A deep learning-based liquid crystal display screen defect automatic classification method

Abstract

This invention belongs to the field of artificial intelligence, specifically relating to an automatic defect classification method for liquid crystal displays based on deep learning. It aims to solve the problems of low recognition rate, high false alarm rate, and poor production line adaptability of traditional methods for subtle defects. The method includes: acquiring high-resolution corrected images; constructing a multi-scale feature extraction network with an encoder-decoder structure; embedding a spatial-channel joint attention module to enhance the response of key defect areas; employing a weighted optimization of focus loss and boundary awareness loss to alleviate class imbalance; and finally outputting the probabilities of eight defect classes and calibrating the confidence level through a fully connected classification head. In the post-processing stage, morphological operations and connected component analysis are combined to generate a structured report, while physical synthesis enhancement and transfer learning are introduced to improve small-sample generalization ability. This solution achieves a classification accuracy of ≥98.5%, a false negative rate of <0.3%, a single-screen detection cycle of <1 second, and a daily production capacity of over 10,000 units, significantly improving the automation efficiency and robustness of defect recognition on production lines.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an automatic classification method for defects in liquid crystal displays based on deep learning. Background Technology

[0002] With the rapid development of intelligent manufacturing and industrial vision inspection technologies, LCD screens, as core display components of modern electronic devices, directly impact the market competitiveness of end products through their production yield and defect control accuracy. Currently, production lines generally rely on manual visual inspection or semi-automatic inspection systems based on traditional image processing algorithms. These systems struggle to meet the demands for accurate identification of high-resolution, minute defects and complex textures, resulting in high false negative rates and poor classification consistency, severely hindering the efficiency and quality stability of large-scale automated production. Especially in high-generation, high-speed production lines, the diverse forms and intertwined causes of defects pose unprecedented challenges to the semantic understanding, feature generalization, and real-time decision-making capabilities of inspection systems.

[0003] However, traditional machine vision methods rely on manually designed feature operators, which are poorly robust to interference factors such as changes in illumination, edge blurring, and local deformation, and are prone to misclassifying process noise as real defects. Furthermore, the distribution of defect samples is highly uneven, and mainstream classification models, lacking structured prior guidance, are prone to getting stuck in local optima, making it difficult to distinguish subtle but crucial defect categories (such as murmurs and minor scratches). In addition, existing deep learning solutions mostly use general convolutional architectures, failing to model the spatial distribution characteristics and physical causes of display defects, resulting in a disconnect between feature representation and defect semantics, leading to inflated classification confidence but insufficient actual generalization ability.

[0004] Therefore, a method for automatic defect classification of liquid crystal displays based on deep learning is desired. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic defect classification method for liquid crystal displays based on deep learning, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An automatic defect classification method for liquid crystal displays (LCDs) based on deep learning includes the following specific steps: Step 1: Acquire raw image data of the LCD: Images of the LCD surface are acquired using a high-resolution industrial camera under standard lighting conditions, obtaining raw images with a resolution of at least 4096 x 3072 pixels. White balance and gamma corrections are then applied to the images to eliminate uneven illumination and color deviation. Step 2: Construct a multi-scale defect perception feature extraction network: A deep convolutional neural network with an encoder-decoder structure is used. The encoder consists of 5 downsampled residual blocks, each containing 2 convolutional layers and 1 skip connection. The decoder fuses multi-level features through transposed convolutions and skip connections, outputting a defect saliency map with the same size as the input image. Step 3: Introduce a physically prior-guided attention mechanism: An empty... The inter-channel joint attention module dynamically adjusts the feature weights of different regions based on the pixel arrangement pattern of the LCD screen and the spatial distribution characteristics of typical defects, strengthening the focus on key defect areas such as murmurs, scratches, bright spots, and dark spots. Step four performs class imbalance perception loss function optimization: a weighted combination of focus loss and boundary perception loss is adopted, where focus loss focuses on hard-to-classify samples, and boundary perception loss improves the sensitivity to subtle boundary defects by calculating the gradient response of defect edge pixels. The loss weight coefficients are set to 0.7 and 0.3, respectively. Step five realizes automatic defect classification and confidence calibration: the feature map output by the network is input into the fully connected classification head, which outputs the preset probability distribution of 8 defect categories, and the confidence of the softmax output is calibrated by temperature scaling to ensure the reliability and interpretability of the classification results.

[0008] Preferably, in step one, the high-resolution industrial camera uses a global shutter CMOS sensor with a frame rate of no less than 30 frames per second, lens distortion of less than 0.1%, and the light source is an LED surface light source with uniformity greater than 95%, the color temperature is controlled at 6500 Kelvin, and the image acquisition process is carried out in a constant temperature and humidity environment, with the ambient temperature fluctuation not exceeding ±1 degree Celsius and the humidity controlled at 50% ±5%.

[0009] Preferably, in step two, the number of output channels of the 1st to 5th residual blocks of the encoder are 64, 128, 256, 512, and 1024 respectively. Each convolutional layer is followed by batch normalization and modified linear unit activation function. The decoder gradually restores the spatial resolution through 4 upsampling operations. After each upsampling, the channel is concatenated with the feature map of the corresponding encoder layer, and finally, a defect saliency map with 8 channels is output.

[0010] Preferably, in step three, the spatial-channel joint attention module first calculates the channel attention weights, generates the importance coefficients of each channel through global average pooling and a two-layer fully connected network, and then combines the spatial attention mechanism to generate a spatial weight map using a 7x7 convolution kernel. The two are multiplied and applied to the backbone features. This module is deployed after the output of the 3rd, 4th and 5th layers of the encoder to take into account both local details and global semantics.

[0011] Preferably, in step four, the focusing parameter γ of the focus loss is set to 2.0. The edge-aware loss calculates the edge gradient of the defect mask using the Sobel operator and performs mean square error calculation with the edge response predicted by the network. During training, a cosine annealing learning rate scheduling strategy is adopted, with an initial learning rate of 0.001, a minimum learning rate of 0.00001, and 200 training rounds.

[0012] Preferably, the eight defect categories preset in step five include Mura, scratches, bright spots, dark spots, foreign objects, indentations, light leaks, and pixel loss. The fully connected classification head contains two hidden layers with 512 and 256 nodes, respectively. The output layer uses softmax activation. The temperature scaling parameter T is determined by minimizing the expected calibration error on the validation set, with a typical value of 1.8.

