A method and system for identifying a demoulding overturning state of a chain block

Abstract

The application relates to the technical field of image recognition, and discloses a chain block demolding overturning state image recognition method and system. The method establishes a closed-loop control mechanism from image quality evaluation to root cause analysis to parameter self-adaptive adjustment, effectively avoids the problem that in the prior art, optical degradation leads to logical misjudgment, further causes mechanical stress, and finally forms a self-reinforcing failure closed loop.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a method and system for recognizing the image of a chain block in the demolding and flipping state. Background Technology

[0002] In automated industrial production, especially for precast concrete products such as interlocking blocks that are produced in large quantities, ensuring product quality and production efficiency is crucial. Modern production lines typically integrate advanced image recognition systems for real-time quality checks, such as assessing demolding integrity, precise flipping angles, and surface quality.

[0003] However, these sophisticated systems frequently encounter various challenges in dynamic and often harsh industrial environments. One significant and often overlooked problem is the gradual degradation of optical components such as industrial cameras due to the accumulation of airborne particles and moisture. This subtle yet persistent environmental disturbance severely impacts the accuracy and reliability of visual inspection, leading to a series of operational problems and ultimately undermining the effectiveness of automation itself. Summary of the Invention

[0004] This invention provides a method and system for image recognition of the demolding and flipping state of interlocking blocks, aiming to solve the technical problems in the field of automated industrial production, such as the degradation of optical components, the increase in the misjudgment rate of image recognition systems, and the resulting increase in mechanical system stress, which lead to a decrease in the accuracy of product quality inspection, a decrease in production efficiency, and a loss of system reliability.

[0005] The technical solution of this application is as follows: In a first aspect, this application discloses a method for image recognition of the demolding and flipping state of interlocking blocks, specifically including: Images of the interlocking blocks to be inspected are acquired using an industrial camera; Perform visual quality optimization processing on the image, which includes at least one of noise reduction processing, contrast enhancement processing, and motion deblurring processing. Obtain visual quality information of the optimized image, including sharpness, contrast, and motion blur. Based on visual quality information, the feature extraction strategy is adaptively adjusted to obtain the visual features of the linked blocks to be detected. The adjustment of the strategy includes at least one of feature descriptor type switching, edge detection operator sensitivity adjustment, and feature fusion weight adjustment. Determine the quality status of the interlocking block to be detected based on visual characteristics; Based on visual quality information and stability indicators of the feature extraction process, a reliability score is generated for the quality status judgment result. Acquire the qualified product yield, reliability score, sharpness information, and production line operation status information of the production line, including mechanical vibration intensity, lens contamination intensity, and light intensity. Based on the qualified product yield, reliability score, clarity information, and production line operating status information, determine the reasons for the decline in the qualified product yield; Depending on the cause, the operating parameters of the production line are adjusted through the production line parameter control interface. These operating parameters affect the acquisition parameters of the industrial camera, the mechanical motion parameters of the production line, or the lighting parameters.

[0006] Furthermore, the steps for adaptively adjusting the feature extraction strategy include: The image is divided into multiple local regions according to a preset grid division method. The size of each local region is N×N pixels, where N is an integer greater than 1. For each local region, the Laplacian variance of the local region is calculated as the local sharpness, the standard deviation of the pixel grayscale of the local region is calculated as the local contrast, and the width of the motion blur kernel of the local region is calculated as the local motion blur degree. Based on the local sharpness, local contrast, and local motion blur of each local region, the feature extraction strategy for the local region is adjusted in a region-specific manner. Specifically, when the local sharpness is lower than the preset sharpness threshold, the ratio of the high and low thresholds of the Canny operator in the edge detection operator is adjusted from the default 1:2 to 1:3; When the local contrast is lower than the preset contrast threshold, the threshold of Otsu global threshold segmentation is shifted towards the lower grayscale direction by a preset offset amount. When the degree of local motion blur exceeds the preset blur threshold, the feature descriptor is switched from the directional gradient histogram to the scale-invariant feature transformation descriptor, which is scale-invariant to motion blur.

[0007] Based on this, the steps for region-specific adjustment of the feature extraction strategy for local regions also include: Visual quality information of each local region is acquired at a preset sampling period; Calculate the magnitude and frequency of change of visual quality information in each local region within a preset time window; When the magnitude of change is less than the preset first threshold and the frequency of change is greater than the preset second threshold, the feature extraction strategy for the current local region remains unchanged. When the magnitude of change exceeds a preset first threshold and the frequency of change is less than a preset second threshold, the feature extraction strategy for the local region is adjusted. When the magnitude of the change exceeds a preset first threshold and the frequency of the change exceeds a preset second threshold, an abnormal warning signal for the image acquisition system is issued.

[0008] In some preferred embodiments, the steps for determining the cause of a decrease in the yield of qualified products include: During production line downtime, acquire a background image that does not contain interlocking blocks; The gray-level co-occurrence matrix texture features and spot distribution analysis are performed on the background image to obtain the texture feature vector and spot distribution feature vector of the background image, respectively. The texture feature vector and the spot distribution feature vector are input into a pre-trained contamination classifier. The contamination classifier outputs the contamination type and contamination index. The contamination type includes at least one of lens stain contamination, light source attenuation contamination and environmental dust contamination. Obtain the pollution index; When the yield of qualified products is lower than the target threshold, if the contamination index is greater than the preset contamination tolerance threshold, and the area of ​​overlap between the contaminated area and the preset key detection area in the image coordinate system is greater than the preset area threshold, then the reason for the decrease in the yield of qualified products is the decrease in the reliability of the image recognition system.

[0009] As an optional approach, the steps to determine the cause of the decline in the yield of qualified products include: A sliding window trend analysis was performed on the yield of qualified products, reliability score, and clarity information to calculate the first-order difference mean and duration of change of each indicator within a preset time window. The mechanical vibration of the production line is monitored in real time by using an accelerometer to obtain the root mean square value of the vibration intensity in the time domain and the main frequency amplitude in the frequency domain. The degree of contamination on the lens surface is assessed in real time by a light-scattering sensor pre-installed on the side of the camera lens, and the contamination intensity data is obtained. Input the qualified product output rate, reliability score, clarity information, root mean square value of vibration intensity data in the time domain and dominant frequency amplitude in the frequency domain, as well as pollution intensity data into the multi-dimensional correlation analysis model; The multidimensional correlation analysis model calculates the Pearson correlation coefficient matrix between each factor, and the Granger causality test is used to analyze the time lag order of each factor relative to the qualified product output rate. The factor with the largest absolute value of the correlation coefficient and the Granger causality test p value is less than 0.05 is identified as the dominant factor. Based on the dominant factors, determine the reasons for the decline in the yield of qualified products.

[0010] To enhance functionality, the quality status of the chain block to be detected is determined through a convolutional neural network model, which includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and an output layer connected in sequence. Convolutional neural network models are obtained through pre-training using the following steps: Obtain a training sample set containing images of qualified and unqualified linked blocks, with each training sample image labeled with a corresponding quality status label; The training sample set is input into the convolutional neural network model for forward propagation calculation to obtain the predicted quality status. Based on the difference between the predicted quality state and the quality state label, the network parameters of the convolutional neural network are updated using the backpropagation algorithm. The forward and backward propagation steps described above are executed iteratively until the loss function value between the predicted quality state and the quality state label is less than the preset loss threshold.

[0011] To improve the solution, the steps for generating a reliability score for the quality status assessment results include: Obtain sharpness, contrast, and motion blur values ​​from visual quality information; Obtain stability metrics for the feature extraction process, including feature point repetition rate, feature descriptor matching success rate, and standard deviation of feature extraction time. Input the sharpness value, contrast value, motion blur value, feature point repetition rate, feature descriptor matching success rate, and standard deviation of feature extraction time into the pre-trained reliability regression model; The reliability regression model outputs a reliability score as the result of the quality status assessment. The reliability score ranges from 0 to 1.

