Liquid crystal screen image defect identification method and system

Abstract

The invention discloses a liquid crystal screen image defect identification method and a liquid crystal screen image defect identification system, and the liquid crystal screen image defect identification method comprises the following steps: a reference image acquisition step, an environment analysis step, an abnormity identification step, an abnormity discrimination step, and an identification classification step. By introducing the reference image acquisition and environment analysis steps, the illumination uniformity degree and the equipment stability degree of the image acquisition environment can be acquired and quantified in real time. The innovative design overcomes the limitation of misjudgment and missing detection caused by environmental factors in the prior art, and the accuracy and the stability of defect identification are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of image processing and machine vision technology, and more specifically, to a method and system for identifying defects in LCD screen images. Background Technology

[0002] In modern industrial production, especially in the manufacturing process of LCD panels, strict quality control is crucial. Traditional defect detection methods, whether relying on manual visual inspection or simple image processing techniques, face challenges such as low efficiency, susceptibility to errors, and inability to provide timely feedback on production status. With the rise of smart manufacturing and the Industrial Internet, the demand for real-time analysis of massive amounts of production data and accurate identification of minute defects is increasing. However, existing image defect recognition systems often struggle to operate stably and reliably in complex production environments due to external interference, which not only affects product yield but also hinders the improvement of production efficiency.

[0003] Specifically, on automated production lines for LCD panels, machine vision-based defect detection systems can reliably and efficiently replace manual visual inspection under ideal conditions, ensuring consistency and speed in inspection. However, during long-term continuous production, some initially unconsidered environmental factors begin to interfere with the accuracy of the detection system. For example, industrial light sources used to provide uniform illumination experience slow light decay in their internal light-emitting units over time, causing the light illuminating the panel to become less uniform. This subtle unevenness in illumination creates slightly darker areas or gradual grayscale gradients in images captured by the camera, characteristics very similar to some real-world display unevenness defects. This leads the system to frequently misjudge these artifacts caused by illumination problems as defects in the panel itself, resulting in numerous false alarms.

[0004] Furthermore, other large equipment in the production workshop transmits weak mechanical vibrations through the ground and equipment frames during start-up, shutdown, or high-speed operation, which are sufficient to cause high-frequency slight shaking of the camera mounting brackets. Therefore, in the actual environment of an LCD screen production line, facing two dynamic and irregular environmental disturbances—uneven light intensity caused by long-term use of lighting sources and mechanical vibrations caused by equipment operation—existing image defect recognition systems face the challenge of effectively distinguishing whether anomalies in images originate from genuine defects in the panel itself or artifacts caused by uneven lighting or momentary vibrations. This necessitates ensuring the accuracy and stability of the recognition results while maintaining high-precision manufacturing quality control requirements. This is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] This invention discloses a method and system for identifying defects in LCD screen images, aiming to solve the problem that existing LCD screen image defect identification systems suffer from misjudgments and missed detections due to environmental interference such as uneven lighting and mechanical vibration in complex production environments, thereby affecting product yield and production efficiency.

[0006] The technical solution of the present invention is as follows: This invention discloses a method for identifying defects in liquid crystal display (LCD) images, comprising the following steps: The reference image acquisition steps involve acquiring reference images to evaluate the image acquisition environment status. These reference images include a white reference plate image for evaluating the uniformity of illumination and a calibration pattern image for evaluating the stability of the equipment. The environmental analysis steps involve determining a first environmental state parameter characterizing the uniformity of illumination during image acquisition based on the white reference plate image, and determining a second environmental state parameter characterizing the stability of the equipment during image acquisition based on the calibration pattern image. The anomaly identification step involves acquiring an image of the LCD screen to be inspected and analyzing the image to identify potential abnormal areas. The anomaly detection step identifies potential anomaly areas based on the first and second environmental state parameters to distinguish between artifacts caused by environmental interference and real defects in the LCD screen. The identification and classification process identifies and classifies real defects, outputting the identification results of real defects, the discrimination results of environmental interference artifacts, and the corresponding environmental state information.

[0007] Furthermore, in the above-mentioned LCD screen image defect recognition method, the anomaly detection step includes: Obtain a first environmental state parameter characterizing the uniformity of illumination during image acquisition, and a second environmental state parameter characterizing the stability of the device during image acquisition; Image features of potential abnormal regions are extracted, including brightness distribution, grayscale gradation, edge sharpness, and morphological structure. The extracted image features are compared with the preset features of uneven illumination artifacts and mechanical vibration artifacts; Based on the comparison results and the first and second environmental state parameters, the type of potential abnormal areas is determined according to the classification and identification strategy.

[0008] Based on this, the classification and identification strategy is as follows: when the brightness distribution and grayscale gradation characteristics of the potential abnormal area match the characteristics of uneven illumination artifacts, and the first environmental state parameter indicates that uneven illumination exists, the potential abnormal area is determined to be an illumination artifact; when the edge sharpness of the potential abnormal area decreases, the morphological structure becomes blurred or stretched, and the second environmental state parameter indicates that mechanical vibration exists, the potential abnormal area is determined to be a vibration artifact; when the potential abnormal area simultaneously meets the characteristics of uneven illumination artifacts and mechanical vibration artifacts, and both the first and second environmental state parameters indicate the existence of corresponding interference, the potential abnormal area is determined to be a composite environmental interference artifact; when the characteristics of the potential abnormal area do not match the characteristics of uneven illumination artifacts and mechanical vibration artifacts, or although there is a partial match but the corresponding environmental state parameter does not indicate the existence of interference, the potential abnormal area is determined to be a real defect of the LCD screen.

[0009] Through this technical solution, the present invention provides a detailed and comprehensive classification and identification strategy, which can accurately determine the nature of potential abnormal areas based on a combination of multiple features and environmental parameters, effectively avoiding misjudgment and missed detection.

[0010] In one implementation, the environmental analysis step includes: Images of a white reference board were acquired under both standard lighting conditions and slightly varied lighting conditions. Perform pixel-level ratio calculations on two white reference board images to obtain a ratio image; The first environmental state parameters are determined based on the uniformity of brightness distribution in the ratio image.

[0011] Based on the above, the environmental analysis steps include: Within a very short time window for acquiring images from the LCD screen, two calibration pattern images are acquired consecutively. Two consecutive calibration pattern images are processed to identify corresponding feature points on the calibration pattern; Based on the corresponding feature points, calculate the pixel-level displacement of the corresponding feature points between the two calibration pattern images; Based on pixel-level displacement, analyze the distribution and relative differences of pixel-level displacement to construct a relative displacement field; Based on the relative displacement field, a second environmental state parameter characterizing the stability of the device during image acquisition is determined.

[0012] To enhance functionality, a region separation operation is performed on potentially abnormal regions to obtain multiple sub-regions, and an independent anomaly detection step is performed on each sub-region.

[0013] More specifically, in some implementation schemes, image feature analysis is performed on potential anomalous regions to obtain information on the brightness, color, and texture distribution within the potential anomalous regions; Based on brightness, color, and texture distribution information, pixel clusters within potential abnormal regions are identified; Determine the boundaries between pixel clusters based on pixel clustering; Based on the boundaries between pixel clusters, potential anomaly regions are divided into multiple connected regions, with each connected region serving as a sub-region.

[0014] Based on the above, the identification and classification steps include: Multi-level feature extraction is performed on real defects, including the size, shape, brightness, color, and texture information of the defects; Weights are assigned to multi-level features to highlight features with high discriminative power in defect type identification; Based on the characteristics of the assigned weights, the actual defects are compared with the preset defect types; Based on the comparison results, the types of real defects are identified and classified.

[0015] Preferably, the sample data volume for identifying defect types is used to determine whether there are defect types with a small sample data volume. When there are defect types with a small amount of sample data, perform initial weight assignment on multi-level features; Based on the preliminary identification results of defects after the initial weight allocation, the weights of the multi-level features are adjusted. Based on the adjusted weights, evaluate the accuracy and recall of defect type identification; Based on the evaluation results of precision and recall, the weights of multi-level features are adjusted incrementally.

[0016] In one implementation, the step of weighting multi-level features to highlight features with high discriminative power for defect type identification includes: Obtain the process parameters and production time information for the current production batch; Based on process parameters and production time information, retrieve defect sample data that are similar to the current production conditions; Analyze defect sample data to identify highly distinguishable features between different defect types; Based on the identified high-discrimination features, initial weights are assigned to the multi-level features. Monitor the accuracy and false negative rate of defect identification; Based on the monitoring results and the current operating status of the production line, the weights of the multi-level features are adjusted incrementally.