[0013] Preferably, the method also includes post-processing optimization of the classification results: performing morphological closing operations and connected component analysis on the defect saliency map output by the network, removing isolated noise regions with an area of ​​less than 10 pixels, merging adjacent defect regions of the same type, and finally generating a structured defect detection report, which includes defect category, location coordinates, area and confidence score.

[0014] Preferably, it also includes constructing a defect sample enhancement and transfer learning mechanism: for defect categories with scarce samples, a physical model-based synthesis method is used to generate realistic defect images, including simulating Gaussian blur overlay of Mura, linear structural perturbation of scratches, and local brightness enhancement of bright spots, and using weights pre-trained on large-scale general image datasets for transfer initialization to improve the model's generalization ability in small sample scenarios.

[0015] Preferably, it also includes deploying a lightweight inference engine: the trained deep neural network is compressed in size by channel pruning and 8-bit integer quantization, with a pruning ratio of 40%. After quantization, the inference latency of the model on the embedded GPU is less than 200 milliseconds and the memory usage is less than 300 megabytes, which meets the real-time detection requirements of the production line.

[0016] Preferably, the method is integrated into a fully automated LCD screen inspection production line, linked with robotic arms, conveyor belts and sorting systems. The single-screen inspection cycle is less than 1 second, the daily inspection capacity is greater than 10,000 pieces, the classification accuracy is greater than or equal to 98.5%, the missed detection rate is less than 0.3%, and the false alarm rate is less than 0.8%.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] High-precision defect classification capability:

[0019] Breakthrough in identifying minute defects: By using a multi-scale feature extraction network and a physical prior-guided attention mechanism, the semantic differences between Mura and minor defects such as slight scratches are effectively captured, and the classification accuracy is improved to over 98.5%, which is more than 25 percentage points higher than traditional manual feature methods.

[0020] Strong robustness: It can maintain stable classification performance under interference conditions such as light fluctuations, complex screen textures and blurred edges, with a false positive rate controlled below 0.8%, which is significantly better than existing general-purpose deep learning models.

[0021] Efficiently adapt to the actual needs of the production line:

[0022] Excellent real-time performance: The lightweight compressed model has an inference latency of less than 200 milliseconds on embedded devices and a single-screen detection cycle of less than 1 second, meeting the high-speed production cycle of high-generation lines;

[0023] High degree of automation integration: Seamlessly integrates with production line robotic arms and sorting systems to achieve full-process automation from image acquisition to defect classification and result output, with a daily inspection capacity of over 10,000 pieces, significantly improving production efficiency.

[0024] Its intelligence and generalization capabilities are outstanding.

[0025] Effective handling of class imbalance: Through joint optimization of focus loss and boundary awareness loss, the recognition performance of rare defective categories is significantly improved, and the standard deviation of recall rate for each category is less than 2%, avoiding model bias towards mainstream categories;

[0026] Strong generalization ability with small samples: Combining physical model-driven defect synthesis and transfer learning strategies, a high-precision classification model can be trained with only a few hundred real samples, reducing data annotation costs by more than 70%, and is suitable for the rapid import and detection of new types of displays. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the overall technical solution architecture of the present invention;

[0028] Figure 2This is a flowchart illustrating the principle of the physical prior-guided multi-scale defect perception and attention fusion mechanism in this invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0030] Currently, in the manufacturing process of liquid crystal displays (LCDs), defect detection and classification mainly rely on manual visual inspection or automated systems based on traditional image processing algorithms. This presents technical bottlenecks such as low detection accuracy, high false negative rates, and difficulty in adapting to complex defect morphologies and high-speed production lines. To address these issues, this invention proposes an automatic defect classification method for LCDs based on deep learning, and applies it to a fully automated LCD inspection production line, achieving high-precision, high-efficiency, and highly robust defect identification and classification.

[0031] Please refer to Figure 1 and Figure 2In the aforementioned deep learning-based automatic defect classification method for LCD screens, step one involves acquiring the original image data of the LCD screen: A high-resolution industrial camera is used to acquire images of the LCD screen surface under standard lighting conditions, obtaining original images with a resolution of at least 4096 x 3072 pixels. White balance and gamma corrections are then performed on the images to eliminate uneven lighting and color deviations. Specifically, in step one, the high-resolution industrial camera uses a global shutter CMOS sensor with a frame rate of at least 30 frames per second, lens distortion of less than 0.1%, and an LED surface light source with a uniformity greater than 95%. The color temperature is controlled at 6500 Kelvin, and the image acquisition process is conducted in a constant temperature and humidity environment, with ambient temperature fluctuations not exceeding ±1 degree Celsius and humidity controlled at 50% ±5%. This step first activates the industrial camera control system, fixing the camera to the end of the robotic arm or above the inspection station, ensuring the camera's optical axis is perpendicular to the LCD screen surface, the focal length is set to 150 mm, and the working distance is 800 mm to ensure the entire display area is within the clear imaging range. The camera trigger signal is synchronously issued by the production line PLC system to ensure precise alignment of each acquisition action with the conveyor belt's movement cycle, avoiding image quality degradation due to motion blur. Before image acquisition, the system executes an automatic exposure and gain adjustment program, dynamically adjusting the exposure time to 1 / 1000 second and setting the gain value to 1.0 based on the current ambient light intensity to maintain stable image brightness and prevent overexposure. The acquired raw image data is stored in RAW format on the local solid-state drive, with each image approximately 38 megabytes in size, containing uncompressed raw pixel values ​​and preserving complete color information. Subsequently, the system calls the image preprocessing module to perform white balance correction. By calculating the average RGB channel values ​​in the neutral gray area of ​​the image, a three-channel scaling factor matrix is ​​constructed to compensate for the color of the entire image, eliminating color cast caused by light source color temperature shifts. Next, gamma correction is performed, converting the input pixel values... Mapped to output value The transformation relationship is as follows: ,in A value of 2.2 is used to restore the true brightness distribution as perceived by the human eye, improving the accuracy of subsequent feature extraction. All preprocessed images are converted to 8-bit grayscale or 24-bit true-color images and stored uniformly in a designated path for subsequent network training or inference. The data flow structure for this step is: camera hardware → image acquisition driver → memory buffer → preprocessing algorithm module → storage database. Efficient communication between each stage is achieved through shared memory or message queues, ensuring that the latency of a single image processing operation is less than 50 milliseconds.

