Liquid crystal screen defect detection method based on improved YOLO-DEA

Abstract

This invention discloses a liquid crystal display (LCD) screen defect detection method based on an improved YOLO-DEA model. The method includes: acquiring an image to be detected; inputting the image to be detected into a trained improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: inputting the image to be detected into the processing layers of the backbone network for sequential processing to obtain multi-scale feature maps; inputting the multi-scale feature maps into the feature fusion network to obtain multi-scale fused features; and inputting the multi-scale fused features into the detection head to obtain the LCD screen defect detection result. This invention can alleviate the problems of insufficient feature representation and information distortion in traditional detection methods based on rigidity assumptions under curved scenarios, achieving accurate localization and stable identification of minute defects under curved interface conditions, and improving the accuracy and stability of LCD screen defect detection.

Description

Technical Field

[0001] This invention relates to the field of screen defect detection technology, and in particular to a liquid crystal screen defect detection method, system, storage medium, and electronic device based on an improved YOLO-DEA. Background Technology

[0002] As a core display component of modern electronic products, the manufacturing quality of LCD screens directly affects the display effect and user experience of the final product. During automated production, factors such as equipment processing precision, fluctuations in process parameters, and environmental disturbances can easily lead to various defects on the surface of LCD screens, including point defects, line defects, leakage, breakage, and cracks. Traditional manual visual inspection methods suffer from low efficiency, high subjectivity, high missed detection rates, and high labor costs, making them unsuitable for the quality inspection needs of modern large-scale, automated production.

[0003] With the rapid development of computer vision and deep learning technologies, deep learning-based defect detection methods are gradually being applied in industrial scenarios, demonstrating significant advantages in detection speed, stability, and consistency. However, facing the trend of LCD manufacturing towards flexibility and diversification, including complex application scenarios such as different sizes, planar and curved multi-form structures, and multi-angle installation postures, existing detection algorithms still face many challenges in practical applications.

[0004] In recent years, deep learning research on LCD defect detection has been relatively limited. However, in other industrial surface defect detection fields, scholars both domestically and internationally have conducted extensive research and achieved a series of results. For example, the CE-SGNet network enhances the correlation modeling ability between defect regions by introducing a graph attention mechanism, but its model structure is complex and its inference speed is less than 15 FPS, making it difficult to meet the application requirements of real-time industrial production lines. Some scholars have proposed a lightweight detection method for steel surface defects by integrating VanillaNet modules with a coordinate attention mechanism, which significantly improves target localization accuracy while reducing model complexity.

[0005] While the aforementioned research has made significant progress in its respective application fields, the inspection of flexible LCD screens still possesses unique complexities. The core challenge lies in the need for inspection systems to overcome their dependence on fixed sizes and rigid structures, achieving universal adaptability to various LCD screen shapes. This goal is constrained by factors such as algorithm design, hardware computing power, and dynamic inspection scenarios. First, LCD screens of different sizes and their defects exhibit significant pixel-level scale differences in the imaging space. Traditional convolutional neural networks, due to their fixed receptive fields, struggle to effectively capture sub-pixel-level defect features across scales. Furthermore, during deep feature transmission, information about small targets is easily weakened or even lost, leading to missed detections of scattered or subtle defects. Second, the geometric distortion and imaging artifacts caused by irregularly shaped screens further exacerbate the inspection difficulty: the stretching effect caused by curvature changes at the edges of curved screens easily forms light and dark stripes resembling scratches; the shadowed areas in the hinge region of folding screens are easily misjudged as dark line defects. Algorithms with dynamic compensation capabilities are urgently needed to effectively distinguish between real defects and deformation artifacts. Meanwhile, the non-display areas in irregularly shaped screens consume a lot of detection resources. The algorithm needs to segment the effective detection area in real time under complex backgrounds and accurately model the defects and distortions on the curved surface. Traditional anchor-frame-based detection mechanisms are less adaptable to such nonlinear features. Summary of the Invention

[0006] This invention provides a liquid crystal display (LCD) defect detection method, system, storage medium, and electronic device based on an improved YOLO-DEA. It can alleviate the problems of insufficient feature representation and information distortion in traditional detection methods based on rigidity assumptions under bending scenarios, achieve accurate positioning and stable identification of minute defects under bending interface conditions, and improve the accuracy and stability of LCD defect detection.

[0007] This invention provides a defect detection method for liquid crystal displays based on an improved YOLO-DEA, comprising: Acquire the image to be detected; The image to be detected is input into a trained, improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

[0008] Furthermore, based on the above-mentioned improved YOLO-DEA-based LCD screen defect detection method, the first processing layer of the backbone network includes a CBS module; the second, third, and fourth processing layers of the backbone network each include a CBS module and a defect perception module connected in sequence; the fifth processing layer of the backbone network includes a CBS module, a defect perception module, and a pyramid pooling and feature extraction module connected in sequence; the feature fusion network includes an upsampling module, a defect perception module, and a CBS module; and the detection head includes an adaptive enhanced dual-path detection module. The processing steps of the improved YOLO-DEA model include: The input image to be detected is processed through five processing layers of the backbone network to obtain the first scale feature map, the second scale feature map, the third scale feature map, the fourth scale feature map and the fifth scale feature map respectively; The fifth-scale feature map is upsampled and then concatenated with the fourth-scale feature map before being input into a defect perception module to obtain the first intermediate feature. The first intermediate feature is then concatenated with the third-scale feature map before being input into a defect perception module to obtain the first-scale fusion feature. The first scale fusion feature is concatenated with the first intermediate feature and then input into a defect perception module to obtain the second scale fusion feature; The second-scale fusion feature is input into the CBS module and then stitched together with the fifth-scale feature map before being input into a defect perception module to obtain the third-scale fusion feature. The first-scale fusion feature, the second-scale fusion feature, and the third-scale fusion feature are respectively enhanced by an adaptive enhanced dual-path detection module to obtain a first enhanced feature, a second enhanced feature, and a third enhanced feature; the first enhanced feature, the second enhanced feature, and the third enhanced feature are subjected to dual-branch detection to obtain dual-branch detection results at three scales; the dual-branch detection results at three scales are merged to obtain the final LCD screen defect detection result.