[0012] For specific situations, the steps to determine the cause of a decrease in the yield of qualified products also include: Obtain the identification information of the current production shift and the identification information of the current operator; Based on the identification information of the production shift and the operator, the benchmark qualified product output rate, benchmark reliability score and benchmark clarity information corresponding to the current production shift and the current operator are extracted from the pre-stored benchmark library of production line operating parameters. The benchmark qualified product yield, benchmark reliability score, and benchmark clarity information are used as reference benchmarks to determine the reasons for the decline in qualified product yield. When the instantaneous value of the qualified product output rate is lower than the benchmark qualified product output rate, the instantaneous value of the reliability score is lower than the benchmark reliability score, and the instantaneous value of the clarity information is lower than the benchmark clarity information, the reason for the decrease in the qualified product output rate is determined to be a decrease in the reliability of the image recognition system.

[0013] To further resolve the issue, the steps to adjust the production line's operating parameters include: When the reason for the decline in the yield of qualified products is determined to be the decline in the reliability of the image recognition system, the priority of parameter adjustment is determined based on the reliability score and sharpness information. If the sharpness information is lower than the first sharpness threshold, the exposure time and gain parameters of the industrial camera will be adjusted first through the production line parameter control interface. If the lens contamination intensity is higher than the first contamination threshold, the automatic lens cleaning device will be activated first through the production line parameter control interface. If the light intensity is lower than the first light threshold or higher than the second light threshold, the lighting parameters will be adjusted first through the production line parameter control interface. If the mechanical vibration intensity is higher than the first vibration threshold, the mechanical movement speed of the production line will be reduced first through the production line parameter control interface.

[0014] Secondly, this application also discloses an image recognition system for the demolding and flipping state of interlocking blocks, specifically including: An industrial camera is set at a predetermined position in the interlocking block demolding and flipping station to acquire images of the interlocking block to be inspected; The image preprocessing module, whose input is connected to the output of an industrial camera, is used to perform visual quality optimization processing on the image; The visual quality information acquisition module has its input end connected to the output end of the image preprocessing module, and is used to acquire the visual quality information of the optimized image. The adaptive feature extraction strategy adjustment module has its first input end connected to the output end of the visual quality information acquisition module, which is used to adjust the feature extraction strategy according to the visual quality information, and its output end outputs the visual features of the chain block to be detected. The image recognition module has its input end connected to the output end of the adaptive feature extraction strategy adjustment module, and is used to determine the quality status of the interlocking block to be detected based on visual features. The reliability score generation module has its first input connected to the output of the visual quality information acquisition module and its second input connected to the output of the adaptive feature extraction strategy adjustment module. It is used to generate a reliability score based on visual quality information and the stability index of the feature extraction process. The production line data acquisition interface is used to collect data on the production line's qualified product yield, mechanical vibration intensity, lens contamination intensity, and light intensity. The root cause analysis module has its first input connected to the output of the reliability score generation module and its second input connected to the output of the production line data acquisition interface. It is used to determine the reasons for the decline in the qualified product output rate based on the qualified product yield, reliability score, clarity information and production line operating status information. The production line parameter control interface has its input end connected to the output end of the root cause analysis module, which is used to adjust the operating parameters of the production line according to the cause. The output terminals of the production line parameter control interface are connected to the control terminals of the industrial camera, the mechanical motion control terminals of the production line, and the lighting control terminals, respectively.

[0015] Beneficial effects The image recognition method and system for the demolding and flipping state of interlocking blocks disclosed in this application, after acquiring an image of the interlocking block to be detected using an industrial camera, first performs visual quality optimization processing on the image, including noise reduction, contrast enhancement, and motion deblurring, effectively improving the usability of the image. Subsequently, it acquires visual quality information such as the sharpness, contrast, and motion blur degree of the optimized image, and adaptively adjusts the feature extraction strategy based on this information, such as switching feature descriptor types, adjusting the sensitivity of edge detection operators, or feature fusion weights, thereby ensuring that stable and reliable visual features are obtained under different image quality conditions. Based on these visual features, the system can accurately determine the quality state of the interlocking block. More importantly, this application introduces a reliability score generation mechanism, which integrates visual quality information and the stability index of the feature extraction process to quantify the reliability of the quality state judgment result.

[0016] Through the above technical solution, this application overcomes the shortcomings of existing image recognition systems, such as susceptibility to interference in dynamic industrial environments, high misjudgment rates, and difficulty in adapting to complex working conditions. It not only improves the accuracy and robustness of interlocking block quality detection, but more importantly, by establishing a closed-loop control mechanism from image quality assessment to root cause analysis and then to adaptive parameter adjustment, it effectively avoids the problem in existing technologies where optical degradation leads to logical misjudgments, which in turn triggers mechanical stress and ultimately forms a self-reinforcing failure loop. Accordingly, this application can significantly improve the overall efficiency and product qualification rate of production lines, reduce rework and recall costs, and provide a more stable, reliable, and intelligent solution for automated industrial production. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for recognizing the demolding and flipping state image of an interlocking block according to an embodiment of the present invention. Figure 2 This is a flowchart of a method for adaptively adjusting feature extraction strategies provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an image recognition system for the demolding and flipping state of interlocking blocks provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 , Figure 1This is a flowchart illustrating an image recognition method for the demolding and flipping state of an interlocking block according to an embodiment of the present invention, including: S11, acquire an image of the interlocking block to be inspected using an industrial camera; S12, perform visual quality optimization processing on the image, the visual quality optimization processing includes at least one of noise reduction processing, contrast enhancement processing and motion deblurring processing; S13, Obtain the visual quality information of the optimized image, including sharpness, contrast and motion blur. S14, based on the visual quality information, adaptively adjust the feature extraction strategy to obtain the visual features of the chain block to be detected, wherein the adjustment of the strategy includes at least one of feature descriptor type switching, edge detection operator sensitivity adjustment, and feature fusion weight adjustment. S15, determine the quality status of the interlocking block to be detected based on the visual features; S16, Based on the visual quality information and the stability index of the feature extraction process, generate a reliability score for the quality state; S17, obtain the qualified product output rate of the production line, the reliability score, the sharpness information and the production line operation status information, wherein the production line operation status information includes mechanical vibration intensity, lens contamination intensity and light intensity; S18, determine the reason for the decrease in the qualified product output rate based on the qualified product output rate, the reliability score, the clarity information, and the production line operation status information; S19, based on the stated reason, adjust the operating parameters of the production line through the production line parameter control interface, wherein the operating parameters act on the acquisition parameters of the industrial camera or the mechanical motion parameters or lighting parameters of the production line.

[0020] This application effectively solves the problems of decreased recognition accuracy and reduced production efficiency of traditional image recognition systems in complex industrial environments caused by factors such as optical degradation and mechanical vibration, by introducing visual quality optimization processing, adaptive feature extraction strategy adjustment, reliability score generation, and root cause analysis and parameter adjustment mechanism based on multi-dimensional information.

[0021] The method proposed in this application aims to improve the accuracy and reliability of image recognition for the demolding and flipping state of interlocking blocks, and to adaptively address common environmental changes and equipment degradation issues in industrial production lines. Here, "interlocking block" refers to a precast concrete product, typically designed with an interlocking structure, widely used in road paving, retaining walls, and other engineering projects. "Demolding and flipping state" refers to the specific posture and surface condition of the interlocking block during production, when it detaches from the mold and flips to enter the next process; this is a crucial step in quality inspection. "Industrial camera" is a high-performance digital camera specifically designed for industrial environments, featuring high resolution, high frame rate, and high stability, commonly used in machine vision systems. "Visual quality optimization processing" refers to a series of preprocessing operations on the image to improve its visual effect and the accuracy of subsequent analysis, such as noise removal, contrast enhancement, and blur elimination. "Visual quality information" refers to indicators that quantify image quality, such as sharpness, contrast, and motion blur. "Feature extraction strategy" refers to the methods and parameter configurations used in image processing to identify and extract meaningful features from an image, such as selecting different feature descriptors and adjusting the parameters of edge detection operators. A "feature descriptor" is a vector used to describe the features of a local region in an image, such as Histogram of Oriented Gradients (HOG) and Scale Invariant Feature Transform (SIFT). An "edge detection operator" is an algorithm used to identify the boundaries of objects in an image, such as the Canny operator and the Sobel operator. A "reliability score" is a quantitative indicator that measures the credibility of image recognition results, typically ranging from 0 to 1. "Production line operating status information" includes mechanical vibration intensity, lens contamination intensity, and illumination intensity, reflecting the environmental and equipment conditions of the production line.