[0032] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step two involves constructing a multi-scale defect-aware feature extraction network. This network employs a deep convolutional neural network with an encoder-decoder structure. The encoder consists of five downsampled residual blocks, each containing two convolutional layers and one skip connection. The decoder fuses multi-level features through transposed convolutions and skip connections, outputting a defect saliency map with the same size as the input image. Specifically, in step two, the number of output channels for the first to fifth residual blocks of the encoder are 64, 128, 256, 512, and 1024, respectively. Each convolutional layer is followed by batch normalization and a modified linear unit activation function. The decoder gradually restores the spatial resolution through four upsampling operations. After each upsampling, the feature map is concatenated with the corresponding encoder layer's feature map, ultimately outputting a defect saliency map with eight channels. This network architecture design follows a hierarchical feature learning paradigm of "dimensionality reduction → abstraction → dimensionality increase → restoration." The encoder's first layer receives a 3-channel input image, which undergoes initial feature extraction using a 7×7 convolutional kernel with a stride of 2 and padding of 3, resulting in an output of 64 channels and a feature map of size 2048×1536. The next layer is the first residual block, which contains two 3×3 convolutional layers. Each convolutional layer is followed by batch normalization and a rectified linear unit (ReLU) activation function. The first convolutional layer has a stride of 1, and the second has a stride of 2, achieving downsampling. The number of output channels remains 64, but the feature map size is halved to 1024×768. Skip connections directly add the input feature map to the output of the second convolutional layer, mitigating the vanishing gradient problem. The second through fifth residual blocks repeat a similar structure, but the number of output channels doubles progressively, reaching 128, 256, 512, and 1024 respectively, while the spatial resolution continuously decreases. The final encoder output is a high-dimensional semantic feature map of size 64×48. The decoder starts with a 1024-channel feature map and upsamples it using deconvolution, with an upsampling factor of 2 each time to restore spatial resolution. The first upsampling yields a 512-channel feature map with a size of 128×96, which is then concatenated with the 512-channel feature map output from the fourth layer of the encoder to form a 1024-channel fused feature map. This process is repeated in the subsequent three upsampling operations, concatenating with the outputs of the third, second, and first layers of the encoder respectively, achieving deep fusion of multi-scale contextual information. Finally, the decoder outputs an 8-channel defect saliency map, with each channel corresponding to the response intensity of a type of defect, such as mura, scratches, or bright spots. Its spatial dimension is consistent with the input image, i.e., 4096×3072, and it is used for subsequent classification tasks. During network training, the Adam optimizer is used with a batch size of 8, an initial learning rate of 0.001, momentum parameters β1=0.9, β2=0.999, and a weight decay coefficient of 0.0001 to prevent overfitting.The model has a total of approximately 18 million parameters, and the training process is carried out on an NVIDIA A100 GPU, with each training round taking about 120 seconds.

[0033] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step three introduces a physical prior-guided attention mechanism: a spatial-channel joint attention module is embedded in the network. This module dynamically adjusts the feature weights of different regions based on the pixel arrangement patterns of the LCD display and the spatial distribution characteristics of typical defects, strengthening the focus on key defect regions such as murmurs, scratches, bright spots, and dark spots. Specifically, in step three, the spatial-channel joint attention module first calculates the channel attention weights, generating the importance coefficients of each channel through global average pooling and a two-layer fully connected network. Then, combined with the spatial attention mechanism, a spatial weight map is generated using a 7x7 convolutional kernel. The two are multiplied and applied to the backbone features. This module is deployed after the outputs of layers 3, 4, and 5 of the encoder to balance local details and global semantics. The core design of this module is to transform the physical characteristics of the LCD display into learnable attention constraints. For the channel attention branch, the input feature map undergoes global average pooling to compress the spatial dimension to 1×1, resulting in a vector with a length equal to the number of channels, for example, a 1024-dimensional vector at the output of layer 5 of the encoder. The input vector is fed into the first fully connected layer (512 nodes, ReLU activation function), then into the second fully connected layer (1024 nodes, Sigmoid activation function). The output is the attention weight vector for each channel, ranging from [0,1], representing the importance of that channel to the current task. For the spatial attention branch, the input feature map is processed by a 7×7 convolution kernel (1 kernel), and the output is a single-channel spatial weight map, also ranging from [0,1], reflecting the attention intensity at different locations in the image. This spatial weight map is upsampled to the same spatial size as the input feature map using bilinear interpolation. Finally, the channel attention weights and spatial attention weights are multiplied element-wise to obtain a joint attention weight map, which is then multiplied channel-wise with the original feature map to achieve feature reweighting. This module is deployed after the outputs of the encoder's 3rd (256 channels), 4th (512 channels), and 5th (1024 channels), enabling the network to obtain physically prior-guided attention enhancement at different levels of abstraction. For example, in layer 3, the attention mechanism tends to focus on areas with obvious texture changes, which may correspond to early scratches or foreign objects; in layer 5, the attention focuses on areas with overall abnormal brightness, which may correspond to large-area mura or light leakage. This layered deployment strategy effectively improves the network's sensitivity to different types of defects, especially performing well in low-contrast backgrounds.

[0034] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step four involves optimizing the loss function to address class imbalance: a weighted combination of focal loss and boundary-aware loss is used. Focal loss focuses on hard-to-classify samples, while boundary-aware loss improves sensitivity to subtle boundary defects by calculating the gradient response of pixels at defect edges. The weight coefficients are set to 0.7 and 0.3, respectively. Specifically, in step four, the focal parameter γ for focal loss is set to 2.0. The boundary-aware loss calculates the edge gradient of the defect mask using the Sobel operator and performs mean squared error calculation with the edge response predicted by the network. During training, a cosine annealing learning rate scheduling strategy is used, with an initial learning rate of 0.001, a minimum learning rate of 0.00001, and 200 training epochs. This step aims to address the problem of a severe imbalance in the number of defect samples across different classes in the training data; for example, there are far more Mura defect samples than indentation or pixel missing samples. Focal loss is defined as:

[0035]

[0036] in For the model to the first The probability of predicting a pixel as positive. To focus the parameters, a value of 2.0 is used to reduce the weight of easily classified samples, forcing the model to pay more attention to difficult-to-classify samples. Boundary-Aware Loss specifically targets the fine structure of defect edges. First, the Sobel operator is used to convolve the real defect mask, calculating its horizontal and vertical gradients to obtain the edge gradient map. Then, the same Sobel operator is applied to the defect saliency map predicted by the network to obtain the predicted edge response map. The mean squared error (MSE) between the two is used as the boundary-aware loss. The formula for calculating the mean squared error is:

[0037]

[0038] in, Indicates the first Edge gradient map of 1 pixel, Indicates the first Predicted edge response map of pixels.