[0009] Furthermore, in the above-mentioned improved YOLO-DEA-based liquid crystal screen defect detection method, the defect perception module includes a deformable convolution module and a dynamic feature fusion module, and the processing procedure of the defect perception module includes: By using deformable convolution modules to perform adaptive spatial sampling and geometric modeling on the input feature map, convolutional features that are robust to deformation are obtained. The convolutional features are input into the dynamic feature fusion module, where feature enhancement and redundancy reduction are performed through keyframe feature reuse and temporal propagation mechanisms, resulting in the output feature map of the defect perception module.

[0010] Furthermore, in the above-mentioned liquid crystal screen defect detection method based on the improved YOLO-DEA, the processing procedure of the deformable convolution module includes: Initialize a learnable offset, and adaptively sample the input feature map using the learnable offset; The features of the sampling points are weighted and aggregated to obtain convolutional features.

[0011] Furthermore, in the above-mentioned LCD screen defect detection method based on the improved YOLO-DEA, the processing procedure of the dynamic feature fusion module includes: The image to be detected is divided into keyframe images and non-keyframe images; The spatial position on the non-keyframe image is projected onto the corresponding position on the keyframe image using optical flow-guided spatial warp, and bilinear interpolation is performed to obtain approximate features. Scale field features are obtained by adjusting the projected features using the scale field. The scale field features and the approximate features are fused to obtain non-keyframe features; The non-critical features are fused with the convolutional features to obtain the output feature map of the defect perception module.

[0012] Furthermore, based on the above-mentioned improved YOLO-DEA-based LCD screen defect detection method, wherein, let... p=(px, py) Non-keyframe images Spatial position on q For keyframe images At the sampling position on the image, p is back-projected onto the corresponding position in the keyframe image using a two-dimensional optical flow field. It can be expressed by the following formula:

[0013] in, This represents the displacement field from frame i to frame k. Represents an optical estimation network; The approximate feature is obtained through the following formula. :

[0014] Among them, G( ) is the bilinear interpolation kernel. Features of keyframe images; G( It can be broken down into:

[0015] in, , for q The x and y coordinates, , for p The horizontal and vertical coordinates; The projected features are adjusted using the following formula:

[0016] in, For scale field characteristics, S( () is a scaling estimation network; Non-keyframe features are obtained using the following formula. : .

[0017] Furthermore, in the above-mentioned LCD screen defect detection method based on the improved YOLO-DEA, the adaptive dual-path detection module includes an SE attention mechanism, and the processing based on the SE attention mechanism includes: Global average pooling is performed on the input feature map to obtain the global description vector for each channel; The global description vector is excited to obtain the channel weight vector; The channel weight vector is broadcast according to the spatial dimension to obtain recalibrated enhanced features.

[0018] This invention also provides a liquid crystal display defect detection system based on an improved YOLO-DEA, comprising: The image acquisition module is used to acquire the image to be detected; The LCD screen defect detection module is used to input the image to be detected into a trained improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

[0019] The present invention also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute any of the above-described liquid crystal screen defect detection methods based on the improved YOLO-DEA.

[0020] The present invention also provides an electronic device, including a processor and a memory, wherein the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used in the steps of the improved YOLO-DEA-based liquid crystal screen defect detection method described in any of the preceding claims.

[0021] This invention provides a method, system, storage medium, and electronic device for detecting defects in liquid crystal displays (LCDs) based on an improved YOLO-DEA. By reconstructing the backbone network structure and introducing a geometric deformation sensing module, this invention effectively enhances the network's ability to model irregular curved surface features of flexible LCDs while ensuring model computational efficiency and deployment feasibility. This alleviates the problems of insufficient feature representation and information distortion in curved scenarios inherent in traditional detection methods based on rigidity assumptions. Furthermore, this invention designs a curvature-aware feature fusion mechanism. By constructing an adaptive interaction path between shallow detail features and deep semantic features, it specifically improves the feature representation capability of weak defects in the edges and deformation regions of curved screens, achieving accurate localization and stable identification of minute defects under curved interface conditions, thereby further improving overall detection performance and robustness. Attached Figure Description

[0022] The technical solution and other beneficial effects of the present invention will become apparent from the following detailed description of specific embodiments of the invention, in conjunction with the accompanying drawings.

[0023] Figure 1 A flowchart of a liquid crystal screen defect detection method based on an improved YOLO-DEA provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the YOLO-DEA model provided in an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram illustrating how deformable convolution, as provided in an embodiment of the present invention, uses a single thread to process multiple channels in the same group that share a common sampling offset and aggregation weights.

[0026] Figure 4 This is a schematic diagram of the dynamic feature fusion module provided in an embodiment of the present invention.

[0027] Figure 5 A flowchart of the SE attention mechanism provided in an embodiment of the present invention.

[0028] Figure 6 Examples of some typical defects provided for embodiments of the present invention.

[0029] Figure 7 This is a comparison chart showing the visualization detection effects of different detection models provided in this embodiment of the invention on a dataset of LCD screen defects.

[0030] Figure 8 The thermal map experimental results are provided by different models in the embodiments of the present invention.

[0031] Figure 9 This is a schematic diagram of the structure of the liquid crystal screen defect detection system based on the improved YOLO-DEA provided in an embodiment of the present invention.

[0032] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In liquid crystal display (LCD) defect detection tasks, the limitations of industrial computing resources and embedded hardware performance constraints, coupled with the high demands placed on model real-time performance, lightweight design, and detection accuracy by production line inspection, make the design of a high-performance defect detection algorithm that balances detection accuracy and inference efficiency of significant engineering application value. To improve the defect detection capability of flexible LCDs under complex deformation conditions, this invention provides an LCD defect detection method, system, storage medium, and electronic device based on an improved YOLO-DEA. The LCD defect detection system based on the improved YOLO-DEA provided in this invention can be integrated into an electronic device, such as a terminal or server. The terminal can include tablet computers, laptops, personal computers (PCs), microprocessor boxes, or other devices.