[0017] Secondly, the present invention also discloses a liquid crystal screen image defect recognition system for implementing a liquid crystal screen image defect recognition method.

[0018] This invention discloses a liquid crystal display (LCD) image defect identification method. By introducing reference image acquisition and environmental analysis steps, it can acquire and quantify the lighting uniformity and equipment stability of the image acquisition environment in real time. Specifically, it assesses lighting uniformity by acquiring a white reference board image and assesses equipment stability by acquiring a calibration pattern image, thereby determining a first environmental state parameter and a second environmental state parameter. In the anomaly discrimination step, the method can intelligently discriminate potential anomaly areas based on these environmental state parameters, effectively distinguishing artifacts caused by environmental interference such as uneven lighting or mechanical vibration from genuine defects in the LCD screen itself. This innovative design overcomes the limitations of existing technologies where misjudgments and missed detections are caused by environmental factors, significantly improving the accuracy and stability of defect identification. Finally, by identifying and classifying genuine defects and outputting the discrimination results of environmental interference artifacts and corresponding environmental state information, this invention can not only accurately control product quality but also provide valuable environmental interference data for improving production processes, thereby effectively improving production efficiency and yield. Detailed Implementation

[0019] The technical solutions of this invention will be clearly and completely described below with reference to specific embodiments. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] This invention proposes a method for identifying defects in liquid crystal display (LCD) images, comprising: The reference image acquisition steps involve acquiring reference images for evaluating the image acquisition environment status. These reference images include a white reference plate image for evaluating the uniformity of illumination and a calibration pattern image for evaluating the stability of the equipment. The environmental analysis step involves determining a first environmental state parameter characterizing the uniformity of illumination during image acquisition based on the white reference plate image, and determining a second environmental state parameter characterizing the stability of the equipment during image acquisition based on the calibration pattern image. The anomaly identification step involves acquiring an image of the LCD screen to be inspected and analyzing the LCD screen image to identify potential abnormal areas. The anomaly detection step involves identifying the potential anomaly area based on the first environmental state parameter and the second environmental state parameter, in order to distinguish between artifacts caused by environmental interference and real defects in the LCD screen. The identification and classification steps identify and classify the real defects, and output the identification results of the real defects, the discrimination results of the environmental interference artifacts, and the corresponding environmental state information.

[0021] This invention introduces environmental state parameters to identify potential abnormal areas, effectively distinguishing between artifacts caused by environmental interference and real defects in the LCD screen, thereby improving the accuracy and stability of defect identification.

[0022] To facilitate a clearer understanding of the technical solution of this invention, some key terms are explained first. In this invention, a "reference image" refers to a specific image used to evaluate the environmental conditions in an image acquisition environment, including a "white reference board image" and a "calibration pattern image." The "white reference board image" is used to evaluate the uniformity of illumination, while the "calibration pattern image" is used to evaluate the stability of the device. A "first environmental state parameter" is a quantitative indicator characterizing the uniformity of illumination during image acquisition, determined based on the white reference board image. A "second environmental state parameter" is a quantitative indicator characterizing the stability of the device during image acquisition, determined based on the calibration pattern image. A "potentially abnormal area" refers to an area in the image of the liquid crystal screen to be tested that may contain defects or artifacts, identified through preliminary analysis. An "artifact" specifically refers to the visual representation of a non-real defect in the image caused by environmental interference (such as uneven lighting or mechanical vibration). A "real defect" refers to a physical or display defect inherent in the liquid crystal screen itself that needs to be identified and classified.

[0023] In the reference image acquisition step, a reference image needs to be acquired to evaluate the state of the image acquisition environment. This reference image may include a white reference board image for evaluating the uniformity of illumination and a calibration pattern image for evaluating the stability of the equipment. For example, a white reference board image can be obtained by placing a standard white reference board at a fixed position on the production line and capturing it using an image acquisition device. This white reference board image can be used for subsequent analysis of illumination uniformity. Simultaneously, a calibration pattern with a specific geometry or markings can be placed within the field of view of the image acquisition device and captured to obtain a calibration pattern image. This calibration pattern image can be used to evaluate the stability of the image acquisition device, such as the presence of jitter or displacement.

[0024] In the environmental analysis step, a first environmental state parameter characterizing the uniformity of illumination during image acquisition can be determined based on the white reference board image. For example, the uniformity of illumination can be quantified by analyzing the brightness distribution of the white reference board image and calculating indices such as average brightness, standard deviation, or brightness gradient. The more uniform the brightness distribution, the higher the uniformity of illumination indicated by the first environmental state parameter. Simultaneously, a second environmental state parameter characterizing the stability of the equipment during image acquisition can be determined based on the calibration pattern image. For example, the stability of the equipment can be quantified by performing image processing on the calibration pattern image, identifying feature points on the calibration pattern, and analyzing the positional changes or blurring of these feature points in images acquired at different time points. The smaller the displacement of the feature points or the lower the blurring, the higher the stability of the equipment indicated by the second environmental state parameter.

[0025] In the anomaly identification step, images of the LCD screen to be inspected need to be acquired and analyzed to identify potential anomaly areas. For example, images of the LCD screens on the production line can be obtained using a high-resolution camera. Subsequently, image processing algorithms, such as edge detection, thresholding, and background subtraction, can be used to perform preliminary analysis of the LCD screen images. These algorithms can identify areas in the image that differ significantly from the normal background, such as areas with abnormal brightness, color, or texture, and mark these areas as potential anomaly regions.

[0026] In the anomaly detection step, potential anomaly areas are identified based on a first environmental state parameter and a second environmental state parameter to distinguish between artifacts caused by environmental interference and genuine defects in the LCD screen. For example, if the environmental analysis step determines that the first environmental state parameter indicates uneven lighting, and the characteristics of the potential anomaly area (such as brightness distribution and grayscale gradation) are similar to those of artifacts caused by uneven lighting, then the potential anomaly area may be identified as an illumination artifact. Similarly, if the second environmental state parameter indicates that the device is vibrating, and the characteristics of the potential anomaly area (such as blurred edges and stretched shapes) are similar to those of artifacts caused by mechanical vibration, then the potential anomaly area may be identified as a vibration artifact. If the characteristics of the potential anomaly area do not match those of environmental interference artifacts, or although they partially match but the corresponding environmental state parameter does not indicate the presence of interference, then the area is more likely to be identified as a genuine defect in the LCD screen.

[0027] In the identification and classification step, genuine defects are identified and classified, and the system outputs the identification results of genuine defects, as well as the discrimination results of environmental interference artifacts and corresponding environmental state information. For example, for areas identified as genuine defects, machine learning or deep learning models can be used to conduct detailed analysis, identify the specific type of defect (such as bright spots, dark spots, scratches, foreign objects, etc.), and classify them. Simultaneously, for areas identified as environmental interference artifacts, the system outputs the discrimination results along with corresponding environmental state information, such as the degree of uneven lighting and the intensity of equipment vibration. This information can be used to guide environmental adjustments or equipment maintenance on the production line.

[0028] This invention proposes a method for identifying defects in LCD screen images. By introducing environmental state parameters to discriminate potential abnormal areas, it effectively distinguishes between artifacts caused by environmental interference and genuine defects in the LCD screen. Traditional methods often struggle to operate reliably and stably when faced with dynamic and unpredictable environmental interferences, such as uneven light intensity caused by long-term use of lighting sources and mechanical vibrations caused by equipment operation, leading to false alarms or missed detections. This invention acquires images of a white reference board and a calibration pattern in a reference image acquisition step, and determines a first environmental state parameter characterizing the uniformity of lighting and a second environmental state parameter characterizing the stability of the equipment in an environmental analysis step. Subsequently, in the anomaly discrimination step, these environmental state parameters are combined to discriminate the identified potential abnormal areas, thereby accurately distinguishing between artifacts caused by environmental interference and genuine defects in the LCD screen. This method not only improves the accuracy and stability of defect identification and reduces false alarms and missed detections, but also provides valuable guidance for production line environmental adjustments and equipment maintenance by outputting the discrimination results of environmental interference artifacts and corresponding environmental state information, thereby improving product yield and production efficiency.