[0039] The final total loss function is:

[0040]

[0041] This weighted combination strategy focuses on overall classification accuracy in the early stages of training, while emphasizing precise matching of edge details in later stages. The training process employs cosine annealing for learning rate scheduling, with the learning rate decaying according to a cosine function with each training epoch. The formula for calculating the cosine function is:

[0042]

[0043] in =0.001, =0.00001, =200, This represents the current training epoch. This strategy helps the model converge smoothly in the later stages of training, avoiding oscillations. The training dataset contains over 50,000 labeled images, divided into a training set (70%), a validation set (15%), and a test set (15%). Data augmentation techniques (random rotation, scaling, flipping, and brightness perturbation) are used to increase sample diversity and prevent overfitting.

[0044] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step five involves automatic defect classification and confidence calibration: the feature map output by the network is input into a fully connected classification head, which outputs a preset probability distribution of eight defect categories. The softmax output is then calibrated using a temperature scaling method to ensure the reliability and interpretability of the classification results. Specifically, the eight preset defect categories in step five include mura, scratches, bright spots, dark spots, foreign objects, indentations, light leakage, and pixel loss. The fully connected classification head contains two hidden layers with 512 and 256 nodes respectively. The output layer uses softmax activation, and the temperature scaling parameter T is determined by minimizing the expected calibration error on the validation set, typically taking a value of 1.8. This step is the decision endpoint of the entire method. The eight-channel defect saliency map output by the network is flattened into a one-dimensional vector and input into the fully connected classification head. The first fully connected layer contains 512 nodes with ReLU activation, the second layer contains 256 nodes with ReLU activation, and the final output layer contains 8 nodes, corresponding to 8 defect categories, with softmax activation, outputting the probability value of each category, summed to 1. For example, the output of a pixel might be [0.05, 0.12, 0.78, 0.03, 0.01, 0.01, 0.01, 0.01], indicating that the pixel most likely belongs to the "bright spot" category with a confidence level of 78%. However, the softmax output of deep neural networks is often overconfident, meaning that high-confidence predictions may be incorrect. Therefore, temperature scaling is introduced for calibration. Assuming the original softmax output is... ,in For logits. The output after temperature scaling is... ,in This is a temperature parameter. This parameter is determined by minimizing the Expected Calibration Error (ECE) on the validation set, where ECE is defined as:

[0045]

[0046] in This represents the number of confidence intervals. For the first Number of samples within the interval The percentage of correctly predicted values ​​within that interval. This represents the average confidence level within that interval. Optimization can be achieved using grid search or gradient descent. To minimize its ECE. Experiments show that when At a value of 1.8, the ECE reaches its minimum, significantly improving the reliability of the classification results. The calibrated output can be used not only for final decision-making but also as a quality assessment indicator to guide subsequent manual review processes.

[0047] Furthermore, the post-processing optimization of the classification results includes performing morphological closing operations and connected component analysis on the defect saliency map output by the network. Isolated noise regions with an area less than 10 pixels are removed, and adjacent defect regions of the same type are merged to generate a structured defect detection report containing the defect category, location coordinates, area, and confidence score. Specifically, this post-processing workflow first binarizes the defect saliency map, setting a threshold of 0.5, and marking pixels above the threshold as defect regions. Then, morphological closing operations (dilation followed by erosion) are performed on the binary image, using 3×3 square structuring elements to fill small holes inside the defects, maintaining the integrity of the defect outline. Next, connected component analysis is performed, traversing each foreground pixel in the image and using four-neighbor or eight-neighbor connection rules to group interconnected pixels into a connected component. For each connected component, its area (number of pixels) is calculated; if the area is less than 10, it is marked as noise and deleted. For connected components with an area greater than or equal to 10, their geometric center coordinates (centroid) are calculated as the defect location. If the Euclidean distance between two connected components of the same category is less than 50 pixels, they are considered different parts of the same defect, merged, and their total area and center coordinates are updated. Finally, all valid defect information is integrated into a structured report in JSON format, with fields including: defect_id, category, x_center, y_center, area, confidence_score, and bounding_box, facilitating parsing and storage by downstream systems.

[0048] Furthermore, it includes constructing a defect sample enhancement and transfer learning mechanism: for defect categories with scarce samples, a physical model-based synthesis method is used to generate realistic defect images, including simulating Mura's Gaussian blur overlay, linear structural perturbation of scratches, and local brightness enhancement of bright spots. Transfer initialization is performed using weights pre-trained on large-scale general image datasets to improve the model's generalization ability in small-sample scenarios. Specifically, for Mura defects, a two-dimensional Gaussian function is used. ,in As the defect center, The diffusion radius, typically randomly selected between 50 and 200 pixels, is used to overlay a Gaussian image onto a normal image to simulate uneven brightness. For scratches, one or more straight lines or curves with a width of 1 to 3 pixels are generated, and linear gradient perturbations are applied along their paths to change pixel values, simulating the visual effect of scratches. For bright spots, a random location is selected, and brightness is radiated outward from that point to form a locally bright area. These synthetic images are mixed with real images and used to expand the training dataset. Simultaneously, a transfer learning strategy is employed. The weights of a ResNet-50 model pre-trained on the ImageNet dataset are loaded into the encoder part of this invention, freezing only the first few layers and allowing subsequent layers to update freely. This leverages general visual features to accelerate convergence and reduces reliance on large amounts of labeled data.