[0035] Please see Figure 1 , Figure 1 The flowchart illustrates a liquid crystal display (LCD) defect detection method based on an improved YOLO-DEA, provided in an embodiment of the present invention. This method, applied to electronic devices, includes the following steps: S1, acquire the image to be detected.

[0036] S2, the image to be detected is input into the trained improved YOLO-DEA model (YOLO-based Deformable Enhancement and Attention Mechanism) to obtain the LCD screen defect detection result; the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

[0037] Figure 2 This is a schematic diagram of the structure of the YOLO-DEA model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the first processing layer of the backbone network includes a CBS (Conv BatchNorm Activation) module. The second, third, and fourth processing layers of the backbone network each include a CBS module and a Defect-aware Feature Module (DFM) module connected in sequence. The fifth processing layer of the backbone network includes a CBS module, a defect-aware feature module, and a pyramid pooling and feature extraction (SPPF+DFF) module connected in sequence. The feature fusion network includes an upsampling module, a defect-aware feature module, and a CBS module. The detection head includes an Adaptive Enhancement Dual-branch Module (AEDM) module. The improved YOLO-DEA model processing procedure includes: S21, Input the image to be detected, and obtain the first scale feature map, the second scale feature map, the third scale feature map, the fourth scale feature map and the fifth scale feature map through the five processing layers of the backbone network.

[0038] It should be noted that in the backbone network, the output of the previous processing layer is used as the input of the next processing layer for processing layer by layer.

[0039] The CBS module consists of a convolution-batch normalization-activation function connected in sequence.

[0040] In one embodiment, the defect-aware module includes a deformable convolutional (DCNv4) module and a dynamic feature fusion (DFF) module, and the processing procedure of the defect-aware module includes: S211 uses a deformable convolution module to perform adaptive spatial sampling and geometric modeling on the input feature map, resulting in convolutional features that are robust to deformation.

[0041] Specifically, the processing procedure of the deformable convolution module includes the following steps: S2111, Initialize the learnable offset, and adaptively sample the input feature map using the learnable offset.

[0042] S2112, weighted aggregation of the features of the sampling points to obtain convolutional features.

[0043] The DCNv4 module systematically optimizes the sampling offset modeling and feature aggregation methods. Specifically, the DCNv4 module uses deformable convolution operations as its core, and fixes the sampling position in traditional convolution. Introducing learnable spatial offsets This enables adaptive sampling of input features, and its calculation form can be expressed as follows:

[0044] in, For convolutional features, Let X be the initial position. W( represents the input feature map) ) represents the kernel weights, and b represents the bias term offset. With modulation coefficient The input features are jointly predicted through parallel branches, thereby dynamically weighting the feature contributions of different sampling points while adjusting the spatial location.

[0045] Compared to earlier versions of DCN, DCNv4 constrains and optimizes the distribution of sampling points during the offset prediction stage, making... The learning process is more stable, avoiding feature misalignment caused by excessive offset, thus enhancing the convolution kernel's ability to characterize local geometry and edge morphology. Simultaneously, by adjusting the modulation coefficients... Through coordinated design with the feature aggregation process, the network can suppress redundant responses during the feature fusion stage, emphasizing the focus on discriminative regions and effectively improving the compactness and robustness of feature representation. In DCNv4, a single thread is used to process multiple channels sharing the same sampling offset and aggregation weights within the same group. This reduces workloads such as memory reads and bilinear interpolation coefficient calculations, and allows for the merging of multiple memory access instructions. Figure 3 The diagram illustrates how deformable convolution, as provided in this embodiment of the invention, uses a single thread to process multiple channels in the same group that share a common sampling offset and aggregation weights. Figure 3 As shown.

[0046] Furthermore, DCNv4 has further simplified and optimized the computation process for deformable convolution, reducing additional computational overhead while maintaining the aforementioned adaptive modeling capabilities. This allows it to balance detection accuracy and inference efficiency in scenarios with complex geometric deformations, non-rigid targets, and significant local morphological changes. Thanks to these improvements, DCNv4 exhibits stronger adaptability and stability in industrial inspection tasks involving bending and twisting deformations, such as flexible LCD displays.

[0047] S212, the convolutional features are input into the dynamic feature fusion module, and feature enhancement and redundancy reduction are performed through keyframe feature reuse and temporal propagation mechanism to obtain the output feature map of the defect perception module.

[0048] In online inspection scenarios for flexible LCD screens, industrial cameras typically capture product surface images in consecutive video frames. Adjacent frames exhibit strong spatiotemporal correlation; the same defect often appears in close spatial locations within a short period, and its high-level semantic representation is more stable in deep feature maps. Simultaneously, the flexible screen introduces non-rigid deformation, reflective highlights, and uneven brightness during transport, bending, or bonding to curved surfaces, causing appearance disturbances over time. Performing complete deep convolutional network inference for every frame would result in significant computational redundancy, making it difficult to meet the real-time requirements of industrial online inspection. Therefore, this embodiment introduces the Deep Feature Flow (DFF) concept, reusing semantic features of adjacent frames through "keyframe computation and non-keyframe propagation," reducing overall computational overhead while maintaining detection accuracy as much as possible. Figure 4 This is a schematic diagram of the dynamic feature fusion module provided in this embodiment of the invention. The core of DFF is that it only runs the feature network on the sparse keyframe set {k} to obtain keyframe features, and for non-keyframe i, it uses the features of the nearest keyframe to obtain approximate features through spatial propagation, and directly sends them into the task network to complete the detection, thereby avoiding repeated calculation of the deep feature extraction process.

[0049] Specifically, the processing steps of the dynamic feature fusion module include the following: S2121, divide the image to be detected into keyframe images and non-keyframe images.

[0050] For example, the partitioning can be performed using a thresholding method based on frame differences or a clustering-based method.

[0051] S2122, using optical flow-guided spatial warp to project the spatial location on the non-keyframe image to the corresponding location on the keyframe image, and performing bilinear interpolation to obtain approximate features.