[0022] The method of this application first acquires images of the interlocking block to be inspected using an industrial camera. The industrial camera can be installed at a specific location on the production line, such as above or to the side of the interlocking block demolding and flipping station, to ensure clear capture of images of the interlocking block in its critical state. For example, a high-resolution CCD or CMOS camera equipped with a suitable lens can be used to obtain images with rich detail and minimal distortion. Image acquisition parameters, such as exposure time, gain, and frame rate, can be initially set according to the actual operating speed and lighting conditions of the production line.

[0023] Subsequently, visual quality optimization processing is performed on the acquired images. This processing aims to eliminate or mitigate various degrading factors present in the images to improve the accuracy of subsequent feature extraction and recognition. Visual quality optimization processing may include at least one of denoising, contrast enhancement, and motion deblurring. For example, denoising may employ methods such as Gaussian filtering, median filtering, or bilateral filtering to smooth the image and remove random noise. Contrast enhancement may employ methods such as histogram equalization, adaptive histogram equalization (CLAHE), or gamma correction to improve the visual effect and detail discernibility of the image. Motion deblurring may employ Wiener filtering, blind deconvolution, or deep learning-based deblurring algorithms to recover images blurred due to camera or object motion.

[0024] Next, we obtain the visual quality information of the optimized image. This information is a key metric for evaluating image quality, including sharpness, contrast, and motion blur. Sharpness can be quantified by calculating the image's gradient magnitude, Laplacian variance, or frequency domain energy. Contrast can be quantified by calculating the image's pixel grayscale standard deviation, RMS contrast ratio, or Michelson contrast ratio. Motion blur can be quantified by analyzing the image's Fourier spectrum, calculating the width of the blur kernel, or using a specialized blur estimation algorithm. This visual quality information provides the data foundation for subsequent adaptive adjustments to the feature extraction strategy.

[0025] Based on the acquired visual quality information, the feature extraction strategy is adaptively adjusted to obtain the visual features of the linked blocks to be detected. This adaptive adjustment is one of the key innovations of this application, enabling the feature extraction process to be optimized according to changes in image quality, thereby improving the robustness of recognition. The strategy adjustment may include at least one of feature descriptor type switching, edge detection operator sensitivity adjustment, and feature fusion weight adjustment. For example, when the image sharpness is low, a feature descriptor that is insensitive to blur can be switched to, or the sensitivity of the edge detection operator can be reduced to avoid detecting too many noisy edges. When the image contrast is insufficient, the feature fusion weights can be adjusted to reduce the dependence on texture features and increase the dependence on shape features.

[0026] After acquiring visual features, the quality status of the linked blocks to be detected is determined based on these features. This is typically achieved through pre-trained image recognition models, such as Support Vector Machines (SVM), Decision Trees, or Convolutional Neural Networks (CNNs). The model classifies the linked blocks as "qualified" or "unqualified" based on the extracted visual features, or further subdivides them into specific defect types, such as "missing corners," "cracks," or "abnormal flip angles."

[0027] To assess the reliability of the quality status judgment results, this application also generates a reliability score for the quality status judgment results based on visual quality information and stability indicators of the feature extraction process. The reliability score is a quantitative indicator that reflects the credibility of the current judgment result. For example, when the image quality is poor or the feature extraction process is unstable, the reliability score will be low, indicating that the judgment result may have a high risk of misjudgment. Stability indicators of the feature extraction process may include feature point repetition rate, feature descriptor matching success rate, and the standard deviation of feature extraction time.

[0028] In addition, this application also obtains the production line's yield rate, reliability score, sharpness information, and production line operating status information. The production line operating status information includes mechanical vibration intensity, lens contamination intensity, and illumination intensity. The yield rate is a key indicator for measuring production efficiency. Mechanical vibration intensity can be monitored in real time using an accelerometer. Lens contamination intensity can be assessed using a diffuse light sensor or image analysis methods. Illumination intensity can be measured using an illumination sensor. This multi-dimensional data provides comprehensive information for subsequent root cause analysis.

[0029] Based on the obtained product yield, reliability score, sharpness information, and production line operating status information, the cause of the decline in product yield is determined. This is a crucial diagnostic step aimed at identifying the root cause of the decreased production efficiency. For example, if the product yield declines, and the reliability score and sharpness information also decline significantly, and the lens contamination intensity increases, the cause may be determined to be lens contamination leading to a decrease in the reliability of the image recognition system. If the product yield declines, and the mechanical vibration intensity increases, the cause may be mechanical vibration causing blurring of the interlocking block image.

[0030] Finally, based on the identified cause, the production line's operating parameters are adjusted via the production line parameter control interface. This closed-loop control mechanism enables the system to take corresponding corrective measures based on the diagnostic results, thereby restoring the production line to normal operation. Operating parameters can affect industrial camera acquisition parameters (such as exposure time and gain), production line mechanical motion parameters (such as conveyor belt speed and demolding mechanism frequency), or lighting parameters (such as light source brightness and color temperature). For example, if the cause is determined to be lens contamination, an automatic lens cleaning device can be activated via the control interface; if the cause is determined to be insufficient lighting, the lighting brightness can be increased; if the cause is determined to be mechanical vibration, the production line's operating speed can be appropriately reduced.

[0031] The image recognition method for the demolding and flipping state of interlocking blocks in this application integrates multiple steps such as visual quality optimization, adaptive feature extraction, reliability assessment, multi-dimensional root cause analysis, and closed-loop parameter adjustment to form an intelligent and robust image recognition and production line optimization system.

[0032] In summary, this application effectively overcomes the problems of insufficient accuracy and reliability of image recognition in complex industrial environments through a series of innovative technical means, providing strong technical support for the intelligent manufacturing of precast concrete products such as interlocking blocks.

[0033] In some implementations of the above methods, the adaptive feature extraction strategy is based on the overall visual quality information of the image. However, in actual industrial production environments, the visual quality of images is often not uniformly distributed. For example, due to factors such as uneven local illumination, mechanical vibration, or local lens contamination, different regions of the image may exhibit different levels of sharpness, contrast, or blurriness. If the strategy is adjusted globally based solely on the overall visual quality information of the image, it may not be possible to specifically optimize the feature extraction effect in local regions, thereby affecting the final recognition accuracy and reliability.

[0034] In this regard, refer to Figure 2 , Figure 2 This is a flowchart of a method for adaptively adjusting feature extraction strategies provided in an embodiment of the present invention. S14 includes: S141, the image is divided into multiple local regions according to a preset grid division method, and the size of each local region is N×N pixels, where N is an integer greater than 1; S142, for each of the local regions, calculate the Laplacian variance of the local region as the local sharpness, calculate the standard deviation of the pixel grayscale of the local region as the local contrast, and calculate the width of the motion blur kernel of the local region as the local motion blur degree. S143, based on the local sharpness, local contrast and local motion blur of each local region, the feature extraction strategy of the local region is adjusted in a region-specific manner; Wherein, when the local sharpness is lower than the preset sharpness threshold, the ratio of the high and low thresholds of the Canny operator of the edge detection operator is adjusted from the default 1:2 to 1:3; When the local contrast is lower than the preset contrast threshold, the threshold of the Otsu global threshold segmentation is shifted towards the lower grayscale direction by a preset offset amount. When the degree of local motion blur is higher than a preset blur threshold, the feature descriptor is switched from the directional gradient histogram to a scale-invariant feature transformation descriptor that is scale-invariant to motion blur.