[0029] In some embodiments of the present invention described above, the anomaly detection step aims to identify potential anomaly regions based on a first environmental state parameter and a second environmental state parameter, in order to distinguish between artifacts caused by environmental interference and genuine defects in the liquid crystal screen. Specifically, the anomaly detection step includes: Obtain a first environmental state parameter characterizing the uniformity of illumination during image acquisition, and a second environmental state parameter characterizing the stability of the device during image acquisition; Image features of the potential abnormal regions are extracted, including brightness distribution, grayscale gradient, edge sharpness, and morphological structure. The extracted image features are compared with preset features of uneven illumination artifacts and mechanical vibration artifacts; Based on the comparison results and the first and second environmental state parameters, the type of the potential abnormal region is determined according to the classification and identification strategy.

[0030] The acquisition of the first and second environmental state parameters aims to provide crucial environmental background information for subsequent anomaly detection. The first environmental state parameter characterizes the uniformity of illumination during image acquisition, while the second environmental state parameter characterizes the stability of the device during image acquisition. Obtaining these parameters lays the foundation for distinguishing between genuine defects and environmental interference artifacts.

[0031] Furthermore, extracting image features from potential anomalous regions is a crucial step in identifying artifacts or defects. These image features specifically include brightness distribution, grayscale gradation, edge sharpness, and morphological structure. Brightness distribution reflects changes in illumination intensity within a region; grayscale gradation reveals smooth transitions or abrupt changes in pixel values; edge sharpness assesses the sharpness of region boundaries, with blurred edges potentially indicating vibrations; and morphological structure describes the region's geometry and texture. The comprehensive extraction of these features allows for a complete characterization of the visual representation of potential anomalous regions.

[0032] Subsequently, the extracted image features were compared with preset artifact features for uneven illumination and mechanical vibration. These preset artifact features, based on extensive experimental data and accumulated experience, represent the characteristic patterns of image artifacts caused by typical uneven illumination and mechanical vibration, respectively. For example, uneven illumination artifacts may manifest as large areas of uneven or gradual brightness changes, while mechanical vibration artifacts may cause image blurring, edge stretching, or ghosting. Through comparison, it can be preliminarily determined whether potential abnormal regions possess visual features similar to known environmental artifacts.

[0033] Finally, based on the comparison results and the first and second environmental state parameters, the type of the potential abnormal region is determined according to a classification and identification strategy. This means that the discrimination process does not rely solely on the comparison of image features, but combines image features with actual environmental state parameters. For example, even if the image features are similar to uneven lighting artifacts, if the first environmental state parameter indicates uniform lighting, the region is more likely to be a real defect. Conversely, if both the image features and environmental parameters point to some kind of environmental interference, it is determined to be an artifact. This comprehensive discrimination mechanism can effectively improve the accuracy and robustness of the discrimination.

[0034] This invention addresses the inaccuracy of traditional methods in distinguishing between genuine defects and environmental artifacts by combining image feature analysis with environmental state parameters. Specifically, by extracting image features such as brightness distribution, grayscale gradation, edge sharpness, and morphological structure of potential abnormal areas, the characteristics of artifacts or defects can be captured visually. However, image features alone are sometimes insufficient for complete differentiation, as some environmental artifacts may visually resemble genuine defects. The introduction of a first and a second environmental state parameter allows the system to obtain objective environmental information during image acquisition. When image features match preset artifact features, the system further verifies the corresponding environmental state parameters. For example, if the image displays blurry or stretched features, and the second environmental state parameter indicates mechanical vibration, the likelihood of that area being identified as a vibration artifact increases significantly. This dual verification mechanism—the mutual confirmation of image features and environmental parameters—enables the system to more accurately identify artifacts caused by environmental interference, thereby avoiding misjudging them as genuine defects in the LCD screen.

[0035] Through the above technical solution, this invention can significantly improve the accuracy of LCD screen image defect identification. By comprehensively analyzing the image features of potential abnormal areas and the environmental parameters during image acquisition, artifacts caused by environmental factors such as uneven lighting or mechanical vibration can be effectively distinguished from genuine defects in the LCD screen itself. This not only reduces the false alarm rate, avoiding the incorrect rejection of defect-free products, but also lowers the false negative rate, ensuring that genuine defects can be accurately identified and classified. Therefore, it improves the robustness and reliability of the defect detection system, providing a more accurate basis for LCD screen quality control.

[0036] In some embodiments of the present invention described above, a method is proposed for identifying potential abnormal regions to distinguish between artifacts caused by environmental interference and genuine defects in the liquid crystal display screen. This method involves determining the type of potential abnormal regions based on a classification and identification strategy. Specifically, this classification and identification strategy can be further refined.

[0037] The above classification and identification strategy is as follows: when the brightness distribution and grayscale gradient characteristics of the potential abnormal area match the characteristics of uneven illumination artifacts, and the first environmental state parameter indicates that there is uneven illumination, the potential abnormal area is determined to be an illumination artifact; when the edge sharpness of the potential abnormal area decreases, the morphological structure becomes blurred or stretched, and the second environmental state parameter indicates that there is mechanical vibration, the potential abnormal area is determined to be a vibration artifact; when the potential abnormal area simultaneously meets the characteristics of uneven illumination artifacts and mechanical vibration artifacts, and both the first and second environmental state parameters indicate the presence of corresponding interference, the potential abnormal area is determined to be a composite environmental interference artifact; when the characteristics of the potential abnormal area do not match the characteristics of uneven illumination artifacts and mechanical vibration artifacts, or although there is a partial match but the corresponding environmental state parameter does not indicate the presence of interference, the potential abnormal area is determined to be a real defect of the LCD screen.

[0038] Specifically, the uneven illumination artifacts typically manifest as uneven gradients in brightness or color in localized areas of the image. For example, a smooth transition from bright to dark or from dark to bright against a uniform background, or a localized color shift. The mechanical vibration artifacts typically manifest as blurred object edges, stretched lines, or ghosting in the image. For example, the straight edges of a calibration pattern become unclear, or point features are stretched into short lines. The first environmental state parameter quantifies the uniformity of illumination during image acquisition, and can be represented, for example, by the standard deviation of brightness or local brightness differences in the white reference image. The second environmental state parameter quantifies the stability of the equipment during image acquisition, and can be represented, for example, by the displacement of feature points or the fluctuation of the displacement field in the calibration pattern image.

[0039] The present invention achieves accurate differentiation between artifacts and real defects by comparing the image features of potential abnormal areas with preset artifact features and combining them with real-time acquired environmental state parameters. Specifically, when an image feature matches a certain artifact feature, the system further checks whether the corresponding environmental state parameters indicate the presence of corresponding environmental interference. For example, if the brightness distribution and grayscale gradient features are detected to match the characteristics of uneven illumination artifacts, and the first environmental state parameter also clearly indicates uneven illumination, the anomaly can be determined with high confidence to be an illumination artifact. This dual verification mechanism, combining "feature matching" and "environmental verification," effectively avoids misjudging environmental interference as a real defect in the LCD screen, thereby improving the accuracy and reliability of defect identification. For composite environmental interference artifacts, this strategy can identify multiple environmental interferences existing simultaneously and their corresponding image feature manifestations. When the image features do not match any artifact features, or although they partially match but the environmental parameters do not indicate the presence of interference, the possibility of environmental interference is ruled out, thus determining it to be a real defect in the LCD screen.

[0040] Through the above technical solution, this invention can significantly improve the accuracy of LCD screen image defect recognition, effectively avoiding the misjudgment of artifacts caused by environmental factors such as uneven lighting and mechanical vibration as real defects in the LCD screen. This strategy, by combining image feature analysis and environmental state parameters for dual discrimination, makes the defect recognition results more reliable, reduces false alarm and false negative rates, thereby improving the efficiency of production line inspection and the level of product quality control. Especially under complex and changing environmental conditions, this solution can provide more robust and intelligent defect discrimination capabilities.

[0041] In some embodiments of the present invention described above, the environmental analysis step determines a first environmental state parameter characterizing the uniformity of illumination during image acquisition based on a white reference plate image. However, in practical applications, a single white reference plate image may be insufficient to comprehensively and robustly reflect the true uniformity of illumination, especially in the presence of subtle light fluctuations or environmental disturbances, potentially leading to inaccurate assessments of illumination uniformity. Therefore, the present invention further proposes a specific implementation of the aforementioned environmental analysis step, aiming to more accurately assess the uniformity of illumination during image acquisition.