[0049] Furthermore, it includes the deployment of a lightweight inference engine: the trained deep neural network is compressed in size through channel pruning and 8-bit integer quantization, with a pruning ratio of 40%. After quantization, the inference latency on embedded GPUs is less than 200 milliseconds, and the memory usage is less than 300 megabytes, meeting the real-time detection requirements of production lines. Specifically, the channel pruning process is based on the L1 norm, calculating the sum of the absolute values ​​of the weights of the output channels of each convolutional layer, sorting them, and removing the 40% of channels with the smallest contribution. At the same time, the number of input channels in subsequent layers is adjusted to maintain network topology consistency. After pruning, the number of model parameters is reduced to approximately 10.8 million. Subsequently, 8-bit integer quantization is performed, mapping floating-point weights and activation values ​​to the integer range of 0 to 255, and achieving fast computation through a lookup table. The quantization process adopts a post-training quantization strategy, calibrated on a validation set to ensure that the accuracy loss is less than 1%. The final model file size is approximately 120 megabytes, which can run on embedded devices such as NVIDIA Jetson AGX Orin, with an inference latency stable within 180 milliseconds, meeting the requirement of a single-screen detection cycle of less than 1 second.

[0050] Furthermore, the method is integrated into a fully automated LCD screen inspection production line, linking with robotic arms, conveyor belts, and sorting systems. The single-screen inspection cycle is less than 1 second, with a daily inspection capacity exceeding 10,000 pieces, a classification accuracy of ≥98.5%, a false negative rate of less than 0.3%, and a false alarm rate of less than 0.8%. Specifically, the production line control system uses a Siemens S7-1500 PLC as the core controller, communicating with industrial cameras, robotic arms, conveyor belt motors, and sorting cylinders via EtherCAT bus. When the conveyor belt delivers the screen to be inspected to the inspection station, the PLC triggers the camera to take a picture, simultaneously recording the screen number and position information. The image data is transmitted to an edge computing server via gigabit Ethernet, where a lightweight inference engine performs defect classification. The classification results are returned to the PLC via the OPC UA protocol. If a defect is detected, the PLC controls the robotic arm to grab the screen, move it to the sorting area, and push it to the defective product collection box by a cylinder; if no defect is found, it continues to the next process. The entire process is a closed-loop control, ensuring high throughput and high stability. The system processes an average of 12,000 screens per day, with a classification accuracy of 98.7%, a missed detection rate of 0.25%, and a false alarm rate of 0.7%, fully meeting the quality control requirements of high-end display panel production lines.

[0051] Example 2

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0053] Currently, in the manufacturing process of liquid crystal displays (LCDs), defect detection and classification mainly rely on manual visual inspection or automated systems based on traditional image processing algorithms. This presents technical bottlenecks such as low detection accuracy, high false negative rates, and difficulty in adapting to complex defect morphologies and high-speed production lines. To address these issues, this invention proposes an automatic defect classification method for LCDs based on deep learning, and applies it to a fully automated LCD inspection production line, achieving high-precision, high-efficiency, and highly robust defect identification and classification.

[0054] In the aforementioned deep learning-based automatic defect classification method for LCD screens, step one involves acquiring the original image data of the LCD screen: A high-resolution industrial camera is used to acquire images of the LCD screen surface under standard lighting conditions, obtaining original images with a resolution of at least 4096 x 3072 pixels. White balance and gamma corrections are then performed on the images to eliminate uneven lighting and color deviations. Specifically, in step one, the high-resolution industrial camera uses a global shutter CMOS sensor with a frame rate of at least 30 frames per second, lens distortion of less than 0.1%, and an LED surface light source with a uniformity greater than 95%. The color temperature is controlled at 650 Kelvin. The image acquisition process is conducted in a constant temperature and humidity environment, with ambient temperature fluctuations not exceeding ±1 degree Celsius and humidity controlled at 50% ±5%. This step first activates the industrial camera control system, fixing the camera to the end of the robotic arm or above the inspection station, ensuring the camera's optical axis is perpendicular to the LCD screen surface, the focal length is set to 150 mm, and the working distance is 800 mm to ensure the entire display area is within the clear imaging range. The camera trigger signal is synchronously issued by the production line PLC system to ensure precise alignment of each acquisition action with the conveyor belt's movement cycle, avoiding image quality degradation due to motion blur. Before image acquisition, the system executes an automatic exposure and gain adjustment program, dynamically adjusting the exposure time to 1 / 1000 second and setting the gain value to 1.0 based on the current ambient light intensity to maintain stable image brightness and prevent overexposure. The acquired raw image data is stored in RAW format on the local solid-state drive, with each image approximately 38 megabytes in size, containing uncompressed raw pixel values ​​and preserving complete color information. Subsequently, the system calls the image preprocessing module to perform white balance correction. By calculating the average RGB channel values ​​in the neutral gray area of ​​the image, a three-channel scaling factor matrix is ​​constructed to compensate for the color of the entire image, eliminating color cast caused by light source color temperature shifts. Next, gamma correction is performed, converting the input pixel values... Mapped to output value The transformation relationship is as follows: ,in A value of 2.2 is used to restore the true brightness distribution as perceived by the human eye, improving the accuracy of subsequent feature extraction. All preprocessed images are converted to 8-bit grayscale or 24-bit true-color images and stored uniformly in a designated path for subsequent network training or inference. The data flow structure for this step is: camera hardware → image acquisition driver → memory buffer → preprocessing algorithm module → storage database. Efficient communication between each stage is achieved through shared memory or message queues, ensuring that the latency of a single image processing operation is less than 50 milliseconds.