[0052] Since convolutional feature maps correspond to the input image in spatial dimensions, DFF uses optical flow-guided spatial warping to align and propagate keyframe features to non-keyframes. Let... p=(px,py) Non-keyframe images Spatial position on q For keyframe images At the sampling position on the image, p is back-projected onto the corresponding position in the keyframe image using a two-dimensional optical flow field. It can be expressed by the following formula:

[0053] in, This represents the displacement field from frame i to frame k. Represents an optical estimation network; The approximate feature is obtained through the following formula. :

[0054] Among them, G( ) is the bilinear interpolation kernel. Features of keyframe images; G( It can be broken down into:

[0055] in, , for q The x and y coordinates, , for p The horizontal and vertical coordinates.

[0056] This interpolation method involves only a small number of non-zero terms in the local neighborhood, has low computational cost, and allows gradients to propagate, making it easy to optimize the network together with subsequent tasks during end-to-end training.

[0057] S2123, scale field features are obtained by adjusting the projected features through the scale field.

[0058] In flexible LCD video inspection, factors such as non-rigid deformation caused by bending, localized strong reflections, and target occlusion / disappearance can lead to deviations in optical flow estimation, resulting in alignment errors and amplitude fluctuations in feature propagation that relies solely on spatial warps. To improve the stability and robustness of propagated features, DFF introduces a scale field, which has the same spatial dimensions and channel dimensions as the feature map, for position-by-position, channel-by-channel amplitude modulation of the propagated features. Specifically, the projected features are adjusted using the following formula:

[0059] in, For scale field characteristics, S( ) is a scaling estimation network.

[0060] S2124, fuses scale field features and approximate features to obtain non-keyframe features.

[0061] Non-keyframe features are obtained using the following formula. : .

[0062] S2125, non-critical features are fused with convolutional features to obtain the output feature map of the defect perception module.

[0063] S22, the fifth-scale feature map is upsampled and then concatenated with the fourth-scale feature map, and then input into a defect perception module to obtain the first intermediate feature. The first intermediate feature is concatenated with the third-scale feature map and then input into a defect perception module to obtain the first-scale fused feature.

[0064] S23, the first scale fusion feature and the first intermediate feature are concatenated and then input into a defect perception module to obtain the second scale fusion feature.

[0065] S24, the second-scale fused features are input into the CBS module and then concatenated with the fifth-scale feature map before being input into a defect perception module to obtain the third-scale fused features.

[0066] S25, the first-scale fusion feature, the second-scale fusion feature, and the third-scale fusion feature are enhanced by the adaptive enhancement dual-path detection module to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature; the first enhanced feature, the second enhanced feature, and the third enhanced feature are subjected to dual-branch detection to obtain the dual-branch detection results of the three scales; the dual-branch detection results of the three scales are merged to obtain the final LCD screen defect detection result.

[0067] In one embodiment, the adaptive dual-path detection module includes an SE attention mechanism.

[0068] In the feature extraction process of convolutional neural networks, different channels often correspond to different semantic responses and texture patterns. However, under complex backgrounds and imaging perturbations, the contributions of each channel to the defect region vary significantly. Traditional convolution fuses channel information in a fixed manner, which easily causes weak responses related to defects to be submerged by background noise, thus affecting the separability of small defects and the continuous representation of elongated defects. To improve the network's selective enhancement capability for key defect features, this embodiment introduces an SE attention mechanism into the detection network. By adaptively recalibrating the channel dimensions, it achieves "enhancing important channels and suppressing redundant channels," thereby improving the discriminativeness and robustness of feature representation. Figure 5 A flowchart of the SE attention mechanism provided in an embodiment of the present invention.

[0069] The processing steps based on the SE attention mechanism include: S251, global average pooling is performed on the input feature map to obtain the global description vector for each channel.

[0070] Let the input feature map be The SE module consists of two phases: "compression" and "excitation." First, the compression phase aggregates information in the spatial dimension using global average pooling to obtain a global description vector z∈ for each channel. It can be expressed by the following formula:

[0071] in, A global description vector for each channel. This is the input feature map.

[0072] This process statistically models the overall strength of the channel response at very low cost, providing a global context for the adaptive generation of subsequent channel weights.

[0073] S252, excite the global description vector to obtain the channel weight vector.

[0074] During the activation phase, a lightweight gating network is used to learn the nonlinear dependencies between channels and output the channel weight vector s∈ To balance expressive power and parameter efficiency, SE employs a bottleneck structure with a dimensionality reduction ratio r:

[0075] in , For learnable parameters, δ( ) represents ReLU activation, σ( ) represents the Sigmoid function to constrain the weights to [0,1].

[0076] S253, broadcasting the channel weight vector according to the spatial dimension to obtain recalibrated enhanced features.

[0077] Finally, the SE broadcasts the channel weights along the spatial dimension and multiplies them channel by channel with the input features to obtain the recalibrated enhanced features (i.e., the first enhanced feature, the second enhanced feature, and the third enhanced feature):

[0078] By integrating SE into the flexible LCD screen defect detection network, the channel contribution can be dynamically adjusted based on the current input global statistical information: when there is high light reflection and brightness gradient, SE tends to suppress channels that have a strong response to the background but low discrimination value; when a defect area appears, SE can enhance the channel response related to the edge, local texture change and defect morphology, thereby improving the signal-to-noise ratio and separability of the features, improving the classification confidence and boundary regression accuracy of the detection head for defects, and enhancing the robustness to bending deformation and imaging perturbation.

[0079] From a computational perspective, the additional parameters of SE mainly come from two fully connected layers, with a scale of approximately 2 / r, and are calculated only in the channel dimension, without significantly increasing the spatial computation of the feature map. Therefore, without significantly reducing inference speed, SE can effectively improve channel selection capability and is suitable as a lightweight attention enhancement module in detection networks.

[0080] Furthermore, in online defect detection of flexible LCD screens, point-like micro-defects often have an extremely low pixel ratio and are easily affected by reflections, high-brightness overflow, and uneven brightness caused by bending. When the main camera uses a large field of view to cover the entire screen, the details of small targets are further compressed, leading to missed detections and low confidence issues. To improve the observability and detection limit of small target defects, this embodiment proposes a dual-branch detection strategy: on the basis of the main camera completing global detection, an auxiliary camera is added to perform higher resolution monitoring of sensitive areas of small targets, forming a complementary system of "global coverage + local fine inspection".

