LED pixel defect detection method based on convolutional neural network

Abstract

The invention discloses an LED pixel defect detection method based on a convolutional neural network, and the method comprises the following steps: obtaining an image, extracting convolution features, and generating a pixel feature vector set; selecting a node feature vector of the initial node to form an initial topological structure; updating the node feature vector, establishing a connection edge, resetting an initial aging value, and outputting a node updating result and a connection edge aging state; performing aging value progressive increase to generate a connection weakening area; calculating a node feature vector of a new node, establishing a connection edge, setting an initial aging value, generating a connection new region, and forming a weak connection sub-graph; error attenuation is executed, the topological structure is converged, and a stable topological structure is generated; judging a bright spot defect, a dark spot defect or a color cast defect of the pixel according to the stable topological structure and the weak connection subgraph; and recording pixel coordinates of the defect and outputting coordinate information. According to the method, the convolution features and the growth nerve gas are fused, accurate detection of LED pixel defects is realized, and the method has the advantage of high precision.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, and in particular to a method for detecting LED pixel defects based on convolutional neural networks. Background Technology

[0002] As the resolution of LED displays continues to improve, brightness consistency, color consistency, and pixel-level stability have become key factors affecting display quality. Existing LED pixel defect detection methods largely rely on traditional image processing algorithms, using threshold segmentation, local contrast analysis, or classification models based on convolutional neural networks to identify defects such as bright spots, dark spots, and color shifts. However, these methods typically only focus on the brightness variations or local texture features of the pixel itself, making it difficult to simultaneously model the topological relationships and structural evolution characteristics between pixels. This makes it challenging to achieve stable detection of weak brightness anomalies, subtle color shifts, and minor defects affected by neighboring interference.

[0003] On the other hand, while convolutional neural networks can extract deep, multi-dimensional features, their local convolutional mechanisms inherently lack sensitivity to pixel connectivity, topological stability, and structural abrupt changes, making it difficult to maintain high robustness in scenes with exceptionally sparse and irregularly distributed pixels. Existing methods generally lack a detection mechanism that can simultaneously utilize convolutional features and dynamic changes in neighborhood structure, failing to adaptively describe structural features such as edge aging, node error accumulation, and weakly connected regions in pixel groups. This results in insufficient accuracy in identifying brightness anomalies, color anomalies, and small defects.

[0004] Therefore, how to provide a method for detecting LED pixel defects based on convolutional neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an LED pixel defect detection method based on convolutional neural networks. This invention integrates convolutional features and growing neural gas to achieve accurate detection of LED pixel defects, and has the advantage of high precision.

[0006] A method for detecting LED pixel defects based on a convolutional neural network according to an embodiment of the present invention includes the following steps: The original image is acquired, input into a convolutional neural network, and the feature map is read to extract convolutional features. The convolutional features are combined with channel brightness, brightness gradient, chromaticity difference and neighborhood consistency in a fixed order to generate a set of pixel feature vectors. Select the node feature vectors of the initial nodes from the set of pixel feature vectors to form the initial topology; Input pixel feature vectors, determine the winning node and neighboring nodes, update node feature vectors, establish connections and reset initial aging values, and output node update results and connection aging status. The aging value is incremented, and edges that reach the threshold are deleted, creating weakened connection regions. Select the node with the largest node error, calculate the node feature vector, establish the connection and set the initial aging value, generate the new connection region and combine it with the weakened connection region to form a weak connection subgraph. Error decay is performed on node errors, and the topology converges based on node feature vectors and edge relationships to generate a stable topology. Based on the stable topology and weakly connected subgraph, the bright spot defects, dark spot defects, or color shift defects of pixels are determined by node error, isolated nodes, growth density changes, and node indices in the weakly connected subgraph. Record the coordinates of pixels that are identified as bright spot defects, dark spot defects, or color cast defects, and output the coordinate information.

[0007] Optionally, the generation of the pixel feature vector set includes: The original image of the LED display screen is obtained and input into a pre-trained convolutional neural network. Feature maps are read from the intermediate layers of the convolutional neural network. The feature vectors corresponding to each pixel position in the image are extracted from the feature maps as convolutional features and stored as a convolutional feature set. Read the red channel brightness value, green channel brightness value and blue channel brightness value of each pixel in the original image, and calculate the brightness gradient; The chromaticity difference is calculated based on the luminance values ​​of the red channel, green channel, and blue channel. Neighborhood consistency is calculated based on a fixed neighborhood window; The convolutional features corresponding to the current pixel in the convolutional feature set are combined with the red channel brightness value, green channel brightness value, blue channel brightness value, brightness gradient, chromaticity difference and neighborhood consistency in a fixed order to form a pixel feature vector.