[0042] The above environmental analysis steps include: acquiring images of the white reference board under standard lighting conditions and slightly varied lighting conditions respectively; performing pixel-level ratio calculations on the two white reference board images to obtain a ratio image; and determining the first environmental state parameter based on the brightness distribution uniformity of the ratio image.

[0043] Specifically, acquiring white reference board images under standard lighting conditions and slightly varied lighting conditions means obtaining a first white reference board image when the image acquisition system is operating normally and the lighting environment is considered ideal or at a baseline state. Subsequently, while keeping other conditions constant, a slight lighting change (e.g., minor brightness fluctuations, local shadows, or color temperature shifts) is introduced or simulated, and a second white reference board image is acquired. The purpose is to more sensitively capture the non-uniformity characteristics of lighting by comparing images under different lighting conditions. Standard lighting conditions can be understood as a pre-set or calibrated lighting environment considered uniform and stable; slightly varied lighting conditions refer to slight, imperceptible lighting deviations that may affect image quality compared to standard lighting conditions.

[0044] Performing pixel-level ratio calculations on the two aforementioned white reference images to obtain a ratio image involves dividing each pixel value in the second white reference image by the corresponding pixel value in the first white reference image, thus generating a new image, the ratio image. Its purpose is to eliminate or reduce the influence of overall brightness differences and highlight the relative changes in illumination uniformity within the image. For example, if the brightness ratio of the two images in a certain area deviates significantly from 1, it indicates a problem with the illumination uniformity in that area.

[0045] Determining the first environmental state parameter based on the uniformity of brightness distribution in the ratio image refers to quantifying the uniformity of brightness distribution in the ratio image by analyzing the statistical characteristics of pixel values, such as calculating their standard deviation, variance, entropy, or performing histogram analysis. The purpose is to transform the degree of lighting uniformity reflected by the ratio image into a quantifiable numerical parameter for subsequent environmental state assessment. For example, the more concentrated the pixel values ​​in the ratio image are around a specific value (such as 1), and the smaller the variance, the better the lighting uniformity.

[0046] The present invention acquires images of a white reference board under standard lighting conditions and under slightly varied lighting conditions, and performs pixel-level ratio calculations on these two images. This effectively eliminates or reduces the interference of overall brightness fluctuations on the evaluation of lighting uniformity. The generation of the ratio image allows the system to focus on detecting relative spatial differences in lighting, rather than absolute brightness changes. Therefore, by analyzing the brightness distribution uniformity of the ratio image, the first environmental state parameter characterizing the lighting uniformity at the time of image acquisition can be determined more accurately and robustly, thus providing more reliable environmental information for subsequent anomaly detection.

[0047] Through the above technical solution, this invention can more accurately assess the uniformity of lighting in the image acquisition environment and effectively distinguish between artifacts caused by uneven ambient lighting and real defects in the LCD screen. This ratio-based image analysis method significantly improves the accuracy and anti-interference capability of environmental state parameters, thereby enhancing the overall reliability of LCD screen image defect identification.

[0048] In some preferred embodiments, the specific implementation is as follows: First, under normal operating conditions of the LCD screen image acquisition system, a white reference board image is acquired using a high-precision camera, serving as a reference image under standard lighting conditions. Then, without changing the camera position or the white reference board position, a slight lighting change is simulated by fine-tuning the power of the light source or placing a thin, semi-transparent filter in front of the light source, and another white reference board image is acquired as an image under slightly changed lighting conditions. Next, each pixel value of the second image is divided by the corresponding pixel value of the first image to generate a ratio image. Finally, this ratio image is statistically analyzed, for example, by calculating the standard deviation of its pixel values. If the standard deviation is below a preset threshold, the lighting uniformity is considered good, and the first environmental state parameter indicates uniform illumination; if the standard deviation is above the preset threshold, the lighting is considered uneven, and the first environmental state parameter indicates uneven illumination.

[0049] In some embodiments of the present invention described above, a second environmental state parameter characterizing the stability of the device during image acquisition is proposed based on a calibration pattern image. However, in its implementation, relying solely on a single calibration pattern image or a simple comparison method may make it difficult to capture the instantaneous, minute mechanical vibrations that occur within the extremely short time of image acquisition. Such instantaneous vibrations may cause slight blurring or deformation of the image, which could then be misjudged as genuine defects in the LCD screen during anomaly identification, increasing the false alarm rate and potentially requiring additional complex algorithms to distinguish these artifacts.

[0050] In response, this invention further proposes that the above-mentioned environmental analysis steps include: Within a very short time window for acquiring images of the LCD screen, two calibration pattern images are acquired consecutively; The two consecutive calibration pattern images are processed to identify the corresponding feature points on the calibration pattern; Based on the corresponding feature points, calculate the pixel-level displacement of the corresponding feature points between the two calibration pattern images; Based on the pixel-level displacement, the distribution and relative differences of the pixel-level displacement are analyzed to construct a relative displacement field; Based on the relative displacement field, a second environmental state parameter characterizing the stability of the device during image acquisition is determined.

[0051] Specifically, acquiring two calibration pattern images consecutively within an extremely short time window of acquiring the LCD screen image means that, in actual testing, to ensure the real-time performance and accuracy of environmental parameters, two calibration pattern images are acquired consecutively at an extremely short interval (e.g., milliseconds) before, after, or simultaneously with the acquisition of the LCD screen image. This is intended to capture minute equipment jitter or vibration that may occur within a very short time, which could affect image quality.

[0052] The process of processing the two consecutive calibration pattern images to identify corresponding feature points on the calibration pattern can be understood as using image processing algorithms, such as corner detection (e.g., Harris corner, Shi-Tomasi corner), edge detection (e.g., Canny operator), or feature descriptors (e.g., SIFT, SURF, ORB), to find unique and traceable feature points in the two consecutively acquired calibration pattern images. These feature points are typically pixel regions on the calibration pattern with obvious geometric structures or texture changes, and their purpose is to provide an accurate reference benchmark for subsequent displacement calculations.

[0053] In practical applications, the pixel-level displacement of the corresponding feature points between the two calibration pattern images is calculated based on these feature points. Specifically, this involves using a feature matching algorithm to associate feature points in the first calibration pattern image with their corresponding feature points in the second calibration pattern image, and calculating the pixel coordinate difference in the X and Y directions for each pair of matched feature points. These differences represent the pixel-level displacement, reflecting the relative motion of the device or calibration pattern on the image plane during the interval between the two image acquisitions.

[0054] Furthermore, based on the pixel-level displacements, the distribution and relative differences of these displacements are analyzed to construct a relative displacement field. This may include statistical analysis of all calculated pixel-level displacements, such as calculating the average displacement, the standard deviation of the displacement, and the directional distribution of the displacement vectors. Through these analyses, a two-dimensional displacement field can be generated, where each point represents the displacement vector of the corresponding region on the calibration pattern. This displacement field can visually demonstrate the overall motion trend and local deformation of the device during image acquisition.

[0055] Therefore, based on the relative displacement field, a second environmental state parameter characterizing the stability of the device during image acquisition is determined. For example, the stability of the device can be quantified based on the average displacement magnitude, maximum displacement value, dispersion of the displacement vector, or the presence of significant displacement in a specific direction. If the displacement field shows a large average displacement or a highly discrete displacement vector, it indicates that the device has significant mechanical vibration or instability, thus setting the second environmental state parameter to indicate an unstable state of the device.

[0056] This invention captures instantaneous, minute mechanical vibrations that are difficult to detect with traditional single-image acquisition or long-interval acquisition by continuously acquiring two calibration pattern images within an extremely short time window and performing precise feature point matching and displacement calculation on these two images. It is precisely this high temporal resolution displacement detection that allows for the accurate quantification of the device's true stable state at the moment of image acquisition. By constructing a relative displacement field, the device's motion patterns within the entire field of view can be comprehensively analyzed, thus avoiding errors that may arise from judging based solely on local information. This method transforms the second environmental state parameter of device stability from a fuzzy qualitative judgment into a precise quantitative indicator, providing a reliable environmental basis for subsequent anomaly detection.