[0055] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step two involves constructing a multi-scale defect-aware feature extraction network. This network employs a deep convolutional neural network with an encoder-decoder structure. The encoder consists of five downsampled residual blocks, each containing two convolutional layers and one skip connection. The decoder fuses multi-level features through transposed convolutions and skip connections, outputting a defect saliency map with the same size as the input image. Specifically, in step two, the number of output channels for the first to fifth residual blocks of the encoder are 64, 128, 256, 512, and 1024, respectively. Each convolutional layer is followed by batch normalization and a modified linear unit activation function. The decoder gradually restores the spatial resolution through four upsampling operations. After each upsampling, the feature map is concatenated with the corresponding encoder layer's feature map, ultimately outputting a defect saliency map with eight channels. This network architecture design follows a hierarchical feature learning paradigm of "dimensionality reduction → abstraction → dimensionality increase → restoration." The encoder's first layer receives a 3-channel input image, which undergoes initial feature extraction using a 7×7 convolutional kernel with a stride of 2 and padding of 3, resulting in an output of 64 channels and a feature map of size 2048×1536. The next layer is the first residual block, which contains two 3×3 convolutional layers. Each convolutional layer is followed by batch normalization and a rectified linear unit (ReLU) activation function. The first convolutional layer has a stride of 1, and the second has a stride of 2, achieving downsampling. The number of output channels remains 64, but the feature map size is halved to 1024×768. Skip connections directly add the input feature map to the output of the second convolutional layer, mitigating the vanishing gradient problem. The second through fifth residual blocks repeat a similar structure, but the number of output channels doubles progressively, reaching 128, 256, 512, and 1024 respectively, while the spatial resolution continuously decreases. The final encoder output is a high-dimensional semantic feature map of size 64×48. The decoder starts with a 1024-channel feature map and upsamples it using deconvolution, with an upsampling factor of 2 each time to restore spatial resolution. The first upsampling yields a 512-channel feature map with a size of 128×96, which is then concatenated with the 512-channel feature map output from the fourth layer of the encoder to form a 1024-channel fused feature map. This process is repeated in the subsequent three upsampling operations, concatenating with the outputs of the third, second, and first layers of the encoder respectively, achieving deep fusion of multi-scale contextual information. Finally, the decoder outputs an 8-channel defect saliency map, with each channel corresponding to the response intensity of a type of defect, such as mura, scratches, or bright spots. Its spatial dimension is consistent with the input image, i.e., 4096×3072, and it is used for subsequent classification tasks. During network training, the Adam optimizer is used with a batch size of 8, an initial learning rate of 0.001, momentum parameters β1=0.9, β2=0.999, and a weight decay coefficient of 0.0001 to prevent overfitting.The model has a total of approximately 18 million parameters, and the training process is carried out on an NVIDIA A100 GPU, with each training round taking about 120 seconds.

[0056] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step three introduces a physical prior-guided attention mechanism: a spatial-channel joint attention module is embedded in the network. This module dynamically adjusts the feature weights of different regions based on the pixel arrangement patterns of the LCD display and the spatial distribution characteristics of typical defects, strengthening the focus on key defect regions such as murmurs, scratches, bright spots, and dark spots. Specifically, in step three, the spatial-channel joint attention module first calculates the channel attention weights, generating the importance coefficients of each channel through global average pooling and a two-layer fully connected network. Then, combined with the spatial attention mechanism, a spatial weight map is generated using a 7x7 convolutional kernel. The two are multiplied and applied to the backbone features. This module is deployed after the outputs of layers 3, 4, and 5 of the encoder to balance local details and global semantics. The core design of this module is to transform the physical characteristics of the LCD display into learnable attention constraints. For the channel attention branch, the input feature map undergoes global average pooling to compress the spatial dimension to 1×1, resulting in a vector with a length equal to the number of channels, for example, a 1024-dimensional vector at the output of layer 5 of the encoder. The input vector is fed into the first fully connected layer (512 nodes, ReLU activation function), then into the second fully connected layer (1024 nodes, Sigmoid activation function). The output is the attention weight vector for each channel, ranging from [0,1], representing the importance of that channel to the current task. For the spatial attention branch, the input feature map is processed by a 7×7 convolution kernel (1 kernel), and the output is a single-channel spatial weight map, also ranging from [0,1], reflecting the attention intensity at different locations in the image. This spatial weight map is upsampled to the same spatial size as the input feature map using bilinear interpolation. Finally, the channel attention weights and spatial attention weights are multiplied element-wise to obtain a joint attention weight map, which is then multiplied channel-wise with the original feature map to achieve feature reweighting. This module is deployed after the outputs of the encoder's 3rd (256 channels), 4th (512 channels), and 5th (1024 channels), enabling the network to obtain physically prior-guided attention enhancement at different levels of abstraction. For example, in layer 3, the attention mechanism tends to focus on areas with obvious texture changes, which may correspond to early scratches or foreign objects; in layer 5, the attention focuses on areas with overall abnormal brightness, which may correspond to large-area mura or light leakage. This layered deployment strategy effectively improves the network's sensitivity to different types of defects, especially performing well in low-contrast backgrounds.

[0057] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step four involves optimizing the loss function to address class imbalance: a weighted combination of focal loss and boundary-aware loss is used. Focal loss focuses on hard-to-classify samples, while boundary-aware loss improves sensitivity to subtle boundary defects by calculating the gradient response of pixels at defect edges. The weight coefficients are set to 0.7 and 0.3, respectively. Specifically, in step four, the focal parameter γ for focal loss is set to 2.0. The boundary-aware loss calculates the edge gradient of the defect mask using the Sobel operator and performs mean squared error calculation with the edge response predicted by the network. During training, a cosine annealing learning rate scheduling strategy is used, with an initial learning rate of 0.001, a minimum learning rate of 0.00001, and 200 training epochs. This step aims to address the problem of a severe imbalance in the number of defect samples across different classes in the training data; for example, there are far more Mura defect samples than indentation or pixel missing samples. Focal loss is defined as:

[0058]

[0059] in For the model to the first The probability of predicting a pixel as positive. To focus the parameters, a value of 2.0 is used to reduce the weight of easily classified samples, forcing the model to pay more attention to difficult-to-classify samples. Boundary-Aware Loss specifically targets the fine structure of defect edges. First, the Sobel operator is used to convolve the real defect mask, calculating its horizontal and vertical gradients to obtain the edge gradient map. Then, the same Sobel operator is applied to the defect saliency map predicted by the network to obtain the predicted edge response map. The mean squared error (MSE) between the two is used as the boundary-aware loss. The formula for calculating the mean squared error is:

[0060]

[0061] in, Indicates the first Edge gradient map of 1 pixel, Indicates the first Predicted edge response map of pixels.

[0062] The final total loss function is:

[0063]

[0064] This weighted combination strategy focuses on overall classification accuracy in the early stages of training, while emphasizing precise matching of edge details in later stages. The training process employs cosine annealing for learning rate scheduling, with the learning rate decaying according to a cosine function with each training epoch. The formula for calculating the cosine function is:

[0065]

[0066] in =0.001, =0.00001, =200, This represents the current training epoch. This strategy helps the model converge smoothly in the later stages of training, avoiding oscillations. The training dataset contains over 50,000 labeled images, divided into a training set (70%), a validation set (15%), and a test set (15%). Data augmentation techniques (random rotation, scaling, flipping, and brightness perturbation) are used to increase sample diversity and prevent overfitting.