[0008] Optionally, the formation of the initial topology includes: Two sets of pixel feature vectors are selected from the set of pixel feature vectors in a fixed index order, and the selected pixel feature vectors are used as the node feature vectors of the initial node. Set an initial node error value for each initial node, establish initial connections between initial nodes, and set an initial aging value for each initial connection. Set the node insertion interval and the number of iterations; The initial topology is formed by combining the initial nodes, node feature vectors, initial node error values, initial edges, initial aging values, node insertion intervals, and iteration counts.

[0009] Optionally, the generation of the node update result and the edge aging state includes: Input pixel feature vectors sequentially, and calculate the distance between the input pixel feature vector and the feature vectors of all nodes; The winning node is determined based on the principle of minimum distance, and the neighboring nodes are determined based on the edge relationships of the winning node in the topology. The node feature vector of the winning node is updated based on the brightness deviation between the input pixel feature vector and the winning node; Update the node feature vectors of neighboring nodes based on the neighborhood consistency between the input pixel feature vector and neighboring nodes; The node index of the winning node and the node index corresponding to the input pixel feature vector are combined into a pair of index groups. This index group is added to the edge set, and an aging value is set for the edge. The aging value is then reset to the initial aging value. Output the update results of the winning node, the update results of the neighboring nodes, and the aging status of the edges.

[0010] Optionally, the generation of the connection weakening region includes: The set of edges is traversed. For each edge traversed, the current aging value of the edge is read. The current aging value is added to a preset fixed increment. The aging value is then added to the fixed increment, and the result of the addition operation is used to overwrite the aging value of the edge to form an updated aging value. The updated aging value is compared with the set aging threshold. If the updated aging value is greater than or equal to the set aging threshold, the edge is deleted from the edge set. When the edge is deleted, the node indices of the two nodes corresponding to the edge are recorded and the node index pair is used as part of the connection weakening region. After completing the aging value increment and edge deletion for all connections, the node index pairs obtained from the edge deletion will form a weakened connection region.

[0011] Optionally, the generation of the weak connection subgraph includes: Read the node error of all nodes in the current iteration, add the node error of each node to the node error of the previous iteration, and use the result of the addition operation as the dynamic error value of the node. Perform a comparison operation on all dynamic error values, determine the node with the largest dynamic error value as the target node, and use the index of this node as the input index for generating new nodes and inserting subsequent edges; Read the node feature vector of the nearest node of the target node, perform addition operations on the node feature vector of the target node and the node feature vector of the nearest node respectively, take the average value, and use the average value as the node feature vector of the new node. Insert the new node into the growing neural gas network, establish an edge between the new node and the target node, set the aging value of the edge to the initial aging value, and use the edge formed between the new node and the target node as a component of the connection of the new growth region. The newly formed regions and the weakened regions are combined and merged into a group of regions, and the merged region is used as a weak connection subgraph.

[0012] Optionally, the generation of the stable topology includes: Read the node error of all nodes in the current iteration, multiply the node error of each node with the preset decay coefficient, and use the result of the multiplication operation as the updated node error of that node. Read the node feature vectors and edge sets of all nodes. For each edge connecting two nodes, read their node feature vectors. Perform addition on the two node feature vectors respectively, take the average value, and use the average value as the update reference value for the two nodes. The above updated reference values ​​are added to the corresponding node feature vectors in turn and divided by two, and the result is used as the iterative update value of the node feature vectors. After all node feature vectors have been updated, the updated set of node feature vectors and the updated set of edges are merged into a set of structures, and this set of structures is used as a temporary topological structure for topological convergence. Multiple rounds of iterative updates are performed on the temporary topology. The change in the node feature vector after each round of iteration is compared with a preset change threshold. The iteration result with a change less than the preset change threshold is taken as the stable topology.