[0057] Through the above technical solution, this invention can more accurately and in real-time assess the stability of image acquisition equipment. Compared to relying solely on a single calibration image or making a rough judgment, this solution, through continuous acquisition and pixel-level displacement analysis, can effectively identify minute mechanical vibrations occurring within a very short time, thereby significantly improving the accuracy of the second environmental state parameter. This precise equipment stability assessment enables a more effective distinction between artifacts caused by equipment vibration and genuine defects in the LCD screen during subsequent anomaly detection steps, reducing the false positive rate and improving the reliability and efficiency of defect identification.

[0058] The following is an illustration using a specific example. Assume a high-speed camera is used to capture images of the LCD screens on a production line. To monitor equipment stability in real time, before each image capture of the LCD screen to be inspected, the system continuously triggers the camera to acquire two standard calibration pattern images within a very short time window of 50 milliseconds. These two calibration pattern images are sent to the image processing unit. The image processing unit first uses an ORB feature detector to extract hundreds of feature points from the two images, and then uses a FLANN matcher to match these feature points, obtaining corresponding feature point pairs. For each matched point pair, its pixel displacement in the X and Y directions is calculated. For example, if a feature point's coordinates are (100, 150) in the first image and (101, 152) in the second image, its displacement vector is (1, 2). The system counts the displacement vectors of all feature points and calculates the average magnitude and standard deviation of these displacement vectors. If the average displacement magnitude exceeds a preset threshold (e.g., 2 pixels), or the standard deviation of the displacement vector is too large (e.g., 1 pixel), it is determined that the equipment is experiencing mechanical vibration, and the second environmental state parameter is set to "vibration present". Conversely, if the displacement is extremely small and uniformly distributed, the equipment is determined to be stable, and the second environmental state parameter is set to "stable". When a potentially abnormal area of ​​blurriness or stretching is subsequently detected in the LCD screen image, the system can accurately identify it as a vibration artifact rather than a real defect in the LCD screen by combining this second environmental state parameter of "vibration present".

[0059] In some embodiments of the present invention described above, a method for identifying potential anomalous regions and performing anomaly discrimination is proposed. However, in practical applications, a potential anomalous region may be large in area or complex in structure, and may contain various anomalies of different natures, such as the coexistence of real defects and environmental interference artifacts, or the presence of multiple types of real defects. If a single overall discrimination is performed on such a complex region, the accuracy of the discrimination result may decrease, and it may be impossible to finely distinguish the specific anomaly types of different parts within the region.

[0060] In response, the present invention further proposes to perform a region separation operation on the above-mentioned potential abnormal regions to obtain multiple sub-regions, and to perform an independent anomaly discrimination step on each sub-region.

[0061] Specifically, region separation refers to dividing a large or complex potential anomaly region into several smaller, more uniform sub-regions. The purpose is to decompose complex anomaly regions into units that are easier to analyze and distinguish. For example, a potential anomaly region may contain a real defect and some artifacts caused by environmental interference, or it may contain two different types of real defects. Through region separation, each sub-region can be ensured to have higher specificity when performing anomaly discrimination. The independent anomaly discrimination step means that each separated sub-region is processed according to the above anomaly discrimination steps, that is, the sub-region is judged based on the first environmental state parameter and the second environmental state parameter to distinguish between artifacts caused by environmental interference and real defects of the LCD screen. This independent discrimination can avoid the characteristics of one sub-region interfering with the discrimination results of another sub-region, thereby improving the overall accuracy and precision of the discrimination.

[0062] The present invention introduces a region separation operation to decompose complex potential anomaly regions into multiple more manageable sub-regions. This decomposition allows subsequent anomaly detection steps to perform more precise analysis of the local features of each sub-region. For example, when a potential anomaly region simultaneously contains artifacts caused by uneven illumination and genuine dead pixels on the LCD screen, a single detection logic may struggle to accurately distinguish between these two different types of anomalies without separation. By separating the region into multiple sub-regions—for example, one sub-region primarily containing dead pixel features and another primarily containing illumination artifact features—independent detection of each sub-region can more accurately identify its true type. This avoids misjudgments or omissions caused by region complexity, improving the precision and reliability of defect identification.

[0063] Through the above technical solution, this invention effectively solves the problem of decreased accuracy that may result from single-discrimination when processing complex or large potential anomaly areas. By refining the potential anomaly area into multiple sub-regions and performing independent discrimination, it is possible to more accurately identify and distinguish anomalies of different natures within the area. For example, it can more finely distinguish between real defects and environmental interference artifacts, or identify multiple different types of real defects existing within the same area. This significantly improves the precision and accuracy of LCD screen image defect recognition, especially when facing complex and varied defect morphologies and environmental interference, providing more reliable discrimination results.

[0064] In some preferred embodiments, it is assumed that a large potential anomalous region is identified during the anomaly identification step. This region may span multiple pixels of the LCD screen, and its brightness, color, and texture distribution are not uniform. For example, the left side of the region may exhibit a clear brightness gradient, while the right side may contain several obvious dark spots. If the region is not separated and anomaly detection is performed directly, the detection result may not accurately reflect the true situation of different parts within the region. To address this, the present invention first performs a region separation operation on the potential anomalous region. Specifically, image segmentation algorithms, such as pixel-based clustering or edge detection methods, can be used to divide the large region into several connected sub-regions. For example, the portion with a brightness gradient is divided into one sub-region, while the portion with dark spots is divided into another one or more sub-regions. Subsequently, an independent anomaly detection step is performed on each sub-region. For example, for the brightness gradient sub-region, combined with a first environmental state parameter (such as illumination uniformity), it may be identified as an illumination artifact; while for the dark spot sub-region, combined with a first environmental state parameter and a second environmental state parameter (such as device stability), it may be identified as a real dead pixel of the LCD screen. In this way, even within a complex region of potential anomalies, accurate identification and differentiation of different types of anomalies can be achieved, thereby avoiding ambiguity in overall judgment and improving the accuracy and reliability of defect identification.

[0065] In some embodiments of the present invention described above, a region separation operation is proposed to perform on potential anomalous regions to obtain multiple sub-regions, and each sub-region is independently identified as anomaly. However, in practical applications, potential anomalous regions may contain defects or artifacts of various different natures. Without a refined region separation mechanism, inaccurate sub-region division may occur, affecting the accuracy and efficiency of subsequent anomaly detection. Therefore, the present invention further proposes a specific implementation method for performing region separation on the aforementioned potential anomalous regions, achieving more accurate sub-region division through in-depth analysis of image features.

[0066] Specifically, image feature analysis is performed on the potential anomaly region to obtain brightness, color, and texture distribution information within the potential anomaly region; pixel clusters within the potential anomaly region are identified based on the brightness, color, and texture distribution information; the boundaries between the pixel clusters are determined based on the pixel clusters; and the potential anomaly region is divided into multiple connected regions based on the boundaries between the pixel clusters, with each connected region serving as a sub-region.

[0067] Image feature analysis of potential anomaly regions involves extracting the brightness values, color channel information (e.g., RGB, HSV), and texture features (e.g., gray-level co-occurrence matrix, local binary pattern) of pixels within the region using various image processing algorithms. These features comprehensively describe the visual attributes of the potential anomaly region. Identifying pixel clusters within the potential anomaly region can be understood as grouping pixels with similar characteristics into different pixel groups based on the obtained brightness, color, and texture distribution information using clustering algorithms (e.g., K-means, mean-shift clustering, or density-based clustering methods). The purpose is to distinguish the components with different properties within the potential anomaly region. In practical applications, determining the boundaries between pixel clusters specifically refers to accurately delineating the dividing lines between different pixel clusters after identification using edge detection algorithms (e.g., Canny operator, Sobel operator) or morphological operations (e.g., dilation, erosion). The purpose is to provide a clear basis for subsequent region division. Thus, based on the boundaries between pixel clusters, the potential anomaly region is divided into multiple connected regions, each of which serves as a sub-region. This means that once the boundaries are determined, the potentially anomalous region is logically divided into several independent regions with consistent internal characteristics. Each connected region is treated as an independent sub-region for subsequent separate anomaly detection.

[0068] The present invention comprehensively captures the visual characteristics of potential anomalous regions by performing detailed image feature analysis, including acquiring information on brightness, color, and texture distribution. Based on these features, the system can identify pixel clusters with similar attributes within a region, thereby effectively distinguishing anomalous components of different natures. Furthermore, by determining the boundaries between these pixel clusters, the present invention can accurately segment potential anomalous regions into multiple independent connected sub-regions. This refined region separation mechanism allows subsequent independent anomaly discrimination steps for each sub-region to be more focused and accurate, avoiding the confusion and misjudgment that may occur when treating complex anomalies as a whole, thus improving the robustness and accuracy of defect identification.

