection according to claim 1, characterized in that, When embedding high-dimensional semantic feature tensors into low-dimensional structural feature tensors, the graph reconstruction unit uses the low-dimensional structural feature tensor as the global spatial reference and aligns the high-dimensional semantic feature tensor pixels according to the original spatial coordinates to maintain the geometric topological association between heterogeneous semantic features and the background.

8. The image feature extraction system for visual inspection according to claim 1, characterized in that, The graph reconstruction unit performs Gaussian smoothing filtering at the junction of the high-dimensional semantic feature tensor and the low-dimensional structural feature tensor. By adjusting the standard deviation parameter of the Gaussian kernel function, it smooths out feature jumps at the junction.

9. The image feature extraction system for visual inspection according to claim 1, characterized in that, The system also includes an adaptive compensation module, which monitors ambient light intensity fluctuations and applies compensation to a preset decision threshold based on these fluctuations to offset the entropy shift of pixel blocks under unbalanced lighting conditions, thus ensuring that the asymmetric feature routing unit maintains consistent recognition of semantic distribution.

Images