[0035] Specifically, to achieve refined analysis of local image regions, the image is divided into multiple local regions according to a preset grid division method. The size of each local region is set to N×N pixels, where N is an integer greater than 1. Its specific value can be flexibly configured based on the size of the interlocking blocks to be detected, the image resolution, and the sensitivity requirements to local quality changes in the actual application scenario. For example, N can be 16, 32, or 64, etc., to ensure that each local region can effectively reflect its internal visual quality characteristics.

[0036] For each local region, its visual quality information is calculated independently. Local sharpness is obtained by calculating the Laplacian variance of that local region. The Laplacian operator is sensitive to gray-level changes in the image; a larger variance generally indicates richer image details, i.e., higher sharpness. Local contrast is characterized by calculating the standard deviation of pixel gray levels in that local region. A larger standard deviation indicates a wider gray-level distribution and higher contrast in that region. The degree of local motion blur is determined by estimating the width of the motion blur kernel in that local region; a larger motion blur kernel width indicates more severe motion blur in that region.

[0037] In practical applications, the feature extraction strategy for each local region is adjusted region-specifically based on the calculated local sharpness, local contrast, and local motion blur. Specifically, when the local sharpness is below a preset sharpness threshold, the high-low threshold ratio of the Canny operator in the edge detection operator is adjusted from the default 1:2 to 1:3 to better extract edge information from blurred areas. This adjustment allows the Canny operator to detect more weak edges in low-sharp regions, thus avoiding the loss of important structural features due to image blur. When the local contrast is below a preset contrast threshold, the threshold of Otsu global thresholding is shifted towards lower grayscale values ​​by a preset amount to achieve more accurate image segmentation. This helps to better separate the target from the background in low-contrast images, especially when the target object has a low grayscale value or the background has a high grayscale value. When the local motion blur exceeds a preset blur threshold, to address the impact of motion blur on feature description, the feature descriptor is switched from a histogram of oriented gradients to a scale-invariant feature transform descriptor that is scale-invariant to motion blur. The directional gradient histogram descriptor is sensitive to the direction and magnitude of image gradients, and its performance may degrade under motion blur conditions. In contrast, the scale-invariant feature transform descriptor, due to its scale invariance and rotation invariance, can still maintain good feature matching ability when there is a certain degree of motion blur.

[0038] This application's solution overcomes the limitations of traditional global visual quality assessment by meticulously dividing an image into multiple local regions and independently evaluating the visual quality of each region. Through this technical solution, the application significantly improves the robustness and accuracy of the image recognition method for the demolding and flipping state of interlocking blocks in complex industrial environments. Compared to solutions that rely solely on global visual quality information for adjustment, this application optimizes refined, region-specific feature extraction strategies for local visual quality problems in different regions of the image (such as local blurring, local low contrast, etc.). Therefore, even under conditions of uneven lighting, mechanical vibration, or local lens contamination leading to non-uniform image quality distribution, it ensures the extraction of more accurate and stable visual features from each key region, effectively reducing false recognition and false rejection rates and improving the reliability of the final quality status judgment. This fine-grained adaptive adjustment capability makes the entire image recognition system more adaptable to environmental changes, further ensuring the qualified product output rate of the production line.

[0039] In some of the embodiments described above in this application, a scheme for region-specific adjustment of feature extraction strategies based on visual quality information of local regions is proposed. However, in real-world production environments, the visual quality information of images may not be static but rather dynamically change due to various factors. If strategy adjustments are made solely based on instantaneous visual quality information, it may lead to overly frequent and unstable adjustments, or failure to effectively identify potential system anomalies, thereby affecting the accuracy of image recognition and the overall stability of the system.

[0040] In this regard, this application further proposes that the steps for region-specific adjustment of the aforementioned feature extraction strategy for local regions also include: Visual quality information of each local region is acquired at a preset sampling period; Calculate the magnitude and frequency of change of visual quality information in each local region within a preset time window; When the change magnitude is less than a preset first threshold and the change frequency is greater than a preset second threshold, the feature extraction strategy for the current local region remains unchanged. When the magnitude of the change is greater than a preset first threshold and the frequency of the change is less than a preset second threshold, the feature extraction strategy for the local region is adjusted. When the magnitude of the change exceeds a preset first threshold and the frequency of the change exceeds a preset second threshold, an abnormal warning signal for the image acquisition system is issued.

[0041] Specifically, acquiring visual quality information for each local region at a preset sampling period means that the system does not acquire visual quality data at a single point in time, but rather periodically and continuously samples visual quality indicators such as sharpness, contrast, and motion blur for each local region. This preset sampling period can be set according to the production line's operating speed, image acquisition frequency, and system response speed requirements; for example, it can be set to acquire data once per second or once per frame. The purpose is to accumulate sufficient time-series data for subsequent analysis of visual quality trends.

[0042] The calculation of the amplitude and frequency of changes in visual quality information in each local area within a preset time window can be understood as performing statistical analysis on the collected time-series data. The amplitude of change can be measured using indicators such as the difference between the maximum and minimum values, standard deviation, or mean absolute deviation, reflecting the fluctuation range of the visual quality index over a period of time. The frequency of change can be obtained by measuring the number of times the index value crosses a certain threshold per unit time, or through frequency domain analysis methods such as Fourier transform, reflecting the speed of fluctuation in the visual quality index. The length of the preset time window should be selected according to the actual application scenario and the required time scale of analysis; for example, it can be set to data from the most recent 5 seconds, 10 seconds, or longer. Its purpose is to quantify the dynamic characteristics of visual quality and provide a basis for intelligent decision-making.

[0043] In practical applications, when the magnitude of the change is less than a preset first threshold and the frequency of the change is greater than a preset second threshold, the feature extraction strategy for the current local region remains unchanged. This typically corresponds to rapid but minor fluctuations in visual quality, such as those caused by slight random noise. In such cases, frequent adjustments to the strategy may introduce instability, so maintaining the current strategy is chosen to preserve the robustness of the system.

[0044] Furthermore, when the magnitude of the change exceeds a preset first threshold and the frequency of the change is less than a preset second threshold, the feature extraction strategy for the local region is adjusted. This indicates a significant and continuous change in visual quality, such as due to slow changes in lighting conditions or slight lens contamination. In this case, the system needs to adjust its strategy accordingly to adapt to the new image quality conditions and ensure the effectiveness of feature extraction.

[0045] Furthermore, when the magnitude of the change exceeds a preset first threshold and the frequency of the change exceeds a preset second threshold, an abnormal warning signal for the image acquisition system is issued. This situation typically indicates that the image acquisition system may have a serious, rapidly changing anomaly, such as severe vibration of an industrial camera, sudden flickering of the light source, or severe obstruction of the lens. In this case, the system not only needs to adjust its strategy but also needs to issue a timely warning to the operator so that manual intervention or troubleshooting can be carried out to prevent a sharp increase in the recognition error rate.

[0046] This application's solution overcomes the limitations of adjusting strategies solely based on instantaneous visual quality by introducing time-series analysis of local visual quality information. Through this technical solution, the application achieves more intelligent and stable adaptive adjustment of feature extraction strategies in image recognition systems. This solution not only effectively addresses the dynamic changes in image visual quality in production environments, avoiding frequent and unstable strategy switching due to instantaneous fluctuations, but also promptly identifies and warns of potential anomalies in the image acquisition system. Consequently, it significantly improves the robustness, reliability, and fault diagnosis capabilities of the image recognition method for the interlocking block demolding and flipping state, ensuring that the image recognition system can continuously and stably provide high-precision judgment results in complex and ever-changing industrial environments, thereby effectively guaranteeing the qualified product output rate of the production line.