[0069] Through the above technical solution, this invention can achieve fine segmentation of complex potential anomaly regions, effectively separating defects or artifacts of different natures. This pixel clustering and boundary determination method based on image features ensures the consistency of features within each sub-region, thereby significantly improving the accuracy and reliability of subsequent independent anomaly discrimination steps. Compared with schemes without a clear region separation mechanism, this invention can more effectively handle composite defects or environmental interference artifacts, avoiding misjudgments or missed judgments caused by improper region division, thus improving the overall performance and efficiency of LCD screen image defect recognition.

[0070] In some embodiments of the present invention, steps for identifying and classifying real defects are proposed. However, in practical applications, there are many types of real defects in LCD screens, and there may be overlapping features or subtle differences between different defects. If only a general or single classification method is used, it may be difficult to achieve high-precision and high-robust defect identification and classification, thereby affecting the efficiency and accuracy of defect detection.

[0071] In response, the present invention further proposes that the above-mentioned identification and classification steps include: Multi-level feature extraction is performed on the real defects, and the features include the size, shape, brightness, color and texture information of the defects; Weights are assigned to the multi-level features to highlight features with high discriminative power for defect type identification; Based on the characteristics of the assigned weights, the actual defects are compared with preset defect types; Based on the comparison results, the types of the actual defects are identified and classified.

[0072] Specifically, the multi-level feature extraction refers to the quantitative description of image information of real defects from different dimensions and granularities. For example, size features may include the defect's area, perimeter, aspect ratio, etc.; shape features may include roundness, rectangularity, convexity, etc.; brightness features may include average brightness, maximum brightness, minimum brightness, brightness standard deviation, etc.; color features may include average color value and color variance in the RGB or HSV color space; and texture features may include contrast, energy, homogeneity, etc., extracted from the gray-level co-occurrence matrix (GLCM). The extraction of these features aims to comprehensively capture the visual attributes of defects, providing a rich data foundation for subsequent classification.

[0073] The purpose of weighting the multi-level features is to enhance the influence of features that have stronger discriminative power for specific defect types, while reducing the interference of features with lower discriminative power. For example, for point defects, size and brightness features may have higher discriminative power; while for linear or planar defects, shape and texture features may be more critical. Weighting can be implemented through feature selection or feature weighting mechanisms using machine learning algorithms (such as support vector machines and neural networks), or it can be preset and adjusted through expert experience or statistical analysis.

[0074] In practical applications, comparing the actual defects with preset defect types based on the weighted features involves inputting the defect feature vectors, after feature extraction and weight allocation, into a pre-trained classification model. This classification model can be a rule-based classifier, Support Vector Machine (SVM), Decision Tree, Random Forest, or a deep learning model (such as a Convolutional Neural Network (CNN)). Preset defect types refer to various known defect categories defined during the training phase, such as bright lines, dark lines, bright spots, dark spots, murras, and foreign objects. The comparison process aims to determine which preset defect type best matches the currently detected actual defect.

[0075] Therefore, identifying and classifying the type of the actual defect based on the comparison results means assigning the current defect to the corresponding defect category according to the discrimination result output by the classification model. For example, if the comparison results show that the defect features highly match the features of the "bright spot" type, then the defect is identified as a bright spot. The classification result includes not only the defect type but also a confidence score to assess the reliability of the classification.

[0076] This invention effectively addresses the challenges of traditional single-feature or indiscriminate feature processing in complex defect classification by introducing a multi-level feature extraction and weight allocation mechanism. Specifically, multi-level feature extraction comprehensively characterizes the essential attributes of defects from multiple dimensions, avoiding misjudgments due to insufficient feature information. Furthermore, weight allocation of these features highlights the most crucial features for distinguishing different defect types, thereby enhancing the classification model's discriminative power. For example, when distinguishing between "bright spots" and "mura," brightness distribution and size features may be given higher weights, while when distinguishing between "scratches" and "foreign objects," shape and texture features may be more important. This dynamic or optimized weight allocation mechanism allows the classifier to more accurately capture subtle differences between different defect types, thus improving overall classification accuracy and robustness.

[0077] Through the above technical solution, this invention can significantly improve the accuracy and precision of identifying and classifying real defects in LCD screens. Compared to basic solutions that only perform general identification and classification, this invention extracts multi-level features from defects and assigns weights to the distinguishability of defect types based on these features. This allows the classification model to more effectively utilize key information, thereby more accurately distinguishing different types of real defects when faced with complex and diverse defect types, reducing false alarms and missed alarms. This refined classification result not only helps production lines conduct more precise source analysis and quality control of defects, but also provides a more reliable basis for subsequent repair or scrapping decisions, thus improving production efficiency and product quality.

[0078] In some preferred embodiments, a specific example is given below. Suppose that when inspecting an LCD screen, the anomaly detection step identifies a potential anomaly area. After the anomaly discrimination step confirms that it is a genuine defect, the identification and classification step will be initiated. First, multi-level feature extraction is performed on the genuine defect, for example, calculating its area as 50 square pixels (size feature), aspect ratio as 1.2 (shape feature), average brightness value as 200 (brightness feature), reddish color (color feature), and its edges as having obvious jaggedness (texture feature).

[0079] Subsequently, according to a pre-defined weighting strategy, for example, brightness and size features have higher weights for point defects, while shape and texture features have higher weights for linear defects. Assume the feature vector of the current defect is input into a deep learning-based classification model, which has already learned the feature patterns and corresponding weights of different defect types (such as bright spots, dark spots, scratches, and murras) during training. The model then compares the feature vector of the current defect with the pre-defined defect types based on these weighted features.

[0080] If the comparison results show that the weighted features of the defect highly match the feature patterns of the "bright spot" type—for example, its high brightness and small size features show a very strong match after weighting—then the defect is identified as a "bright spot." Simultaneously, the system will output the defect type as "bright spot," possibly accompanied by a confidence score, such as 95%. This refined identification and classification allows production personnel to quickly understand the nature of the defect and take appropriate action. For example, for a bright spot defect, it may be necessary to inspect the backlight module or pixel driving circuit.

[0081] In some embodiments of the present invention described above, after multi-level feature extraction of real defects, weights are assigned to these multi-level features to highlight features with high discriminative power for defect type identification. However, in practical applications, the amount of sample data for different defect types may vary significantly, especially for some rare defect types where the amount of sample data may be very scarce. If weights are assigned only once based on limited sample data, the weight settings may be inaccurate, thus affecting the accuracy and recall of identifying these defect types with limited sample data. To address this, the present invention further proposes a method for optimizing multi-level feature weight allocation, aiming to improve the identification performance of various defects, especially defects with limited sample data, by dynamically adjusting the weights.

[0082] In response, this invention further proposes the following steps for identifying and classifying genuine defects: Identify the sample data volume of the defect type to determine whether there are defect types with a small sample data volume; When there is a defect type with a small amount of sample data, an initial weight assignment is performed on the multi-level features; Based on the preliminary identification results of the defects after the initial weight allocation, the weights of the multi-level features are adjusted; Based on the adjusted weights, the accuracy and recall of the defect type identification are evaluated; Based on the evaluation results of the accuracy and recall, the weights of the multi-level features are adjusted incrementally.

[0083] Specifically, the sample data volume for identifying the defect type refers to the number of training samples or historical samples corresponding to each known or preset defect type obtained through statistical analysis of all known or preset defect types before defect identification and classification. The purpose is to identify defect types that are insufficiently representative in the training dataset and have a significantly small sample size. For example, a threshold can be set; when the sample size for a certain defect type is lower than this threshold, it is determined to be a defect type with a small sample data volume.

[0084] When there are defect types with limited sample data, initial weight allocation for the multi-level features can be understood as assigning initial weight values ​​to each level of features such as size, shape, brightness, color, and texture information of the defect based on preset rules, expert experience, or preliminary statistical analysis. These initial weights aim to provide a starting point for subsequent iterative optimization, ensuring a preliminary identification basis even for defects with limited sample data.