[0013] Optionally, the determination of bright spot defect pixels, dark spot defect pixels, or color shift defect pixels includes: Read the node errors of all nodes in the stable topology, compare the node errors with a preset error threshold, mark the nodes whose node errors are greater than the preset error threshold as error abnormal nodes, and use the error abnormal nodes as the initial content of the abnormal set. Read the set of edges in the stable topology, filter the nodes with zero edges, mark the nodes with zero edges as isolated nodes, and add the isolated nodes to the anomaly set; Read the number of node updates in the topology during the iteration process, compare the number of node updates with a preset growth density threshold, mark the nodes with an update count greater than the preset growth density threshold as abnormal growth density nodes, and add the abnormal growth density nodes to the anomaly set. Read the node index in the weak connection subgraph, mark the node corresponding to the node index in the weak connection subgraph as a weak connection exception node, and add the weak connection exception node to the exception set; Perform a correspondence matching between the node feature vectors of all nodes in the anomaly set and the node feature vectors of the stable topology. Classify the nodes according to their corresponding pixel positions and mark the corresponding pixels as bright spot defect pixels, dark spot defect pixels, or color shift defect pixels.

[0014] Optionally, the output of the final defect coordinate information includes: Read the node indices of pixels identified as bright spot defects, dark spot defects, and color cast defects, match the node indices with the pixel positions in the original image, and find and record the horizontal and vertical positions of the corresponding pixels as pixel coordinates based on the node indices. The recorded pixel coordinates are categorized into bright spot defect pixels, dark spot defect pixels, and color cast defect pixels, forming bright spot defect coordinate sets, dark spot defect coordinate sets, and color cast defect coordinate sets, respectively. Perform an integration operation on the set of bright spot defect coordinates, the set of dark spot defect coordinates, and the set of color cast defect coordinates, merge the three sets of coordinates into one output list, and output this output list as the final defect coordinate information.

[0015] The beneficial effects of this invention are: This invention combines the deep representation capabilities of convolutional neural networks (CNNs) with the topological self-organizing mechanism of growing neural gas to form a highly sensitive, stable, and interpretable detection process for pixel-level LED defect detection. CNNs can extract multi-scale, multi-channel deep features from the original image, which, together with traditional RGB brightness values, brightness gradients, chromaticity differences, and neighborhood consistency, constitute a rich pixel feature vector, providing a more discriminative input foundation for subsequent topological structure construction. Through this fusion of convolutional features and manually designed features, this invention effectively overcomes the limitations of traditional thresholding methods, such as insufficient feature dimensions and inability to express complex pixel distribution relationships, achieving full capture of subtle brightness changes and nonlinear color shift features.

[0016] During the topology construction process, this invention utilizes the adaptive node update, edge aging, node insertion, and error accumulation mechanisms of the growing neural gas to gradually reveal the potential structural relationships between pixels in the topological network. Pixels with abnormal brightness or color, due to their inherent differences from normal pixels in the feature space, exhibit significant topological behaviors during iteration, such as accelerated error accumulation, accelerated edge aging, and the deletion or reconnection of connections. This results in the formation of weakened connection regions, newly formed connection regions, and weakly connected subgraphs in the network structure. This structural abrupt change becomes a key basis for identifying minor defects such as weak bright spots, weak dark spots, and color shifts, enabling this invention to effectively identify pixel sparsity anomalies that are difficult for traditional convolutional neural networks to handle individually.