[0067] In the aforementioned deep learning-based automatic defect classification method for LCD displays, step five involves automatic defect classification and confidence calibration: the feature map output by the network is input into a fully connected classification head, which outputs a preset probability distribution of eight defect categories. The softmax output is then calibrated using a temperature scaling method to ensure the reliability and interpretability of the classification results. Specifically, the eight preset defect categories in step five include mura, scratches, bright spots, dark spots, foreign objects, indentations, light leakage, and pixel loss. The fully connected classification head contains two hidden layers with 512 and 256 nodes respectively. The output layer uses softmax activation, and the temperature scaling parameter T is determined by minimizing the expected calibration error on the validation set, typically taking a value of 1.8. This step is the decision endpoint of the entire method. The eight-channel defect saliency map output by the network is flattened into a one-dimensional vector and input into the fully connected classification head. The first fully connected layer contains 512 nodes with ReLU activation, the second layer contains 256 nodes with ReLU activation, and the final output layer contains 8 nodes, corresponding to 8 defect categories, with softmax activation, outputting the probability value of each category, summed to 1. For example, the output of a pixel might be [0.05, 0.12, 0.78, 0.03, 0.01, 0.01, 0.01, 0.01], indicating that the pixel most likely belongs to the "bright spot" category with a confidence level of 78%. However, the softmax output of deep neural networks is often overconfident, meaning that high-confidence predictions may be incorrect. Therefore, temperature scaling is introduced for calibration. Assuming the original softmax output is... ,in For logits. The output after temperature scaling is... ,in This is a temperature parameter. This parameter is determined by minimizing the Expected Calibration Error (ECE) on the validation set, where ECE is defined as:

[0068]

[0069] in This represents the number of confidence intervals. For the first Number of samples within the interval The percentage of correctly predicted values ​​within that interval. This represents the average confidence level within that interval. Optimization can be achieved using grid search or gradient descent. To minimize its ECE. Experiments show that when At a value of 1.8, the ECE reaches its minimum, significantly improving the reliability of the classification results. The calibrated output can be used not only for final decision-making but also as a quality assessment indicator to guide subsequent manual review processes.

[0070] Furthermore, the post-processing optimization of the classification results includes performing morphological closing operations and connected component analysis on the defect saliency map output by the network. Isolated noise regions with an area less than 10 pixels are removed, and adjacent defect regions of the same type are merged to generate a structured defect detection report containing the defect category, location coordinates, area, and confidence score. Specifically, this post-processing workflow first binarizes the defect saliency map, setting a threshold of 0.5, and marking pixels above the threshold as defect regions. Then, morphological closing operations (dilation followed by erosion) are performed on the binary image, using 3×3 square structuring elements to fill small holes inside the defects, maintaining the integrity of the defect outline. Next, connected component analysis is performed, traversing each foreground pixel in the image and using four-neighbor or eight-neighbor connection rules to group interconnected pixels into a connected component. For each connected component, its area (number of pixels) is calculated; if the area is less than 10, it is marked as noise and deleted. For connected components with an area greater than or equal to 10, their geometric center coordinates (centroid) are calculated as the defect location. If the Euclidean distance between two connected components of the same category is less than 50 pixels, they are considered different parts of the same defect, merged, and their total area and center coordinates are updated. Finally, all valid defect information is integrated into a structured report in JSON format, with fields including: defect_id, category, x_center, y_center, area, confidence_score, and bounding_box, facilitating parsing and storage by downstream systems.

[0071] Furthermore, it includes constructing a defect sample enhancement and transfer learning mechanism: for defect categories with scarce samples, a physical model-based synthesis method is used to generate realistic defect images, including simulating Mura's Gaussian blur overlay, linear structural perturbation of scratches, and local brightness enhancement of bright spots. Transfer initialization is performed using weights pre-trained on large-scale general image datasets to improve the model's generalization ability in small-sample scenarios. Specifically, for Mura defects, a two-dimensional Gaussian function is used. ,in As the defect center, The diffusion radius, typically randomly selected between 50 and 200 pixels, is used to overlay a Gaussian image onto a normal image to simulate uneven brightness. For scratches, one or more straight lines or curves with a width of 1 to 3 pixels are generated, and linear gradient perturbations are applied along their paths to change pixel values, simulating the visual effect of scratches. For bright spots, a random location is selected, and brightness is radiated outward from that point to form a locally bright area. These synthetic images are mixed with real images and used to expand the training dataset. Simultaneously, a transfer learning strategy is employed. The weights of a ResNet-50 model pre-trained on the ImageNet dataset are loaded into the encoder part of this invention, freezing only the first few layers and allowing subsequent layers to update freely. This leverages general visual features to accelerate convergence and reduces reliance on large amounts of labeled data.

[0072] Furthermore, it includes the deployment of a lightweight inference engine: the trained deep neural network is compressed in size through channel pruning and 8-bit integer quantization, with a pruning ratio of 40%. After quantization, the inference latency on embedded GPUs is less than 200 milliseconds, and the memory usage is less than 300 megabytes, meeting the real-time detection requirements of production lines. Specifically, the channel pruning process is based on the L1 norm, calculating the sum of the absolute values ​​of the weights of the output channels of each convolutional layer, sorting them, and removing the 40% of channels with the smallest contribution. At the same time, the number of input channels in subsequent layers is adjusted to maintain network topology consistency. After pruning, the number of model parameters is reduced to approximately 10.8 million. Subsequently, 8-bit integer quantization is performed, mapping floating-point weights and activation values ​​to the integer range of 0 to 255, and achieving fast computation through a lookup table. The quantization process adopts a post-training quantization strategy, calibrated on a validation set to ensure that the accuracy loss is less than 1%. The final model file size is approximately 120 megabytes, which can run on embedded devices such as NVIDIA Jetson AGX Orin, with an inference latency stable within 180 milliseconds, meeting the requirement of a single-screen detection cycle of less than 1 second.