[0085] Furthermore, adjusting the weights of the multi-level features based on the preliminary identification results of the defects after the initial weight allocation means using these weights to perform a preliminary identification and classification of the defects after the initial weight allocation is completed. Based on the results of this preliminary identification, especially for defect types with poor identification performance (e.g., high false positive or false negative rates), particularly those with small sample sizes, the feature weights are specifically modified. For example, the weights of features with high discriminative power but previously low weights can be increased, or the weights of noise-sensitive features can be decreased.

[0086] The evaluation of the precision and recall of defect type identification based on the adjusted weights refers to re-identifying defects using the adjusted weights after weight adjustment and calculating the precision and recall of the identification results. Precision measures how many of the identified defects are true defects, while recall measures how many of all true defects are successfully identified. These metrics are key quantitative standards for evaluating the effectiveness of the current weight allocation.

[0087] In practical applications, the weights of the multi-level features are incrementally adjusted based on the evaluation results of accuracy and recall. For example, iterative optimization algorithms such as gradient descent, genetic algorithms, or reinforcement learning can be used. The feature weights are adjusted continuously in small steps according to the changing trends of accuracy and recall. The aim is to gradually converge to an optimal or near-optimal weight allocation scheme through multiple iterations, thereby maximizing the overall defect recognition performance, especially in balancing the recognition effects of common and rare defects.

[0088] This invention addresses the problem of imbalanced sample size by introducing the identification of the sample data volume for different defect types. When a defect type with a small sample data volume is identified, the system performs an initial weight allocation, providing a foundation for subsequent optimization. Subsequently, analysis of the preliminary identification results reveals shortcomings in the current weight allocation, particularly for difficult-to-identify defect types. Based on accuracy and recall evaluations, the system quantifies the effectiveness of the current weight allocation and uses these quantitative indicators to guide incremental adjustments to the weights. This iterative and feedback mechanism allows feature weights to dynamically adapt to the identification needs of different defect types, thus overcoming the limitations of static weight allocation in handling imbalanced sample size.

[0089] Through the above technical solution, this invention effectively solves the problem of decreased recognition performance caused by the uneven amount of sample data for different defect types during defect identification and classification. Especially for defect types with limited sample data, by adjusting the sample size, initial weight allocation, weight adjustment based on preliminary recognition results, and evaluation and incremental adjustment of accuracy and recall, the feature weights can be more refined and adaptively matched to the recognition needs of various defects. This significantly improves the accuracy and recall of defect type identification, especially by ensuring effective identification of rare defect types while maintaining good recognition performance for common defects, thereby enhancing the robustness and practicality of the overall LCD screen image defect recognition method.

[0090] In some preferred embodiments, it is assumed that four common defect types exist during the LCD screen manufacturing process: "bright spots," "dark spots," "scratches," and "foreign objects," as well as a very rare "Mura effect" defect. Historically, the sample size for the "Mura effect" is far lower than that of the other four defects.

[0091] First, the system identifies the "Mura effect" as a defect type with limited sample data. Next, it assigns initial weights to the multi-level features of all defects (such as size, shape, brightness, color, and texture). For example, brightness might initially be given a higher weight because bright spots and dark spots are more noticeable in terms of brightness.

[0092] The system then uses these initial weights to perform preliminary identification on a batch of LCD screen images to be detected. The preliminary identification results show that the recall rate for the "Mura effect" is very low, and many genuine Mura effect defects are missed.

[0093] Based on this initial identification result, the system will adjust the feature weights. For example, it may increase the weights of texture and color gradient features, since the Mura effect often manifests as uneven color or brightness gradient regions, and these features may have low weights in the initial assignment.

[0094] The system then uses the adjusted weights to identify the "Mura effect" again and evaluates the precision and recall. If the recall improves but is still not ideal, the system will incrementally adjust the weights based on the evaluation results, such as further fine-tuning the weights of texture and color features, and possibly introducing higher weights for region uniformity features.

[0095] Through this iterative and incremental adjustment, the accuracy and recall of the "Mura effect" were eventually brought to an acceptable level, without significantly affecting the recognition performance of other common defects.

[0096] In some embodiments of the present invention described above, a method is proposed to extract multi-level features from real defects and then assign weights to these multi-level features to highlight features with high discriminative power for defect type identification. However, in actual LCD screen production, the type and manifestation of defects, as well as the process parameters and time information of the production environment, may dynamically change. If the feature weight allocation method is static or based solely on historical general data, it may not be able to adequately adapt to these dynamic changes, affecting the accuracy and robustness of defect identification, especially when facing new defects or fluctuations in production conditions, where the identification effect may be poor.

[0097] To address this, the present invention further proposes the following steps for weighting the aforementioned multi-level features to highlight features with high discriminative power in defect type identification: Obtain the process parameters and production time information for the current production batch; Based on the process parameters and the production time information, retrieve defect sample data similar to the current production conditions; The defect sample data is analyzed to identify highly distinguishable features between different defect types; Based on the identified high-discrimination features, an initial weight allocation is performed on the multi-level features; Monitor the accuracy and false negative rate of defect identification; Based on the monitoring results and the current operating status of the production line, the weights of the multi-level features are adjusted incrementally.

[0098] Specifically, obtaining the process parameters and production time information of the current production batch refers to the automatic or manual input of the production process parameters of the batch of LCD screens currently being tested, such as production line number, equipment model, material batch, temperature, humidity, and the specific production time of that batch. This information forms the basis for subsequent retrieval of similar defect sample data. The process parameters may include, but are not limited to, production line speed, coating thickness, curing temperature, and exposure time, while the production time information can be accurate to the date, shift, or even a specific time period.

[0099] Furthermore, based on the process parameters and production time information, defect sample data similar to the current production conditions is retrieved. This can be understood as utilizing defect sample data stored in a historical database, comparing the current production conditions with historical production conditions, and filtering out defect samples with high similarity in process parameters and production time. For example, a distance-based similarity measurement method can be used to calculate the similarity score between the current production conditions and historical records, and select sample data with scores higher than a preset threshold. The purpose is to ensure that the sample data used for weight allocation is highly relevant to the current actual production situation, thereby improving the effectiveness of weight allocation.

[0100] In practical applications, the defect sample data is analyzed to identify highly discriminative features between different defect types. Specifically, statistical analysis and machine learning algorithms (such as feature selection algorithms, decision tree analysis, and principal component analysis) are used to deeply mine the retrieved similar defect sample data to identify features that show significant differences in distinguishing different defect types. For example, for a "bright spot" defect, its brightness feature may be highly discriminative; for a "scratch" defect, its shape and texture features may be more discriminative. The aim is to determine which multi-level features are most critical for defect classification under current production conditions.

[0101] Based on this, the initial weight allocation for the multi-level features, according to the identified high-discrimination features, refers to assigning initial weight values ​​to different multi-level features (such as size, shape, brightness, color, texture, etc.) based on the above analysis results. High-discrimination features will be given higher weights to enable them to play a greater role in subsequent defect identification. For example, expert experience, analytic hierarchy process (AHP), or information gain-based algorithms can be used to determine the initial weights.

[0102] Simultaneously, monitoring the accuracy and false negative rate of defect identification can be understood as continuously collecting the system's identification results of real defects when running the defect identification system on the actual production line, and comparing them with the results of manual re-inspection or subsequent quality inspection, thereby evaluating the system's accuracy (the proportion of correctly identified defects) and false negative rate (the proportion of real defects that were not identified) in real time. This monitoring can be conducted using methods such as sampling inspection and full inspection verification.

[0103] Finally, based on the monitoring results and the current operating status of the production line, the weights of the multi-level features are incrementally adjusted. This means that when the monitoring accuracy decreases or the false negative rate increases, or when the operating status of the production line changes (e.g., equipment wear, material replacement, changes in ambient temperature and humidity), the system will make small, iterative adjustments to the previously assigned feature weights according to a preset adjustment strategy or adaptive learning algorithm (e.g., reinforcement learning, Bayesian optimization). The purpose is to enable the feature weights to dynamically adapt to changes in the production environment and continuously optimize the performance of defect identification.

[0104] The present invention incorporates process parameters and production time information of production batches, and dynamically retrieves similar defect sample data based on this information, thereby ensuring that the basis for feature weight allocation is highly relevant to the current actual production conditions. By analyzing this relevant sample data, features with high discriminative power for defect types in the current production environment can be identified, and initial weight allocation can be performed accordingly. Furthermore, by continuously monitoring the accuracy and false negative rate of defect identification, and combining this with the real-time operating status of the production line, the system can incrementally adjust the feature weights. This dynamic and adaptive weight adjustment mechanism allows the defect identification model to continuously learn and adapt to changes in the production environment, avoiding the performance degradation that may result from static weight allocation.