[0017] Furthermore, this invention addresses the challenge of simultaneously identifying pixel anomalies from both local feature and structural evolution perspectives by jointly analyzing weakly connected subgraphs, growth density variations, node error distribution, and isolated node states. This enables refined detection of pixel-level defects. The generation of a stable topology ensures the algorithm's high robustness to noise disturbances, maintaining stable output even under conditions of uneven illumination, brightness fluctuations, or high image noise. Ultimately, this invention can identify bright spot defects, dark spot defects, and color shift defects with a lower false positive rate in high-resolution LED display inspection, significantly improving detection accuracy, reliability, and generalization ability. This provides technical support for automated production line inspection, high-precision calibration, and long-term stable monitoring. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an LED pixel defect detection method based on convolutional neural network and topological evolution of growing neural gas proposed in this invention; Figure 2 This is a schematic diagram illustrating the determination of bright spot defects, dark spot defects, and color shift defects based on weak connection subgraphs, node errors, isolated nodes, and growth density changes in an LED pixel defect detection method based on convolutional neural networks proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A method for detecting LED pixel defects based on convolutional neural networks includes the following steps: The original image of the LED display screen is obtained and input into a pre-trained convolutional neural network. The feature map output by the intermediate layer of the convolutional neural network is read. The convolutional features corresponding to each pixel position in the image are extracted from the feature map. The convolutional features of each pixel are combined with the red channel brightness, green channel brightness, blue channel brightness, brightness gradient, chromaticity difference and neighborhood consistency calculated according to a fixed neighborhood window in a fixed order to generate a set of pixel feature vectors containing convolutional features. Two sets of pixel feature vectors are selected from the set of pixel feature vectors as the node feature vectors of the initial nodes. The initial node error value is set, the initial connection is established and the initial aging value is set. The node insertion interval and the number of iterations are set to form the initial topology. Input pixel feature vectors sequentially, calculate the distance between the input feature vector and the node feature vector to determine the winning node and neighboring nodes, update the node feature vector of the winning node based on the brightness deviation, update the node feature vector of the neighboring nodes based on the neighborhood consistency, establish a connection between the winning node and the pixel input feature vector and reset the aging value of the connection to the initial aging value, and output the node update result and the connection aging status. The aging value is incremented for all edges, and edges whose aging value reaches the set aging threshold are deleted to generate weakened connection regions. Select the node with the largest cumulative node error based on the node error accumulation, calculate the node feature vector of the new node based on the node feature vector of the node and its nearest node, insert the new node and establish the connection, set the initial aging value, generate the new connection region, and combine the new connection region with the weakened connection region to form a weak connection subgraph. Error decay is performed on the node errors of all nodes, and the topology converges based on the node feature vectors and the edge relationships to generate a stable topology. Based on the stable topology and weakly connected subgraph, the bright spot defects, dark spot defects, or color shift defects of pixels are determined by node error, isolated nodes, growth density changes, and node indices in the weakly connected subgraph. Record the coordinates of pixels that are identified as bright spot defects, dark spot defects, or color cast defects, and output the coordinate information.

[0021] In this embodiment, the generation of the pixel feature vector set includes: The original image of the LED display screen is obtained and input into a pre-trained convolutional neural network. Feature maps are read from the intermediate layers of the convolutional neural network. The feature vectors corresponding to each pixel position in the image are extracted from the feature maps as convolutional features and stored as a convolutional feature set. For each pixel in the original image, read the red channel brightness value, green channel brightness value, and blue channel brightness value, and calculate the brightness gradient. The brightness gradient is obtained by taking the square root of the sum of the squares of the differences between the red channel brightness value of the current pixel and the red channel brightness value of the neighboring pixel, the squares of the differences between the green channel brightness value of the current pixel and the green channel brightness value of the neighboring pixel, and the squares of the differences between the blue channel brightness value of the current pixel and the blue channel brightness value of the neighboring pixel. The chromaticity difference is calculated based on the luminance values ​​of the red channel, green channel, and blue channel. The chromaticity difference is calculated by taking the square root of the sum of the squares of the differences between the red channel and green channel luminance values, the squares of the differences between the red channel and blue channel luminance values, and the squares of the differences between the green channel and blue channel luminance values. Neighborhood consistency is calculated based on a fixed neighborhood window. The neighborhood consistency is calculated by subtracting the red, green, and blue channel brightness values ​​of the corresponding pixels in the neighborhood from the red, green, and blue channel brightness values ​​of the current pixel, respectively, summing the squares of the three brightness differences, and then averaging the summation results of all pixels in the neighborhood. The average value is then used as the neighborhood consistency value of the pixel. The convolutional features corresponding to the current pixel in the convolutional feature set are combined with the red channel brightness value, green channel brightness value, blue channel brightness value, brightness gradient, chromaticity difference and neighborhood consistency in a fixed order to form a pixel feature vector.