[0047] In some embodiments described above, this application proposes determining the cause of a decline in the yield of qualified products based on the yield rate, reliability score, clarity information, and production line operating status information. However, in actual production environments, a decline in the yield rate of qualified products may be caused by a variety of complex factors, among which a decline in the reliability of image recognition systems is often closely related to environmental pollution or equipment degradation. Relying solely on macroscopic indicators may make it difficult to accurately identify the specific type of pollution and its impact on the image recognition system, leading to inefficient root cause analysis or inaccurate diagnosis.

[0048] In this regard, this application further proposes the following steps for determining the reasons for the decline in the yield of qualified products: During production line downtime, acquire a background image that does not contain interlocking blocks; The background image is subjected to gray-level co-occurrence matrix texture feature extraction and spot distribution analysis to obtain the texture feature vector and spot distribution feature vector of the background image, respectively. The texture feature vector and the light spot distribution feature vector are input into a pre-trained pollution classifier. The pollution classifier outputs a pollution type and a pollution index. The pollution type includes at least one of lens stain pollution, light source attenuation pollution, and environmental dust pollution. Obtain the pollution index; When the yield of qualified products is lower than the target threshold, if the contamination index is greater than the preset contamination tolerance threshold, and the overlap area between the contaminated area and the preset key detection area in the image coordinate system is greater than the preset area threshold, then the reason for the decrease in the yield of qualified products is the decrease in the reliability of the image recognition system.

[0049] Specifically, acquiring background images that do not contain interlocking blocks during production line downtime aims to eliminate interference from the characteristics of the interlocking blocks themselves, thereby allowing focus on detecting abnormal conditions in the production line environment or the image acquisition equipment itself. This downtime can be any point in time during production line shift changes, maintenance, or brief production stoppages.

[0050] The purpose of extracting gray-level co-occurrence matrix (GLCM) texture features and analyzing light spot distribution in the background image is to quantify potential contamination or lighting anomalies. GLCM texture features describe the statistical characteristics of pixel gray-level spatial distribution in an image and are highly sensitive to detecting texture changes caused by lens smudges, environmental dust, etc. Light spot distribution analysis focuses on detecting abnormal light spots or uneven brightness in the image, which is usually related to light source attenuation or local occlusion. These two analytical methods comprehensively obtain the texture feature vector and light spot distribution feature vector of the background image, providing a rich data foundation for subsequent contamination classification.

[0051] In practical applications, the texture feature vector and the light spot distribution feature vector are input into a pre-trained pollution classifier. This pollution classifier can be built based on machine learning or deep learning models, such as support vector machines, random forests, or convolutional neural networks. The classifier is trained on a large number of background images labeled with pollution type and pollution index, enabling it to accurately identify the type of pollution, such as lens smudges, light source attenuation pollution, or environmental dust pollution, and output a quantified pollution index. The pollution index can be understood as the severity or confidence level of the pollution; a higher value indicates more severe pollution.

[0052] Furthermore, when the yield of qualified products falls below a preset target threshold, the system will combine a contamination index for judgment. If the contamination index is greater than a preset contamination tolerance threshold, it indicates a significant contamination problem. Simultaneously, to ensure that contamination truly affects the accuracy of image recognition, it is also necessary to determine the overlap area between the contaminated area and a preset key detection area in the image coordinate system. The key detection area refers to the region in the image that is crucial for judging the quality of interlocking blocks. When the overlap area is greater than a preset area threshold, it can be clearly determined that the decrease in the yield of qualified products is due to a decrease in the reliability of the image recognition system, and that this decrease is caused by specific contamination.

[0053] The solution proposed in this application effectively addresses the blind spots that may exist in traditional methods when identifying the specific reasons for the decline in the yield of qualified products by introducing specialized analysis of background images during idle periods of the production line.

[0054] This application further proposes steps for determining the cause of a decline in the yield of qualified products, including: A sliding window trend analysis was performed on the yield of qualified products, reliability score, and clarity information to calculate the first-order difference mean and duration of change of each indicator within a preset time window. The mechanical vibration of the production line is monitored in real time by using an accelerometer to obtain the root mean square value of the vibration intensity in the time domain and the main frequency amplitude in the frequency domain. The degree of contamination on the lens surface is assessed in real time by a light-scattering sensor pre-installed on the side of the camera lens, and the contamination intensity data is obtained. Input the qualified product output rate, reliability score, clarity information, root mean square value of vibration intensity data in the time domain and dominant frequency amplitude in the frequency domain, as well as pollution intensity data into the multi-dimensional correlation analysis model; The Pearson correlation coefficient matrix between each factor is calculated using the multidimensional correlation analysis model, and the time lag order of each factor relative to the qualified product output rate is analyzed by Granger causality test. The factor with the largest absolute value of the correlation coefficient and the p value of Granger causality test is less than 0.05 is identified as the dominant factor. Based on the aforementioned dominant factors, determine the reasons for the decline in the yield of qualified products.

[0055] Specifically, sliding window trend analysis of qualified product output rate, reliability score, and clarity information involves applying a fixed-size time window to a continuous data stream and calculating the first-order difference mean and duration of change within the window. The first-order difference mean reflects the average rate and direction of change of the indicator over a short period, such as whether the output rate is increasing or decreasing, and by how much; the duration of change indicates the stability or persistence of this trend. Its purpose is to capture the dynamic characteristics of these key indicators, providing more time-series input for subsequent correlation analysis.

[0056] Real-time monitoring of mechanical vibration on the production line using accelerometers involves installing accelerometers on key mechanical components of the production line to continuously collect vibration data. The root mean square value of the vibration intensity data in the time domain quantifies the overall energy of the vibration, reflecting the degree of mechanical wear or imbalance; the dominant frequency amplitude in the frequency domain reveals the vibration intensity at a specific frequency, helping to identify specific fault modes, such as bearing damage or gear meshing problems. The aim is to obtain refined data on the mechanical operating status of the production line, which is a crucial factor affecting the stability of the image recognition system and product quality.

[0057] In practical applications, the real-time assessment of lens surface contamination using a diffused light sensor pre-installed on the side of the camera lens involves detecting the intensity of scattered light caused by contaminants such as dust and oil on the lens surface. The contamination intensity data directly reflects the cleanliness of the lens surface; lens contamination directly affects image acquisition quality and consequently, recognition accuracy. The aim is to provide a non-invasive, real-time lens contamination monitoring method to supplement production line operating status information.

[0058] Furthermore, the time-domain root mean square values ​​and frequency-domain dominant frequency amplitudes of the qualified product output rate, clarity information, vibration intensity data, and pollution intensity data are input into a multi-dimensional correlation analysis model. This model aims to process these multi-source heterogeneous data and uncover the deep relationships between them. Specifically, the multi-dimensional correlation analysis model calculates the Pearson correlation coefficient matrix between each factor to quantify the degree of linear correlation between any two variables; the larger the absolute value of the correlation coefficient, the stronger the linear relationship. In addition, Granger causality tests are used to analyze the time lag order of each factor relative to the qualified product output rate. The Granger causality test is a statistical method used to determine whether one time series can predict another, i.e., whether a causal relationship exists (not just a correlation), and to reveal the time lag of this causal relationship. Its purpose is to identify the dominant factors that have a significant impact on the decline in the qualified product output rate and have a causal relationship, considering the dynamic characteristics of the time series from multiple dimensions.

[0059] The proposed solution introduces sliding window trend analysis, which can dynamically capture the changing trends of qualified product output rate, reliability score and clarity information, rather than relying solely on instantaneous values, thereby providing a more time-series data foundation for subsequent correlation analysis.

[0060] This application further proposes that the aforementioned determination of the quality status of the chain block to be detected is achieved through a convolutional neural network model. This convolutional neural network model includes, in sequence, an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and an output layer.