[0081] The specific process is as follows: each enhanced feature is simultaneously fed into the main branch (global detection) and the auxiliary branch (local high-resolution detection) for parallel processing. The detection results output from the two branches are combined using a fusion strategy to make a comprehensive decision, thereby achieving fine detection of small target defects and stable identification of global defects.

[0082] The main branch first outputs the candidate defects and their locations for the entire screen. For small target candidates or suspicious areas with confidence levels close to the threshold or affected by specular interference, the auxiliary branch is called to verify and re-discriminate the corresponding areas. The results from the two paths are then fused for decision-making: if they are consistent, the confidence level is increased and the output is stabilized; if they are inconsistent, the auxiliary path corrects the clearer observation evidence of the small target or makes a decision according to preset rules.

[0083] This design improves the recall rate of minor defects and reduces false positives caused by reflective noise without sacrificing the overall cycle time, while also enhancing the system's robustness under flexible screen bending and complex imaging conditions.

[0084] In object detection, bounding box regression aims to make the predicted bounding box as consistent as possible with the ground truth bounding box in terms of position and shape. Traditional methods are based on... / The regression loss is scale-sensitive and fails to directly reflect the core objective of the detection task: "overlap degree." Using only IoU as the optimization objective also has limitations: gradient vanishes when the predicted bounding box does not overlap with the ground truth bounding box, and even when overlap exists, IoU struggles to simultaneously constrain geometric factors such as center position deviation and aspect ratio inconsistency. To obtain a more stable and geometrically consistent regression signal, this embodiment employs the CIoU loss function as the bounding box regression term. This allows for joint constraints on the predicted bounding box from three aspects: overlapping area, center distance, and shape consistency, thereby improving localization accuracy and convergence stability.

[0085] Let the prediction box be B The real frame is Its intersection-union ratio is defined as:

[0086] Building upon this, CIoU further introduces a center point distance penalty and an aspect ratio consistency penalty in addition to the IoU term. Its metric can be expressed as:

[0087] Among them, b and These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively, ρ( ) represents the Euclidean distance; c is the distance that can simultaneously enclose B and The diagonal length of the minimum bounding rectangle is used to normalize the center distance term, making the penalties comparable for targets of different scales. The shape consistency term v measures the difference in aspect ratio between the two frames and is often written as:

[0088] Among them, (w,h) and ( , ) represent the width and height of the predicted bounding box and the ground truth bounding box, respectively. The weighting coefficient α is used to adaptively balance the influence of the shape term and is generally defined as:

[0089] Therefore, the CIoU loss can be written as:

[0090] In the defect detection task of flexible LCD screens, defect morphology often includes both point-like micro-targets and slender line-like targets. Furthermore, due to bending deformation and reflective interference, bounding boxes are more prone to center shift and aspect ratio drift. CIoU, by explicitly penalizing the center distance, allows the predicted boxes to continue converging towards the true location even when the overlap is similar; simultaneously, it suppresses shape distortion through the aspect ratio term, which is particularly beneficial for the stable fitting of defects such as slender scratches and cracks. Compared to regression targets that rely solely on IoU, CIoU provides more comprehensive geometric constraints and a smoother optimization path during training, helping to improve the accuracy of bounding box regression and the overall detection performance of the model.

[0091] The experimental process for the YOLO-EDA model is described below: (1) Dataset To evaluate the effectiveness of the algorithm proposed in this invention in real-world production scenarios, it is necessary to construct a high-quality LCD screen defect dataset. However, currently, there is a lack of publicly available large-scale data resources in this field, making it difficult to directly meet research needs. Therefore, the data collection work in this embodiment is based on the real production line environment of a smart manufacturing enterprise in Jiangxi Province. Under the highly standardized production process of the enterprise, defective products account for a low proportion of the total output, resulting in a limited number of defect samples that can be obtained per day. To cover as many diverse defect morphologies and imaging conditions as possible, this embodiment collects all defect samples generated by the factory over 14 consecutive days and completes image acquisition and processing. The final data covers multiple typical defects. To facilitate subsequent statistical analysis and model training, this embodiment categorizes all samples into four main defect types. Figure 6 Examples of some typical defects provided for embodiments of the present invention.

[0092] This embodiment collected 2000 original defect samples and expanded the total number of images to 4307 through data augmentation strategies. To address the issue of uneven sample numbers across defect categories, and while ensuring full coverage of typical defect morphologies, the samples were further systematically screened and precisely labeled to construct the final experimental dataset for model training and evaluation. The dataset contains five key display defects: dot defects, line defects, scratches, leakage, and broken screens. The distribution of the number of samples in each category is shown in Table 1.

[0093] Table 1 Defect Types and Corresponding Number of Images

[0094] (2) Evaluation indicators To comprehensively evaluate the actual performance of the proposed algorithm in LCD screen defect detection, precision, recall, and mean precision (mAP) are used as the main performance evaluation metrics. The model's lightweight and real-time performance are also analyzed in conjunction with metrics such as model parameter count, computational complexity, and inference speed. The mAP metric is presented in two forms: mAP50 and mAP50-95. mAP50 evaluates the model's basic detection capability under a more lenient IoU threshold, reflecting the overall level of the model's ability to detect targets. mAP50-95, on the other hand, averages the results over a stricter IoU threshold range, emphasizing localization accuracy and bounding box regression quality under high overlap requirements, thus providing a more objective measure of the model's overall performance in fine-grained localization scenarios. The specific calculation methods for the evaluation metrics are shown below:

[0095] In the formula, TP represents the number of positive samples correctly detected, FP represents the number of negative samples incorrectly detected as positive samples, and FN represents the number of positive samples not detected. Therefore, precision (P) reflects the reliability of the prediction results; a higher value indicates fewer false positives. Recall (R) reflects the completeness of target detection; a higher value indicates fewer false negatives. In actual detection, P and R usually exhibit an inverse relationship as the confidence threshold changes. Therefore, P-R curves are typically plotted based on (P,R) combinations at different thresholds to more comprehensively describe the model's detection performance at different operating points.