[0022] In this embodiment, the formation of the initial topology includes: Two sets of pixel feature vectors are selected from the set of pixel feature vectors in a fixed index order. The selected pixel feature vectors are used as the node feature vectors of the initial nodes for subsequent distance calculation, node update and edge establishment. An initial node error value is set for each initial node. The initial node error value is set by assigning a fixed value to each initial node and storing the fixed value as the error variable of that node. An initial edge is established between the initial nodes. The edge is formed by forming a pair of node index groups by combining the node indices of any two initial nodes and adding the node index groups to the edge set. An initial aging value is set for each initial edge. The initial aging value is set to a fixed initial value. Set a node insertion interval, which is a preset fixed iteration interval value written into a node insertion control variable, so that subsequent iterations trigger node insertion when the iteration interval is reached. Set the number of iterations, which is a preset fixed total number of iterations written into an iteration count control variable, so that each iteration in the topology evolution process increases the count by one round after completing one node update and edge update, until the fixed total number of iterations is reached. The initial topology is formed by combining the initial nodes, node feature vectors, initial node error values, initial edges, initial aging values, node insertion intervals, and iteration counts. Specifically, the formation of the topology involves combining all the above variables into a data structure, storing the node index, node feature vector array, node error array, edge index group, edge aging array, node insertion control variable, and iteration count control variable in a fixed order, and using this structure as the entire content of the initial topology.

[0023] In this embodiment, the generation of the node update result and the edge aging state includes: Input the pixel feature vectors sequentially, and use each pixel feature vector as the input for distance calculation and node feature vector update. Calculate the distance between the input pixel feature vector and all node feature vectors. The distance is calculated by taking the square root of the sum of the squares of the brightness difference of the red channel, the green channel, the blue channel, the brightness gradient difference, the chromaticity difference, and the neighborhood consistency difference. The winning node is determined based on the principle of minimum distance, and the neighboring nodes are determined based on the edge relationships of the winning node in the topology. The node feature vector of the winning node is updated based on the brightness deviation between the input pixel feature vector and the winning node. The generation of the node feature vector of the winning node specifically includes subtracting the corresponding brightness values ​​in the node feature vector of the winning node from the brightness values ​​of the red channel, green channel and blue channel in the input pixel feature vector, multiplying the three brightness differences by a preset update coefficient, and then adding the multiplication result to the value of the corresponding dimension of the node feature vector of the winning node. The node feature vector of the neighboring node is updated based on the neighborhood consistency between the input pixel feature vector and the neighboring node. The generation of the updated node feature vector of the neighboring node specifically includes subtracting the neighborhood consistency value in the neighboring node feature vector from the neighborhood consistency value in the input feature vector, multiplying the difference by a preset neighborhood consistency update coefficient, and then adding the multiplication result to the neighborhood consistency dimension value in the neighboring node feature vector. The node index of the winning node and the node index corresponding to the input pixel feature vector are combined into a pair of index groups. This index group is added to the edge set, and an aging value is set for the edge. The aging value is then reset to the initial aging value. Output the update results of the winning node, the update results of the neighboring nodes, and the aging status of the edges.

[0024] In this embodiment, the generation of the connection weakening region includes: The set of edges is traversed. For each edge traversed, the current aging value of the edge is read. The current aging value is added to a preset fixed increment. The aging value is then added to the fixed increment, and the result of the addition operation is used to overwrite the aging value of the edge to form an updated aging value. The updated aging value is compared with the set aging threshold. If the updated aging value is greater than or equal to the set aging threshold, the edge is deleted from the edge set. When the edge is deleted, the node indices of the two nodes corresponding to the edge are recorded and the node index pair is used as part of the connection weakening region. After completing the aging value increment and edge deletion for all connections, the node index pairs obtained from the edge deletion will form a weakened connection region.

[0025] In this embodiment, the generation of the weak connection subgraph includes: Read the node error of all nodes in the current iteration, perform an addition operation on the node error of each node and its node error in the previous iteration, and use the result of the addition operation as the dynamic error value of the node. The addition operation is performed by reading the node error value of the current iteration and the node error value of the previous iteration from the node error array, adding the two, and writing the addition result to the corresponding index position of the dynamic error array to overwrite the original dynamic error value, so that the dynamic error array becomes the only storage of the cumulative node error value. A comparison operation is performed on all dynamic error values, and the node with the largest dynamic error value is determined as the target node. The index of this node is used as the input index for generating new nodes and inserting subsequent edges. The comparison operation is implemented by traversing the dynamic error array and recording the array index of the item with the largest value. Read the node feature vector of the nearest node of the target node. The nearest node is obtained by reading the index of the adjacent node of the target node in the edge set. Perform addition operation on the node feature vector of the target node and the node feature vector of the nearest node respectively, take the average value, and use the average value as the node feature vector of the new node. Insert the new node into the growing neural gas network, establish an edge between the new node and the target node, set the aging value of the edge to the initial aging value, and use the edge formed between the new node and the target node as a component of the connection of the new growth region. The process involves combining newly formed and weakened connection regions into a single region, and using this combined region as a weak connection subgraph. Specifically, the generation of the weak connection subgraph includes reading both the weakened and newly formed connection regions, merging the two types of node index pairs into a new set of node index pairs in a fixed order, and using this set as the weak connection subgraph.