[0061] Specifically, the input layer receives preprocessed image data of the linked blocks to be detected. The first and second convolutional layers are responsible for extracting features at different levels from the input image, such as edges, textures, and shapes. The convolution operation extracts features from the image by sliding the convolution kernel, generating feature maps. The first and second pooling layers downsample the feature maps output by the convolutional layers to reduce feature dimensionality, improve the computational efficiency of the model, and enhance the model's robustness to local image deformations. The fully connected layer integrates the features output by the pooling layers and maps them to the final classification space. The output layer generates a quality status judgment result for the linked blocks to be detected, such as "qualified" or "unqualified," based on the output of the fully connected layer.

[0062] The convolutional neural network model is obtained through pre-training via the following steps: First, a training sample set containing images of valid and invalid linked blocks is obtained, where each training sample image is labeled with a corresponding quality status label. These labels explicitly indicate the true quality status of the linked blocks in each image.

[0063] Subsequently, the training sample set is input into the convolutional neural network model for forward propagation calculation to obtain the model's prediction quality status for each training sample.

[0064] Based on the difference between the predicted quality state and the quality state label, the network parameters of the convolutional neural network model are updated using the backpropagation algorithm. The backpropagation algorithm calculates the gradient according to the loss function (e.g., cross-entropy loss) and uses an optimizer (e.g., stochastic gradient descent, Adam, etc.) to adjust network parameters such as convolutional kernel weights and biases to minimize the error between the predicted result and the true label.

[0065] The forward and backward propagation steps described above are performed iteratively until the loss function value between the predicted quality state and the quality state label is less than a preset loss threshold. This means that the model has fully learned the feature patterns in the training data and achieved the expected classification performance.

[0066] This application's solution introduces a convolutional neural network model to determine the quality status of the interlocking blocks to be detected, effectively overcoming the limitations of traditional methods in handling complex defect patterns and varying image conditions. Through this technical solution, the application significantly improves the accuracy and robustness of interlocking block quality status assessment. The convolutional neural network model can effectively identify subtle defects and complex texture features that are difficult for traditional methods to capture, reducing the false positive and false negative rates. Furthermore, the model has good generalization ability; after sufficient training, it can adapt to changes in interlocking block images under different batches and production conditions, thereby improving the intelligence level and automated detection efficiency of the entire image recognition system and providing a more reliable technical guarantee for product quality control on the production line.

[0067] The steps described above for generating the reliability score of the quality status assessment result can be further refined.

[0068] Specifically, the steps for generating a reliability score for the quality status assessment result include: Obtain sharpness, contrast, and motion blur values ​​from visual quality information; Obtain stability metrics for the feature extraction process, including feature point repetition rate, feature descriptor matching success rate, and standard deviation of feature extraction time; The sharpness value, the contrast value, the motion blur degree value, the feature point repetition rate, the feature descriptor matching success rate, and the standard deviation of the feature extraction time are input into the pre-trained reliability regression model. The reliability regression model outputs a reliability score for the quality status, and the reliability score ranges from 0 to 1.

[0069] Among these, the sharpness, contrast, and motion blur values ​​in visual quality information are quantitative indicators obtained through analysis of the optimized image. For example, sharpness values ​​can be obtained by calculating the image's gradient magnitude, Laplacian variance, or high-frequency energy in the frequency domain; contrast values ​​can be obtained by calculating the standard deviation of the image's pixel grayscale values ​​or local contrast measures; and motion blur values ​​can be obtained by estimating the image's motion blur kernel width or point spread function parameters. These values ​​provide an objective assessment of the image's inherent quality.

[0070] Furthermore, stability metrics for the feature extraction process aim to quantify the robustness and consistency of the feature extraction algorithm under different image conditions. Feature point repetition rate refers to the proportion of identical feature points detected in different images or under different processing conditions of the same image, reflecting the stability of feature detection. Feature descriptor matching success rate refers to the proportion of correctly matched descriptors out of the total matching attempts during the feature matching process, reflecting the discriminative power and robustness of the feature descriptors. The standard deviation of feature extraction time measures the variability in the time consumed by the feature extraction process, reflecting the stability of the algorithm in actual operation.

[0071] Specifically, a reliability regression model is a pre-trained machine learning model, such as support vector regression, random forest regression, or neural network regression. This model learns the relationship between a large number of image samples and their corresponding visual quality information, feature extraction stability metrics, and manually evaluated reliability scores. This allows it to predict a reliability score for the quality status assessment based on new input data. The reliability score is a value between 0 and 1, where 1 represents extremely high reliability and 0 represents extremely low reliability, providing a quantitative basis for subsequent decision-making.

[0072] The proposed solution generates a reliability score for the quality status assessment result by comprehensively considering the visual quality information of the image itself and the stability index of the feature extraction process. This technical solution provides a more refined and quantifiable mechanism for evaluating the reliability of quality status assessment results.

[0073] In some embodiments described above, this application proposes determining the cause of a decline in the yield of qualified products based on the yield rate, reliability score, image clarity information, and production line operating status information. However, in actual production environments, the operating status of the production line, the operator's skill level, and the characteristics of production shifts may inherently affect the yield rate, reliability score, and image clarity. If judgment is based solely on real-time data, it may be difficult to accurately distinguish between an actual decline in system performance and normal fluctuations due to specific shifts or operators, leading to misjudgments or delayed diagnosis.

[0074] In this regard, this application further proposes that the steps for determining the reasons for the decline in the yield of qualified products also include: Obtain the identification information of the current production shift and the identification information of the current operator; Based on the identification information of the production shift and the identification information of the operator, the benchmark qualified product output rate, benchmark reliability score and benchmark clarity information corresponding to the current production shift and the current operator are extracted from the pre-stored production line operation parameter benchmark library. The benchmark qualified product output rate, the benchmark reliability score, and the benchmark clarity information are used as reference benchmarks to determine the reasons for the decline in the qualified product output rate. When the instantaneous value of the qualified product output rate is lower than the benchmark qualified product output rate, the instantaneous value of the reliability score is lower than the benchmark reliability score, and the instantaneous value of the sharpness information is lower than the benchmark sharpness information, it is determined that the reason for the decrease in the qualified product output rate is the decrease in the reliability of the image recognition system.

[0075] Specifically, obtaining the identification information of the current production shift and the current operator means that the system interacts with the production management system or the operator login system to obtain the currently running production shift number and the unique identifier of the operator currently responsible for that production line. This identification information is a key input for establishing personalized benchmarks.

[0076] The pre-stored production line operating parameter benchmark library can be understood as a database or data model containing historical production data, categorized and statistically analyzed according to different production shifts and operators. For example, for each shift and operator combination, the benchmark library stores the average yield of qualified products under normal operating conditions, the average reliability score of image recognition, and the average image sharpness information. Extracting this benchmark information aims to provide a personalized reference standard for the current production status, consistent with historical performance.

[0077] In practical applications, using the benchmark qualified product output rate, the benchmark reliability score, and the benchmark clarity information as reference benchmarks for determining the cause of the decline in qualified product output rate means that the system no longer relies solely on preset fixed thresholds, but dynamically evaluates performance based on the specific context of the current production environment (shift and operator). When the instantaneous values ​​of the qualified product output rate, the reliability score, and the clarity information are all lower than their corresponding benchmark values, it indicates that the current production status is below the historical normal levels for that specific shift and operator in multiple key indicators. This simultaneous decline in multiple indicators, especially when the reliability score and clarity information, which are directly related to the image recognition system, also decline, can more accurately point to the decline in the reliability of the image recognition system as the main reason for the decline in the qualified product output rate.

[0078] This application's solution establishes dynamic and personalized reference benchmarks for qualified product yield, reliability score, and clarity information by introducing production shift and operator identification information and combining it with a pre-stored benchmark library of operating parameters. When the system detects that the current production indicators (qualified product yield, reliability score, and clarity information) are simultaneously lower than their corresponding personalized benchmarks, it can more accurately determine that the decline in qualified product yield is due to a decrease in the reliability of the image recognition system. This benchmark-based comparison mechanism effectively avoids misjudgments caused by inherent differences between different shifts or operators, making problem diagnosis more accurate. By judging the simultaneous decline of multiple related indicators, the confidence of the diagnostic results is enhanced, ensuring timely and accurate identification of reliability problems in the image recognition system.