[0096] Mean precision (AP) is defined as the area under the P-R curve, used to comprehensively reflect the overall precision performance of the model at different recall levels.

[0097] Furthermore, averaging the AP across all categories yields the mAP metric:

[0098] in Let be the average precision of the i-th class. This indicates the number of categories. A higher mAP value indicates better overall performance of the model in multi-category defect detection tasks. In this paper, mAP50 and mAP50-95 were calculated under the conditions of IoU=0.50 and IoU=0.50:0.95, respectively, to simultaneously consider the evaluation requirements of basic detection capability and high-precision positioning capability.

[0099] (3) Experimental platform and hyperparameter settings To verify the effectiveness and reproducibility of the algorithm proposed in this invention, a unified experimental platform was built in this embodiment, ensuring that all comparative and ablation experiments were conducted under the same hardware and software environment and training configuration. The experimental system used the Ubuntu 18.04 operating system, PyTorch as the deep learning framework, and YOLOv11 as the benchmark detection model. In terms of hardware, the experimental platform was equipped with an NVIDIA RTX 3090 GPU (24GB of VRAM), and the development language was Python 3.8.1. Specific hardware and software environment configurations are shown in Table 2.

[0100] Regarding training settings, to ensure a fair comparison between different methods, all experiments used the same hyperparameter configuration: input image size was set to 640×640, training epochs were 300, initial learning rate was 0.01, batch size was 24, and data loading threads were 8. To comprehensively evaluate the performance improvement of the YOLO-DEA algorithm, this paper conducts experimental verification from two aspects: firstly, comparative experiments with several mainstream object detection models are conducted to verify the overall performance advantage; secondly, module ablation experiments are used to analyze the contribution of each improved component to detection accuracy and robustness.

[0101] Table 2 Experimental Environment

[0102] (4) Comparative experiment To verify the comprehensive performance of the proposed YOLO-DEA model in the defect detection task of flexible LCD screens, this embodiment conducted a systematic comparative experiment with several mainstream object detection algorithms under a unified experimental environment and consistent training parameter settings. The experimental results are shown in Table 3. The comparison models cover typical two-stage detectors such as the Faster R-CNN series, as well as representative single-stage detectors such as SSD, YOLOv3, YOLOv5, and YOLOv8, to ensure the comprehensiveness and objectivity of the evaluation results. To ensure fairness, all comparison models were retrained under the same dataset and the same hyperparameter conditions, thereby eliminating the interference of different implementation methods and experimental environments on the results.

[0103] Experimental results show that YOLO-DEA achieves a better balance between detection accuracy and inference efficiency. Compared to two-stage methods such as Faster R-CNN, although they have certain advantages in feature representation, their computational complexity is high and their inference speed is slow, making it difficult to meet the real-time requirements of industrial online inspection. YOLO-DEA, by fusing DFM and AEDM structures and combining the CIoU loss function to optimize the regression process, effectively controls model complexity while enhancing feature representation capabilities, achieving higher detection efficiency. Compared to single-stage models such as SSD and YOLOv3, YOLO-DEA performs more stably in detecting defects in small targets and slender structures, achieving better results in both mAP50 and mAP50-95 metrics, especially with a more significant improvement in localization accuracy under high IoU conditions. Overall, the improved modules form a synergistic mechanism in feature modeling, detection enhancement, and regression optimization, enabling the model to significantly improve detection accuracy and robustness while ensuring real-time performance, providing an engineering solution that balances performance and efficiency for defect detection in flexible LCD screens.

[0104] Table 3 Comparative Experiments

[0105] Figure 7 This is a comparison of the visualization detection performance of different detection models provided in this invention on a liquid crystal screen defect dataset. The overall results show that traditional two-stage detectors, such as Faster R-CNN, have relatively accurate localization capabilities on detected targets. However, due to the deep network structure and insufficient feature response to small-scale targets, significant missed detections still occur in scenarios with small targets such as point defects. For single-stage detection models, although the detection speed is fast and can cover most defect types, when facing targets with large scale spans, such as line defects, the limited ability of feature representation to model complex deformations often leads to incomplete detection box localization or boundary regression offsets, especially when there is background reflection or uneven brightness.

[0106] In contrast, the YOLO-DEA model proposed in this invention effectively alleviates the aforementioned shortcomings through multi-module collaborative optimization. The DFF module enhances the dynamic fusion capability of features, enabling the network to achieve more sufficient information interaction between shallow details and deep semantics, thereby improving the detection stability of point defects; the DCNv4 module enhances the modeling capability of flexible deformation and non-rigid structures through deformable convolution, enabling the model to maintain high localization accuracy under curved surface bending and complex background conditions; the SE attention mechanism suppresses background interference and strengthens the response to key defects through channel recalibration, improving the feature representation of structural targets such as line defects and cracks; the dual-path detection strategy provides more refined observation information in small target scenes, further reducing the risk of missed detections. Experimental results show that the organic integration of multiple modules enables the model to significantly improve the overall localization accuracy and classification accuracy while maintaining a high recall rate, exhibiting stronger robustness and stability in complex industrial scenarios.

[0107] (4) Ablation test To further verify the detection performance of the algorithm proposed in this invention and analyze the actual contributions of each improved module, this embodiment evaluates the effectiveness and synergistic advantages of the three core innovative modules—the Defect Detection Module (DFM), the Adaptive Enhanced Dual-Path Detection Module (AEDM), and the CIoU loss function—through an ablation experiment system. The experiment used YOLOv11 as the baseline model, and training and testing were conducted under identical hyperparameter settings and hardware / software environments to ensure the fairness and comparability of the results. The experimental results are shown in Table 4, where “√” indicates that the module was introduced in the corresponding experimental group.