[0026] In this embodiment, the generation of the stable topology includes: Read the node errors of all nodes in the current iteration, perform a multiplication operation on the node error of each node with a preset decay coefficient, the multiplication operation uses the node error as the multiplicand and the decay coefficient as the multiplier, and use the result of the multiplication operation as the updated node error of that node. Read the node feature vectors and edge sets of all nodes. For each edge connecting two nodes, read their node feature vectors. Perform addition on the two node feature vectors respectively, take the average value, and use the average value as the update reference value for the two nodes. The above updated reference values ​​are added to the corresponding node feature vectors in turn and divided by two, and the result is used as the iterative update value of the node feature vectors. After all node feature vectors have been updated, the updated set of node feature vectors and the updated set of edges are merged into a set of structures, and this set of structures is used as a temporary topology structure for topology convergence. The generation of the temporary topology structure includes creating a structure, the structure fields of which include a node feature vector array, an edge index array, and an edge aging array. The above values ​​are written into the corresponding fields of the structure, and the structure is stored as a temporary topology structure. Multiple rounds of iterative updates are performed on the temporary topology. The change in the node feature vector after each round of iteration is compared with a preset change threshold. The iteration result with a change less than the preset change threshold is taken as the stable topology.

[0027] In this embodiment, the determination of bright spot defect pixels, dark spot defect pixels, or color cast defect pixels includes: Read the node errors of all nodes in the stable topology, compare the node errors with a preset error threshold, mark the nodes whose node errors are greater than the preset error threshold as error abnormal nodes, and use the error abnormal nodes as the initial content of the abnormal set. Read the set of edges in the stable topology, filter the nodes with zero edges, mark the nodes with zero edges as isolated nodes, and add the isolated nodes to the anomaly set; Read the number of node updates in the topology during the iteration process, compare the number of node updates with a preset growth density threshold, mark the nodes with an update count greater than the preset growth density threshold as abnormal growth density nodes, and add the abnormal growth density nodes to the anomaly set. Read the node index in the weak connection subgraph, mark the node corresponding to the node index in the weak connection subgraph as a weak connection exception node, and add the weak connection exception node to the exception set; Perform a correspondence matching between the node feature vectors of all nodes in the anomaly set and the node feature vectors of the stable topology. Classify the nodes according to their corresponding pixel positions and mark the corresponding pixels as bright spot defect pixels, dark spot defect pixels, or color shift defect pixels.

[0028] In this embodiment, the output of the final defect coordinate information includes: Read the node indices of pixels identified as bright spot defects, dark spot defects, and color cast defects, match the node indices with the pixel positions in the original image, and find and record the horizontal and vertical positions of the corresponding pixels as pixel coordinates based on the node indices. The recorded pixel coordinates are categorized into bright spot defect pixels, dark spot defect pixels, and color cast defect pixels, forming bright spot defect coordinate sets, dark spot defect coordinate sets, and color cast defect coordinate sets, respectively. Perform an integration operation on the set of bright spot defect coordinates, the set of dark spot defect coordinates, and the set of color cast defect coordinates, merge the three sets of coordinates into one output list, and output this output list as the final defect coordinate information.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the routine maintenance and inspection of a high-definition LED display screen in a large indoor broadcasting center. The main background wall of the broadcasting center is composed of multiple high-brightness LED displays, resulting in a high overall resolution. However, after prolonged continuous use, minor defects such as abnormal brightness points, dim pixels, and color shift areas irregularly appear on the screens. These pixel anomalies negatively impact the visual experience of live broadcasts, but due to their sparse distribution and very small area, coupled with the complex lighting conditions inside the broadcasting hall, they are easily missed during manual inspections. The detection method proposed in this invention is clearly applicable in this scenario, effectively solving the problem that traditional detection methods cannot accurately identify subtle bright spots, weak dark spots, and slight color shifts.