[0073] Furthermore, the method is integrated into a fully automated LCD screen inspection production line, linking with robotic arms, conveyor belts, and sorting systems. The single-screen inspection cycle is less than 1 second, with a daily inspection capacity exceeding 10,000 pieces, a classification accuracy of ≥98.5%, a false negative rate of less than 0.3%, and a false alarm rate of less than 0.8%. Specifically, the production line control system uses a Siemens S7-1500 PLC as the core controller, communicating with industrial cameras, robotic arms, conveyor belt motors, and sorting cylinders via EtherCAT bus. When the conveyor belt delivers the screen to be inspected to the inspection station, the PLC triggers the camera to take a picture, simultaneously recording the screen number and position information. The image data is transmitted to an edge computing server via gigabit Ethernet, where a lightweight inference engine performs defect classification. The classification results are returned to the PLC via the OPC UA protocol. If a defect is detected, the PLC controls the robotic arm to grab the screen, move it to the sorting area, and push it to the defective product collection box by a cylinder; if no defect is found, it continues to the next process. The entire process is a closed-loop control, ensuring high throughput and high stability. The system processes an average of 12,000 screens per day, with a classification accuracy of 98.7%, a missed detection rate of 0.25%, and a false alarm rate of 0.7%, fully meeting the quality control requirements of high-end display panel production lines.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic defect classification of liquid crystal displays based on deep learning, characterized in that: The specific steps include the following: Step 1: Acquire raw image data of the LCD screen: Use a high-resolution industrial camera to capture images of the LCD screen surface under standard lighting conditions to obtain raw images with a resolution of no less than 4096 x 3072 pixels. Step 2: Construct a multi-scale defect perception feature extraction network: A deep convolutional neural network with an encoder-decoder structure is adopted. The encoder part consists of 5 downsampled residual blocks, each containing 2 convolutional layers and 1 skip connection. The decoder part fuses multi-level features through transposed convolution and skip connections, and outputs a defect saliency map with the same size as the input image. Step 3 introduces a physical prior-guided attention mechanism: a spatial-channel joint attention module is embedded in the network. This module dynamically adjusts the feature weights of different regions based on the pixel arrangement pattern of the liquid crystal display screen and the spatial distribution characteristics of typical defects. Step 4: Optimize the loss function for class imbalance awareness: Use a weighted combination of focus loss and boundary awareness loss, where focus loss focuses on hard-to-classify samples, and boundary awareness loss calculates the gradient response of defect edge pixels; Step 5: Implement automatic defect classification and confidence calibration: Input the feature map output by the network into the fully connected classification head, and output the preset probability distribution of 8 defect categories; In step three, the spatial-channel joint attention module first calculates the channel attention weights, generates the importance coefficients of each channel through global average pooling and two fully connected networks, and then combines the spatial attention mechanism to generate a spatial weight map using a 7x7 convolution kernel. The two are multiplied and applied to the backbone features. This module is deployed after the output of the 3rd, 4th and 5th layers of the encoder to take into account both local details and global semantics. In step four, the focusing parameter γ of the focus loss is set to 2.

0. The edge-aware loss calculates the edge gradient of the defect mask using the Sobel operator and performs mean square error calculation with the edge response predicted by the network. During training, a cosine annealing learning rate scheduling strategy is adopted, with an initial learning rate of 0.001, a minimum learning rate of 0.00001, and 200 training rounds. The eight defect categories preset in step five include Mura, scratches, bright spots, dark spots, foreign objects, indentations, light leaks, and missing pixels. The fully connected classification head contains two hidden layers with 512 and 256 nodes respectively. The output layer uses softmax activation. The temperature scaling parameter T is determined by minimizing the expected calibration error on the validation set, with a typical value of 1.

8. It also includes post-processing optimization of the classification results: performing morphological closing operations and connected component analysis on the defect saliency map output by the network, removing isolated noise regions with an area of ​​less than 10 pixels, and merging adjacent defect regions of the same type to finally generate a structured defect detection report, including defect category, location coordinates, area and confidence score. It also includes building a defect sample enhancement and transfer learning mechanism: for defect categories with scarce samples, a physical model-based synthesis method is used to generate realistic defect images, including simulating Gaussian blur overlay of Mura, linear structural perturbation of scratches, and local brightness enhancement of bright spots. The weights pre-trained on a large-scale general image dataset are used for transfer initialization to improve the model's generalization ability in small sample scenarios.

2. The method for automatic defect classification of liquid crystal displays based on deep learning according to claim 1, characterized in that: In step one, the high-resolution industrial camera uses a global shutter CMOS sensor with a frame rate of no less than 30 frames per second, lens distortion of less than 0.1%, and a light source with a uniformity of more than 95% LED surface light source. The color temperature is controlled at 6500 Kelvin. The image acquisition process is carried out in a constant temperature and humidity environment, with the ambient temperature fluctuation not exceeding ±1 degree Celsius and the humidity controlled at 50% ±5%.

3. The method for automatic defect classification of liquid crystal displays based on deep learning according to claim 1, characterized in that: In step two, the number of output channels for the 1st to 5th residual blocks of the encoder are 64, 128, 256, 512, and 1024 respectively. Each convolutional layer is followed by batch normalization and a modified linear unit activation function. The decoder gradually restores the spatial resolution through 4 upsampling operations. After each upsampling, the feature map of the corresponding encoder layer is concatenated with the channel data, and finally, a defect saliency map with 8 channels is output.

4. The method for automatic defect classification of liquid crystal displays based on deep learning according to claim 1, characterized in that: It also includes deploying a lightweight inference engine: the trained deep neural network is compressed in size through channel pruning and 8-bit integer quantization, with a pruning ratio of 40%. After quantization, the inference latency of the model on the embedded GPU is less than 200 milliseconds and the memory usage is less than 300 megabytes, meeting the real-time detection requirements of the production line.

5. The method for automatic defect classification of liquid crystal displays based on deep learning according to claim 1, characterized in that: It is also integrated into a fully automated LCD screen inspection production line, working in conjunction with robotic arms, conveyor belts, and sorting systems. The single-screen inspection cycle is less than 1 second, the daily inspection capacity is greater than 10,000 pieces, the classification accuracy is greater than or equal to 98.5%, the missed detection rate is less than 0.3%, and the false alarm rate is less than 0.8%.

Images