[0108] The experimental results show that the detection accuracy of the baseline model still has room for improvement. Introducing DFM into the backbone network significantly improves overall accuracy because the fusion of DFF and DCNv4 enhances the model's ability to represent features of flexible deformation and fine-grained defects. Further addition of the AEDM module effectively suppresses background interference and enhances the detection capability of multi-scale defects through the SE attention mechanism and dual-branch detection strategy, resulting in a more balanced performance across categories. Furthermore, optimizing the bounding box regression using the CIoU loss function further improves the model's localization accuracy under high IoU conditions. In summary, the three modules complement each other in feature modeling, detection enhancement, and regression optimization, exhibiting more significant performance gains when used together, fully validating the effectiveness and rationality of the proposed improvement scheme.

[0109] Table 4 YOLO-DEA Ablation Experiment

[0110] (5) Thermograph Experiment To investigate the feature extraction performance within the model, this embodiment uses heatmaps to explore the output of different models. Figure 8 The images show the heatmap experimental results of different models provided in this embodiment of the invention. Visual analysis of the gradient-weighted feature response results for different models yields an importance score for each pixel to the target category, generating corresponding heatmaps. The visualization results show significant differences in the feature interest regions among the different detection models. Traditional two-stage detection networks exhibit good response capabilities for detecting larger defects, but for small targets such as point defects, their feature response regions are scattered and have weak activation intensity, making it difficult to form stable discrimination regions, resulting in significant missed detections. While some single-stage detection models can respond to multiple defect types, under complex backgrounds or deformation conditions, their interest regions often deviate from the true defect edges, especially for slender structural targets such as line defects and cracks, where the heatmap activation range is incomplete, leading to a large deviation between the predicted and true bounding boxes.

[0111] In contrast, the proposed YOLO-DEA model exhibits a more concentrated and continuous response region in the heatmap. The model can form a clear distribution of high-response features in defect regions at different scales, especially in typical samples such as point defects, line defects, and cracks, where the feature activation regions highly overlap with the actual defect locations. This indicates that by introducing DFM to enhance deformation modeling capabilities, utilizing AEDM to strengthen multi-scale feature representation, and combining CIoU to optimize the regression process, the model achieves significant improvements in both spatial localization and semantic representation. For complex structural targets such as line defects and slender cracks, YOLO-DEA can fully characterize their shape and size features, making the regression box fit the actual boundary more closely, thereby effectively improving overall detection accuracy and localization stability.

[0112] Based on the method described in the above embodiments, this embodiment will further describe the LCD screen defect detection system based on the improved YOLO-DEA. The improved YOLO-DEA LCD screen defect detection system can be implemented as an independent entity or integrated into an electronic device. The electronic device can be a terminal, server, or other device. The terminal can include a tablet computer, a laptop computer, a personal computer (PC), a microprocessor box, or other devices.

[0113] Please see Figure 9 , Figure 9 This invention specifically describes a liquid crystal display defect detection system based on an improved YOLO-DEA, applicable to electronic devices. The improved YOLO-DEA-based liquid crystal display defect detection system may include: The image acquisition module is used to acquire the image to be detected; The LCD screen defect detection module is used to input the image to be detected into a trained improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

[0114] In specific implementation, the above modules and / or units can be implemented as independent entities, or they can be arbitrarily combined and implemented as the same or several entities. For the specific implementation of the above modules and / or units, please refer to the previous method embodiments. For the specific beneficial effects that can be achieved, please also refer to the beneficial effects in the previous method embodiments, which will not be repeated here.

[0115] In addition, this embodiment of the invention also provides an electronic device, which may be a computer, tablet computer, or other similar device. This electronic device can implement the steps of any embodiment of the improved YOLO-DEA-based liquid crystal screen defect detection method provided in this embodiment of the invention. Therefore, it can achieve the beneficial effects that any improved YOLO-DEA-based liquid crystal screen defect detection method provided in this embodiment of the invention can achieve, as detailed in the preceding embodiments, and will not be repeated here.

[0116] Figure 10 A specific structural block diagram of an electronic device provided in an embodiment of the present invention is shown. This electronic device can be used to implement the improved YOLO-DEA-based liquid crystal screen defect detection method provided in the above embodiments. The electronic device 500 can be a terminal, server, or other device. The terminal can include a tablet computer, laptop computer, personal computer (PC), microprocessor box, or other devices.

[0117] The memory 520 can be used to store software programs and modules, such as the program instructions / modules corresponding to those in the above embodiments. The processor 580 executes various functional applications and data processing by running the software programs and modules stored in the memory 520. The memory 520 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 520 may further include memory remotely located relative to the processor 580, and these remote memories can be connected to the electronic device 500 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0118] The input unit 530 can be used to receive input numeric or character information, and to generate a keyboard and mouse related to user settings and function control. Display unit 540 can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces, which can be composed of graphics, text, icons, video, and any combination thereof. Display unit 540 may include display panel 541, which may optionally be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar forms.

[0119] Electronic device 500, through transmission module 570 (e.g., Wi-Fi module), can help users receive requests, send information, etc., providing users with wireless broadband internet access. Although transmission module 570 is shown in the figure, it is understood that it is not an essential component of electronic device 500 and can be omitted as needed without changing the essence of the invention.

[0120] The processor 580 is the control center of the electronic device 500. It connects to various parts of the phone via various interfaces and lines, and performs various functions and processes data of the electronic device 500 by running or executing software programs and / or modules stored in the memory 520, and by calling data stored in the memory 520, thereby providing overall monitoring of the electronic device. Optionally, the processor 580 may include one or more processing cores; in some embodiments, the processor 580 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 580.

[0121] Electronic device 500 also includes a power supply 590 (such as a battery) that supplies power to various components. In some embodiments, the power supply may be logically connected to processor 580 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 590 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0122] Although not shown, the electronic device 500 also includes cameras (such as front-facing cameras and rear-facing cameras), Bluetooth modules, etc., which will not be described in detail here. Specifically, in this embodiment, the display unit of the electronic device is a touch screen display, and the mobile terminal also includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors. One or more programs contain instructions for performing the following operations: Acquire the image to be detected; The image to be detected is input into a trained, improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

[0123] In practice, the above modules can be implemented as independent entities or combined in any way to be implemented as the same or several entities. For the specific implementation of the above modules, please refer to the previous method implementation examples, which will not be repeated here.