[0030] In practical applications, staff first use professional monitoring cameras installed in the central control console of the studio to acquire raw images of the entire LED display screen. Because the displayed content changes in real time, testing is conducted while the background image maintains a static test pattern to ensure that the pixel brightness and chromaticity relationships in the input image are stable. The acquired images are input into the system, which first extracts deep convolutional features for each pixel location using a convolutional neural network. This convolutional neural network, derived from a pre-trained visual model, is used to capture texture features, local structural features, and color distribution patterns at different spatial scales in the displayed image. Subsequently, the system further calculates the red, green, and blue brightness, brightness gradient, chromaticity difference, and neighborhood consistency obtained based on a fixed neighborhood window for each pixel, combining the convolutional features with multi-dimensional artificial features into a more complete pixel representation.

[0031] After completing pixel feature construction, the system initializes the growing neural gas topology according to the method described in the claims. The initial node features are derived from the multidimensional feature vectors of a subset of pixels, and these serve as the starting point for constructing initial connections and initial aging values. As a large number of pixel features from the image are sequentially input into the system, the topology network begins to dynamically evolve based on the differences between pixel features. Each time a new pixel feature vector is input, the system first finds the winning node closest to that feature and simultaneously determines its connected neighboring nodes. The winning node updates its features based on brightness deviation, neighboring nodes adjust based on neighborhood consistency, and a new connection is formed between the winning node and the input pixel, resetting the aging value of that connection.

[0032] As the input process continues, the aging values ​​of some edges in the topology network accumulate. Once they reach a set threshold, they are deleted, and this deletion gradually forms weakened connection regions in the network structure. Simultaneously, because some nodes maintain high errors over multiple update rounds, they are selected by the system as node insertion targets. New edges are formed between the inserted nodes and the target nodes, creating newly formed connection regions around them. The weakened connection regions and newly formed connection regions merge to form a weak connection subgraph. This structure is crucial in topology evolution, reflecting the locations of abrupt changes in feature distribution in the image, often corresponding to bright spots, dark spots, or slightly color-shifted pixels in the real image.

[0033] In actual testing in the studio, the system constructs a stable topology through multiple iterations. The convergence process of the stable topology reflects the gradual convergence of structural relationships among most normal pixels, while abnormal pixels exhibit significant deviations from the structure. For example, pixels with abnormal brightness may cause a continuous increase in node error values ​​due to excessive feature differences; isolated pixels no longer maintain connections with other nodes in the structure due to the extensive deletion of edges; in sparsely distributed color-shifted regions, the number of nodes covered by weakly connected subgraphs is often small, but structural abrupt changes are obvious; changes in growth density also manifest as an abnormally high number of local updates as the number of heterogeneous pixels increases. The system uniformly analyzes these structural anomalies and locates the relevant nodes to their corresponding pixel coordinates, thereby determining whether they are bright spots, dark spots, or color-shifted defects.

[0034] Throughout the implementation process, the inspection was conducted at night to avoid lighting interference in the studio and ensure image consistency. The system can complete pixel-level topology analysis of the entire screen in a short time and determine the precise location of abnormal pixels. Staff can then repair or replace defective pixels based on the coordinates output by the system. Through multiple comparative inspections, it was found that the method of this invention has significant advantages in identifying micro-pixel anomalies that are difficult to detect by traditional methods, such as weak bright spots, slight dark spots, and mild color shifts. This ensures the high brightness and high consistency of the image required in the broadcast environment, further improving the overall display quality and operational stability.

[0035] This invention demonstrates that, in high-resolution LED display environments, a detection method combining convolutional features with the topological evolution of growing neural gas can effectively enhance the system's ability to identify minute pixel anomalies. Mechanisms such as edge aging, node error accumulation, and the formation of weakly connected subgraphs within the topological structure enable the system to automatically capture structural mutations in a complex feature space, thereby achieving higher reliability and precision than traditional methods. This method is not only applicable to broadcasting centers but also to the maintenance of various types of LED display devices, including outdoor advertising screens, conference displays, and exhibition screens, demonstrating broad practical application value.