[0079] This application further proposes the following steps for adjusting the operating parameters of the production line: When it is determined that the decrease in the yield of qualified products is due to a decrease in the reliability of the image recognition system, the priority of parameter adjustment is determined based on the reliability score and the clarity information. If the sharpness information is lower than the first sharpness threshold, the exposure time and gain parameters of the industrial camera are adjusted preferentially through the production line parameter control interface. If the lens contamination intensity is higher than the first contamination threshold, the automatic lens cleaning device will be activated first through the production line parameter control interface; If the light intensity is lower than the first light threshold or higher than the second light threshold, the lighting parameters are adjusted first through the production line parameter control interface; If the mechanical vibration intensity is higher than the first vibration threshold, the mechanical movement speed of the production line will be reduced preferentially through the production line parameter control interface.

[0080] Specifically, when the system determines that the decline in the yield of qualified products is due to a decrease in the reliability of the image recognition system, it will comprehensively consider both the reliability score and sharpness information to determine the priority of parameter adjustments. The reliability score can be understood as the credibility of the image recognition results, while sharpness information directly reflects the image quality. By analyzing these two key indicators, the most pressing problem that needs to be solved can be more accurately identified, thereby guiding subsequent parameter adjustments.

[0081] The first sharpness threshold is a preset standard value used to measure whether the image sharpness meets the recognition requirements. When the actual sharpness information acquired is lower than this threshold, it indicates a serious problem with the image quality. In this case, the exposure time and gain parameters of the industrial camera should be adjusted first. Exposure time refers to the length of time the industrial camera's photosensitive element receives light, and the gain parameter refers to the degree to which the image signal is amplified. By adjusting these two parameters, the brightness and signal-to-noise ratio of the image can be directly affected, thereby improving the image sharpness.

[0082] Furthermore, the first contamination threshold is used to determine whether the lens contamination level meets the standard for requiring cleaning. When the lens contamination intensity exceeds this threshold, it indicates that there are stains on the lens surface, severely affecting image acquisition quality. In this case, the automatic lens cleaning device is activated first through the production line parameter control interface to physically remove contaminants from the lens surface and restore the clarity of image acquisition.

[0083] Furthermore, the first and second illumination thresholds represent the lower and upper limits of illumination intensity, respectively. When the illumination intensity is below the first threshold or above the second threshold, it indicates that the lighting environment of the production line is unsuitable for image recognition. Insufficient illumination will result in an overly dark image, making details difficult to discern; excessive illumination may lead to overexposure and information loss. In this case, the lighting parameters should be adjusted first through the production line parameter control interface, such as adjusting the brightness, color temperature, or illumination angle of the light source, to ensure stable and suitable illumination conditions for image acquisition.

[0084] Furthermore, the first vibration threshold is a standard used to measure whether the intensity of mechanical vibration exceeds an acceptable range. When the mechanical vibration intensity exceeds this threshold, it indicates excessive vibration in the production line, which may lead to blurred images or inaccurate positioning. In this case, the mechanical movement speed of the production line is reduced first through the production line parameter control interface to minimize the impact of vibration on image acquisition and recognition, thereby improving the accuracy and stability of image recognition.

[0085] The solution proposed in this application effectively addresses the potential blindness and inefficiency in parameter adjustment in the basic solution by introducing a parameter adjustment priority mechanism based on reliability scores and clarity information, as well as a refined adjustment strategy for different specific reasons.

[0086] refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of an image recognition system for the demolding and flipping state of interlocking blocks provided in an embodiment of the present invention, including: An industrial camera is set at a predetermined position in the interlocking block demolding and flipping station to acquire images of the interlocking block to be inspected; An image preprocessing module, whose input is connected to the output of the industrial camera, is used to perform visual quality optimization processing on the image; The visual quality information acquisition module has its input end connected to the output end of the image preprocessing module, and is used to acquire the visual quality information of the optimized image. An adaptive feature extraction strategy adjustment module has its first input end connected to the output end of the visual quality information acquisition module, and is used to adjust the feature extraction strategy according to the visual quality information. Its output end outputs the visual features of the chain block to be detected. An image recognition module, whose input is connected to the output of the adaptive feature extraction strategy adjustment module, is used to determine the quality status of the interlocking block to be detected based on the visual features. A reliability score generation module has a first input terminal connected to the output terminal of the visual quality information acquisition module and a second input terminal connected to the output terminal of the adaptive feature extraction strategy adjustment module. It is used to generate a reliability score based on the visual quality information and the stability index of the feature extraction process. The production line data acquisition interface is used to collect data on the production line's qualified product yield, mechanical vibration intensity, lens contamination intensity, and light intensity. The root cause analysis module has a first input terminal connected to the output terminal of the reliability score generation module and a second input terminal connected to the output terminal of the production line data acquisition interface. It is used to determine the reasons for the decline in the qualified product output rate based on the qualified product yield rate, the reliability score, the clarity information, and the production line operating status information. The production line parameter control interface has its input end connected to the output end of the root cause analysis module, and is used to adjust the operating parameters of the production line according to the cause. The output terminals of the production line parameter control interface are respectively connected to the control terminal of the industrial camera, the mechanical motion control terminal of the production line, and the lighting control terminal.

[0087] The image recognition system for the demolding and flipping state of interlocking blocks proposed in this application significantly improves the robustness and adaptability of the system through its modular architecture and intelligent processing flow. This system not only effectively addresses fluctuations in image quality through image preprocessing and adaptive feature extraction strategy adjustment modules, ensuring the acquisition of high-quality visual features under various complex conditions, but also quantitatively evaluates the credibility of the recognition results through a reliability score generation module, avoiding blind acceptance of low-reliability judgments.

[0088] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for recognizing the image of the demolding and flipping state of interlocking blocks, characterized in that, include: Images of the interlocking blocks to be inspected are acquired using an industrial camera; The image is subjected to visual quality optimization processing, which includes at least one of noise reduction processing, contrast enhancement processing, and motion deblurring processing. Obtain visual quality information of the optimized image, including sharpness, contrast, and motion blur. Based on the visual quality information, the feature extraction strategy is adaptively adjusted to obtain the visual features of the chain block to be detected. The adjustment of the strategy includes at least one of feature descriptor type switching, edge detection operator sensitivity adjustment, and feature fusion weight adjustment. The quality status of the interlocking block to be detected is determined based on the visual characteristics. Based on the visual quality information and the stability index of the feature extraction process, a reliability score for the quality state is generated; The qualified product output rate of the production line, the reliability score, the sharpness information, and the production line operation status information are obtained, wherein the production line operation status information includes mechanical vibration intensity, lens contamination intensity, and light intensity. Based on the qualified product output rate, the reliability score, the clarity information, and the production line operating status information, determine the reasons for the decrease in the qualified product output rate; Based on the stated reasons, the operating parameters of the production line are adjusted through the production line parameter control interface. These operating parameters are applied to the acquisition parameters of the industrial camera, the mechanical motion parameters of the production line, or the lighting parameters.

2. The method according to claim 1, characterized in that, The steps of the adaptive feature extraction strategy include: The image is divided into multiple local regions according to a preset grid division method. The size of each local region is N×N pixels, where N is an integer greater than 1. For each of the local regions, the Laplacian variance of the local region is calculated as the local sharpness, the standard deviation of the pixel grayscale of the local region is calculated as the local contrast, and the width of the motion blur kernel of the local region is calculated as the local motion blur degree. Based on the local sharpness, local contrast, and local motion blur of each local region, the feature extraction strategy for the local region is adjusted in a region-specific manner. Wherein, when the local sharpness is lower than the preset sharpness threshold, the ratio of the high and low thresholds of the Canny operator of the edge detection operator is adjusted from the default 1:2 to 1:3; When the local contrast is lower than the preset contrast threshold, the threshold of the Otsu global threshold segmentation is shifted towards the lower grayscale direction by a preset offset amount. When the degree of local motion blur is higher than a preset blur threshold, the feature descriptor is switched from the directional gradient histogram to a scale-invariant feature transformation descriptor that is scale-invariant to motion blur.