[0105] Through the above technical solution, this invention enables dynamic and adaptive management of multi-level feature weights in LCD screen image defect recognition. Compared to traditional static weight allocation methods, this invention's solution can intelligently adjust feature weights based on the specific process parameters and production time information of the current production batch, as well as the real-time operating status of the production line, thereby significantly improving the accuracy and robustness of defect recognition. Especially when facing complex situations such as fluctuating production conditions, the emergence of new defects, or the evolution of defect features over time, this solution can effectively reduce false alarm and false negative rates, ensuring that the defect recognition system always remains in optimal working condition, thereby improving the efficiency and reliability of LCD screen product quality control.

[0106] In some preferred embodiments, a specific example is given below. Suppose that a liquid crystal display (LCD) production line will exhibit slightly different characteristics in its "Mura" defects (a type of uneven brightness defect) depending on the season or the raw materials used from different suppliers. Traditional identification systems might use fixed weights for brightness, color, and texture features. However, the solution of this invention first obtains the process parameters of the current production batch (e.g., raw material supplier A, production temperature 25°C) and production time information (e.g., summer). Based on this information, the system retrieves Mura defect sample data from a historical database generated under the same or similar conditions (e.g., using the same raw materials from supplier A, summer production).

[0107] Next, the system analyzes the retrieved Mura defect sample data and finds that when using raw materials from supplier A and producing in the summer, the "brightness uniformity" feature of the Mura defect has a higher distinguishability than the "texture roughness" feature in differentiating Mura from other defect types. Based on this analysis, the system assigns a higher initial weight to the "brightness uniformity" feature and a relatively lower initial weight to the "texture roughness" feature.

[0108] During actual production, the system continuously monitors the accuracy and false negative rate of defect identification. For example, if the false negative rate of Mura defects increases, and a slight increase in the vibration frequency of a critical piece of equipment on the production line (e.g., a backlight module assembly machine) is detected (a change in production line operating status), the system may determine that the current weight allocation may no longer be optimal. In this case, the system will incrementally adjust the feature weights based on the monitoring results and the production line operating status. For example, it may slightly increase the weight of the vibration-related "edge sharpness" feature, or reassess the weight of the "brightness uniformity" feature, to adapt to the new production conditions, thereby ensuring that the defect identification system can continue to operate efficiently.

[0109] This invention also proposes a liquid crystal display (LCD) image defect recognition system to implement a method for identifying LCD image defects. The system modularizes functions such as image acquisition, environmental analysis, anomaly recognition, anomaly discrimination, and classification, enabling collaborative work among these modules. The reference image acquisition module provides environmental data, the environmental analysis module converts this data into quantifiable parameters, the anomaly recognition module performs initial screening of anomalies, the anomaly discrimination module intelligently distinguishes anomalies using environmental parameters, and finally, the classification module provides precise defect type and environmental interference information. This clearly defined, collaborative modular design allows the system to efficiently and accurately handle complex defect recognition tasks.

[0110] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying defects in LCD screen images, characterized in that, include: The reference image acquisition steps involve acquiring reference images for evaluating the image acquisition environment status. These reference images include a white reference plate image for evaluating the uniformity of illumination and a calibration pattern image for evaluating the stability of the equipment. The environmental analysis step involves determining a first environmental state parameter characterizing the uniformity of illumination during image acquisition based on the white reference plate image, and determining a second environmental state parameter characterizing the stability of the equipment during image acquisition based on the calibration pattern image. The anomaly identification step involves acquiring an image of the LCD screen to be inspected and analyzing the LCD screen image to identify potential abnormal areas. The anomaly detection step involves identifying the potential anomaly area based on the first environmental state parameter and the second environmental state parameter, in order to distinguish between artifacts caused by environmental interference and real defects in the LCD screen. The identification and classification steps identify and classify the real defects, and output the identification results of the real defects, the discrimination results of the environmental interference artifacts, and the corresponding environmental state information.

2. The method for identifying defects in LCD screen images according to claim 1, characterized in that, The anomaly detection step includes: Obtain a first environmental state parameter characterizing the uniformity of illumination during image acquisition, and a second environmental state parameter characterizing the stability of the device during image acquisition; Image features of the potential abnormal regions are extracted, including brightness distribution, grayscale gradient, edge sharpness, and morphological structure. The extracted image features are compared with preset features of uneven illumination artifacts and mechanical vibration artifacts; Based on the comparison results and the first and second environmental state parameters, the type of the potential abnormal region is determined according to the classification and identification strategy.

3. The method for identifying defects in LCD screen images according to claim 2, characterized in that, The classification and identification strategy is as follows: when the brightness distribution and grayscale gradient characteristics of the potential abnormal region match the characteristics of the uneven illumination artifact, and the first environmental state parameter indicates uneven illumination, the potential abnormal region is determined to be an illumination artifact; when the edge sharpness of the potential abnormal region decreases, its shape and structure become blurred or stretched, and the second environmental state parameter indicates mechanical vibration, the potential abnormal region is determined to be a vibration artifact; when the potential abnormal region simultaneously meets the characteristics of both the uneven illumination artifact and the mechanical vibration artifact, and both the first and second environmental state parameters indicate the presence of corresponding interference, the potential abnormal region is determined to be a composite environmental interference artifact; when the characteristics of the potential abnormal region do not match either the characteristics of the uneven illumination artifact or the mechanical vibration artifact, or although there is a partial match but the corresponding environmental state parameter does not indicate the presence of interference, the potential abnormal region is determined to be a real defect of the LCD screen.

4. The method for identifying defects in LCD screen images according to claim 1, characterized in that, The environmental analysis steps include: Images of the white reference board were acquired under both standard lighting conditions and slightly varied lighting conditions. Pixel-level ratio calculations are performed on the two white reference board images to obtain a ratio image; The first environmental state parameter is determined based on the uniformity of brightness distribution in the ratio image.

5. The method for identifying defects in LCD screen images according to claim 5, characterized in that, The environmental analysis steps include: Within a very short time window for acquiring images of the LCD screen, two calibration pattern images are acquired consecutively; The two consecutive calibration pattern images are processed to identify the corresponding feature points on the calibration pattern; Based on the corresponding feature points, calculate the pixel-level displacement of the corresponding feature points between the two calibration pattern images; Based on the pixel-level displacement, the distribution and relative differences of the pixel-level displacement are analyzed to construct a relative displacement field; Based on the relative displacement field, a second environmental state parameter characterizing the stability of the device during image acquisition is determined.

6. The method for identifying defects in LCD screen images according to claim 5, characterized in that, A region separation operation is performed on the potential abnormal region to obtain multiple sub-regions, and an independent anomaly detection step is performed on each sub-region.

7. The method for identifying defects in LCD screen images according to claim 6, characterized in that, Image feature analysis is performed on the potential anomaly region to obtain brightness, color, and texture distribution information within the potential anomaly region; Based on the brightness, color, and texture distribution information, pixel clusters within the potential abnormal regions are identified; Based on the pixel clustering, the boundaries between the pixel clusters are determined; Based on the boundaries between the pixel clusters, the potential abnormal region is divided into multiple connected regions, with each connected region serving as a sub-region.

8. The method for identifying defects in LCD screen images according to claim 1, characterized in that, The identification and classification steps include: Multi-level feature extraction is performed on the real defects, and the features include the size, shape, brightness, color and texture information of the defects; Weights are assigned to the multi-level features to highlight features with high discriminative power for defect type identification; Based on the characteristics of the assigned weights, the actual defects are compared with preset defect types; Based on the comparison results, the types of the actual defects are identified and classified.

9. A method for identifying defects in LCD screen images according to claim 8, characterized in that, Identify the sample data volume of the defect type to determine whether there are defect types with a small sample data volume; When there is a defect type with a small amount of sample data, an initial weight assignment is performed on the multi-level features; Based on the preliminary identification results of the defects after the initial weight allocation, the weights of the multi-level features are adjusted; Based on the adjusted weights, the accuracy and recall of the defect type identification are evaluated; Based on the evaluation results of the accuracy and recall, the weights of the multi-level features are adjusted incrementally.

10. A liquid crystal display screen image defect recognition system for implementing the liquid crystal display screen image defect recognition method as described in any one of claims 1-9.