[0124] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of the present invention provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the improved YOLO-DEA-based liquid crystal screen defect detection method provided by the present invention.

[0125] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0126] Since the instructions stored in the storage medium can execute the steps in any embodiment of the improved YOLO-DEA-based liquid crystal screen defect detection method provided in the embodiments of the present invention, the beneficial effects that any improved YOLO-DEA-based liquid crystal screen defect detection method provided in the embodiments of the present invention can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0127] The above provides a detailed description of a liquid crystal screen defect detection method, system, storage medium, and electronic device based on an improved YOLO-DEA, as provided in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A defect detection method for liquid crystal displays based on an improved YOLO-DEA, characterized in that, The method includes: Acquire the image to be detected; The image to be detected is input into a trained, improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

2. The liquid crystal display defect detection method based on the improved YOLO-DEA according to claim 1, characterized in that, The first processing layer of the backbone network includes a CBS module; the second, third, and fourth processing layers of the backbone network each include a CBS module and a defect perception module connected in sequence; the fifth processing layer of the backbone network includes a CBS module, a defect perception module, and a pyramid pooling and feature extraction module connected in sequence; the feature fusion network includes an upsampling module, a defect perception module, and a CBS module; the detection head includes an adaptive enhanced dual-path detection module. The processing steps of the improved YOLO-DEA model include: The input image to be detected is processed through five processing layers of the backbone network to obtain the first scale feature map, the second scale feature map, the third scale feature map, the fourth scale feature map and the fifth scale feature map respectively; The fifth-scale feature map is upsampled and then concatenated with the fourth-scale feature map before being input into a defect perception module to obtain the first intermediate feature. The first intermediate feature is then concatenated with the third-scale feature map before being input into a defect perception module to obtain the first-scale fusion feature. The first scale fusion feature is concatenated with the first intermediate feature and then input into a defect perception module to obtain the second scale fusion feature; The second-scale fusion feature is input into the CBS module and then stitched together with the fifth-scale feature map before being input into a defect perception module to obtain the third-scale fusion feature. The first-scale fusion feature, the second-scale fusion feature, and the third-scale fusion feature are respectively enhanced by an adaptive enhanced dual-path detection module to obtain a first enhanced feature, a second enhanced feature, and a third enhanced feature; the first enhanced feature, the second enhanced feature, and the third enhanced feature are subjected to dual-branch detection to obtain dual-branch detection results at three scales; the dual-branch detection results at three scales are merged to obtain the final LCD screen defect detection result.

3. The liquid crystal display defect detection method based on the improved YOLO-DEA according to claim 2, characterized in that, The defect perception module includes a deformable convolution module and a dynamic feature fusion module. The processing procedure of the defect perception module includes: By performing adaptive spatial sampling and geometric modeling on the input feature map through deformable convolution modules, convolutional features that are robust to deformation are obtained. The convolutional features are input into the dynamic feature fusion module, where feature enhancement and redundancy reduction are performed through keyframe feature reuse and temporal propagation mechanisms, resulting in the output feature map of the defect perception module.

4. The liquid crystal display defect detection method based on the improved YOLO-DEA according to claim 3, characterized in that, The processing steps of the deformable convolution module include: Initialize a learnable offset, and adaptively sample the input feature map using the learnable offset; The features of the sampling points are weighted and aggregated to obtain convolutional features.

5. The liquid crystal display defect detection method based on the improved YOLO-DEA according to claim 4, characterized in that, The processing steps of the dynamic feature fusion module include: The image to be detected is divided into keyframe images and non-keyframe images; The spatial position on the non-keyframe image is projected onto the corresponding position on the keyframe image using optical flow-guided spatial warp, and bilinear interpolation is performed to obtain approximate features. Scale field features are obtained by adjusting the projected features using the scale field. The scale field features and the approximate features are fused to obtain non-keyframe features; The non-critical features are fused with the convolutional features to obtain the output feature map of the defect perception module.

6. The liquid crystal display defect detection method based on the improved YOLO-DEA according to claim 5, characterized in that, set up p=(px,py) Non-keyframe images Spatial position on q For keyframe images At the sampling position on the image, p is back-projected onto the corresponding position in the keyframe image using a two-dimensional optical flow field. It can be expressed by the following formula: in, This represents the displacement field from frame i to frame k. Represents an optical estimation network; The approximate feature is obtained through the following formula. : Among them, G( ) is the bilinear interpolation kernel. Features of keyframe images; G( It can be broken down into: in, , for q The x and y coordinates, , for p The x and y coordinates; The projected features are adjusted using the following formula: in, For scale field characteristics, S( () is a scaling estimation network; Non-keyframe features are obtained using the following formula. : 。 7. The liquid crystal display defect detection method based on the improved YOLO-DEA according to claim 2, characterized in that, The adaptive dual-path detection module includes an SE attention mechanism, and the processing based on the SE attention mechanism includes: Global average pooling is performed on the input feature map to obtain the global description vector for each channel; The global description vector is excited to obtain the channel weight vector; The channel weight vector is broadcast according to the spatial dimension to obtain recalibrated enhanced features.

8. A liquid crystal display screen defect detection system based on an improved YOLO-DEA, characterized in that, include: The image acquisition module is used to acquire the image to be detected; The LCD screen defect detection module is used to input the image to be detected into a trained improved YOLO-DEA model to obtain the LCD screen defect detection result; wherein the improved YOLO-DEA model includes a backbone network, a feature fusion network, and a detection head, and the backbone network includes five processing layers; the processing procedure of the improved YOLO-DEA model includes: The image to be detected is input into the processing layer of the backbone network and processed sequentially to obtain a multi-scale feature map. The multi-scale feature map is then input into the feature fusion network to obtain a multi-scale fused feature. Finally, the multi-scale fused feature is input into the detection head to obtain the LCD screen defect detection result.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the liquid crystal display defect detection method based on the improved YOLO-DEA as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The device includes a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to execute the steps in the liquid crystal screen defect detection method based on the improved YOLO-DEA as described in any one of claims 1 to 7.

Images