3. The method according to claim 2, characterized in that, The step of performing region-specific adjustments to the feature extraction strategy for the local region further includes: Visual quality information of each local region is acquired at a preset sampling period; Calculate the magnitude and frequency of change of visual quality information of each local region within a preset time window; When the change magnitude is less than a preset first threshold and the change frequency is greater than a preset second threshold, the feature extraction strategy for the current local region remains unchanged. When the magnitude of the change is greater than a preset first threshold and the frequency of the change is less than a preset second threshold, the feature extraction strategy for the local region is adjusted. When the magnitude of the change exceeds a preset first threshold and the frequency of the change exceeds a preset second threshold, an abnormal warning signal for the image acquisition system is issued.

4. The method according to claim 1, characterized in that, The steps for determining the reason for the decrease in the yield of qualified products include: During production line downtime, acquire a background image that does not contain interlocking blocks; The background image is subjected to gray-level co-occurrence matrix texture feature extraction and spot distribution analysis to obtain the texture feature vector and spot distribution feature vector of the background image, respectively. The texture feature vector and the light spot distribution feature vector are input into a pre-trained pollution classifier. The pollution classifier outputs a pollution type and a pollution index. The pollution type includes at least one of lens stain pollution, light source attenuation pollution, and environmental dust pollution. Obtain the pollution index; When the yield of qualified products is lower than the target threshold, if the contamination index is greater than the preset contamination tolerance threshold, and the overlap area between the contaminated area and the preset key detection area in the image coordinate system is greater than the preset area threshold, then the reason for the decrease in the yield of qualified products is the decrease in the reliability of the image recognition system.

5. The method according to claim 1, characterized in that, The steps for determining the reason for the decrease in the yield of qualified products include: A sliding window trend analysis was performed on the qualified product output rate, the reliability score, and the clarity information to calculate the first-order difference mean and the duration of change of each indicator within a preset time window. The mechanical vibration of the production line is monitored in real time by using an accelerometer to obtain the root mean square value of the vibration intensity in the time domain and the main frequency amplitude in the frequency domain. The degree of contamination on the lens surface is assessed in real time by a light-scattering sensor pre-installed on the side of the camera lens, and the contamination intensity data is obtained. The qualified product output rate, the reliability score, the clarity information, the root mean square value of the vibration intensity data in the time domain and the dominant frequency amplitude in the frequency domain, and the pollution intensity data are input into the multi-dimensional correlation analysis model. The Pearson correlation coefficient matrix between each factor is calculated using the multidimensional correlation analysis model, and the time lag order of each factor relative to the qualified product output rate is analyzed using the Granger causality test. The factor with the largest absolute value of the correlation coefficient and a Granger causality test p value of less than 0.05 is identified as the dominant factor. Based on the aforementioned dominant factors, determine the reasons for the decrease in the yield of qualified products.

6. The method according to claim 1, characterized in that, The method further includes: determining the quality state of the chain block to be detected is achieved by a convolutional neural network model, wherein the convolutional neural network model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and an output layer connected in sequence; The convolutional neural network model is obtained through pre-training via the following steps: Obtain a training sample set containing images of qualified and unqualified linked blocks, with each training sample image labeled with a corresponding quality status label; The training sample set is input into the convolutional neural network model for forward propagation calculation to obtain the predicted quality status. Based on the difference between the predicted quality state and the quality state label, the network parameters of the convolutional neural network model are updated using the backpropagation algorithm. The forward and backward propagation steps described above are executed iteratively until the loss function value between the predicted quality state and the quality state label is less than a preset loss threshold.

7. The method according to claim 1, characterized in that, The step of generating the reliability score for the quality status includes: Obtain the sharpness value, contrast value, and motion blur value from the visual quality information; Obtain stability metrics for the feature extraction process, including feature point repetition rate, feature descriptor matching success rate, and standard deviation of feature extraction time; The sharpness value, the contrast value, the motion blur degree value, the feature point repetition rate, the feature descriptor matching success rate, and the standard deviation of the feature extraction time are input into the pre-trained reliability regression model. The reliability score of the quality status is output by the reliability regression model, and the reliability score ranges from 0 to 1.

8. The method according to claim 1, characterized in that, The step of determining the reason for the decrease in the yield of qualified products also includes: Obtain the identification information of the current production shift and the identification information of the current operator; Based on the identification information of the production shift and the identification information of the operator, the benchmark qualified product output rate, benchmark reliability score and benchmark clarity information corresponding to the current production shift and the current operator are extracted from the pre-stored production line operation parameter benchmark library. The benchmark qualified product output rate, the benchmark reliability score, and the benchmark clarity information are used as reference benchmarks to determine the reasons for the decline in the qualified product output rate. When the instantaneous value of the qualified product output rate is lower than the benchmark qualified product output rate, the instantaneous value of the reliability score is lower than the benchmark reliability score, and the instantaneous value of the sharpness information is lower than the benchmark sharpness information, it is determined that the reason for the decrease in the qualified product output rate is the decrease in the reliability of the image recognition system.

9. The method according to claim 1, characterized in that, The steps for adjusting the operating parameters of the production line include: When it is determined that the decrease in the yield of qualified products is due to a decrease in the reliability of the image recognition system, the priority of parameter adjustment is determined based on the reliability score and the clarity information. If the sharpness information is lower than the first sharpness threshold, the exposure time and gain parameters of the industrial camera are adjusted preferentially through the production line parameter control interface. If the lens contamination intensity is higher than the first contamination threshold, the automatic lens cleaning device will be activated first through the production line parameter control interface; If the light intensity is lower than the first light threshold or higher than the second light threshold, the lighting parameters are adjusted first through the production line parameter control interface; If the mechanical vibration intensity is higher than the first vibration threshold, the mechanical movement speed of the production line will be reduced preferentially through the production line parameter control interface.

10. A system for recognizing the demolding and flipping state of interlocking blocks, characterized in that, include: An industrial camera is set at a predetermined position in the interlocking block demolding and flipping station to acquire images of the interlocking block to be inspected; An image preprocessing module, whose input is connected to the output of the industrial camera, is used to perform visual quality optimization processing on the image; The visual quality information acquisition module has its input end connected to the output end of the image preprocessing module, and is used to acquire the visual quality information of the optimized image. An adaptive feature extraction strategy adjustment module has its first input end connected to the output end of the visual quality information acquisition module, and is used to adjust the feature extraction strategy according to the visual quality information. Its output end outputs the visual features of the chain block to be detected. An image recognition module, whose input is connected to the output of the adaptive feature extraction strategy adjustment module, is used to determine the quality status of the interlocking block to be detected based on the visual features. A reliability score generation module has a first input terminal connected to the output terminal of the visual quality information acquisition module and a second input terminal connected to the output terminal of the adaptive feature extraction strategy adjustment module. It is used to generate a reliability score based on the visual quality information and the stability index of the feature extraction process. The production line data acquisition interface is used to collect data on the production line's qualified product yield, mechanical vibration intensity, lens contamination intensity, and light intensity. The root cause analysis module has a first input terminal connected to the output terminal of the reliability score generation module and a second input terminal connected to the output terminal of the production line data acquisition interface. It is used to determine the reasons for the decline in the qualified product output rate based on the qualified product yield rate, the reliability score, the clarity information, and the production line operating status information. The production line parameter control interface has its input end connected to the output end of the root cause analysis module, and is used to adjust the operating parameters of the production line according to the cause. The output terminals of the production line parameter control interface are respectively connected to the control terminal of the industrial camera, the mechanical motion control terminal of the production line, and the lighting control terminal.

Images