[0036] Table 1 Performance Comparison of LED Pixel Defect Detection Methods As shown in Table 1, the method of this invention significantly outperforms the traditional thresholding method and the simple CNN classification method in all detection metrics. In the bright spot detection task, the accuracy of the traditional thresholding method is only 85.3%, the CNN method improves to 88.6%, and the method of this invention, which combines CNN with the topological evolution of growing neural gas, further improves it to 92.8%. This improvement is not a minor optimization, but reflects the significant advantage of the topological growth model in handling abrupt changes in brightness structure. In this invention, edge aging, node error accumulation, and the formation of weakly connected subgraphs can sensitively respond to brightness deviations, thus performing better in handling weak bright spots.

[0037] In the task of dark spot detection, the accuracy of the traditional thresholding method is 81.5%, reflecting its difficulty in distinguishing weak dark spots from low-brightness normal pixels. CNNs, due to their ability to extract local deep features, improve the accuracy to 86.4%. However, the method of this invention, through dynamic tracking of the cumulative value of node errors and continuous monitoring of the consistency of neighborhood features, more clearly represents the "abnormal shrinkage behavior" of dark spots in the topological structure, thereby significantly improving the dark spot detection accuracy to 91.3%, a difference of nearly 10 percentage points compared to the traditional thresholding method.

[0038] In color cast detection, the traditional thresholding method achieves an accuracy of 78.1%, the CNN method 84.2%, while the method of this invention reaches 91.7%. Color cast defects typically do not exhibit obvious local brightness jumps like bright or dark spots, but rather present as slight color shifts, making them difficult to detect using brightness thresholding methods. This invention utilizes convolutional features to enhance color representation capabilities and leverages a growing neural gas network to dynamically track structural changes between nodes, resulting in more differentiated distributions of color-cast pixels in weakly connected subgraphs. Therefore, it can identify even slight color casts that are difficult for CNNs to handle.

[0039] In terms of overall detection accuracy, the accuracy has improved from 82.6% with the traditional thresholding method to 85.7% with CNN, and then to 92.3% with this invention. Each stage reflects an improvement in technical capabilities, but the improvement of this invention is the most significant, with an overall accuracy 9.7 percentage points higher than the thresholding method and 6.6 percentage points higher than using CNN alone. The performance advantages of this invention mainly come from three aspects: first, convolutional features enhance pixel representation capabilities; second, the growing neural gas network can automatically capture abrupt changes in topology; and third, weak connection subgraphs can accurately characterize structural anomalies in fine-grained defect regions.

[0040] In summary, the method of this invention achieves an accuracy rate exceeding 90% in detecting three types of defects: bright spots, dark spots, and color shifts, significantly outperforming traditional methods. These data fully demonstrate the technical advantages of combining convolutional features with the topological evolution of growing neural gas, exhibiting strong sensitivity, strong stability, and high precision in detecting pixel-level weak anomalies.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting LED pixel defects based on a convolutional neural network, characterized in that, Includes the following steps: The original image is acquired, input into a convolutional neural network, and the feature map is read to extract convolutional features. The convolutional features are combined with channel brightness, brightness gradient, chromaticity difference and neighborhood consistency in a fixed order to generate a set of pixel feature vectors. Select the node feature vectors of the initial nodes from the set of pixel feature vectors to form the initial topology; Input pixel feature vectors, determine the winning node and neighboring nodes, update node feature vectors, establish connections and reset initial aging values, and output node update results and connection aging status. The aging value is incremented, and edges that reach the threshold are deleted, creating weakened connection regions. Select the node with the largest node error, calculate the node feature vector, establish the connection and set the initial aging value, generate the new connection region and combine it with the weakened connection region to form a weak connection subgraph. Error decay is performed on node errors, and the topology converges based on node feature vectors and edge relationships to generate a stable topology. Based on the stable topology and weakly connected subgraph, the bright spot defects, dark spot defects, or color shift defects of pixels are determined by node error, isolated nodes, growth density changes, and node indices in the weakly connected subgraph. Record the coordinates of pixels that are identified as bright spot defects, dark spot defects, or color cast defects, and output the coordinate information.

2. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The generation of the pixel feature vector set includes: The original image of the LED display screen is obtained and input into a pre-trained convolutional neural network. Feature maps are read from the intermediate layers of the convolutional neural network. The feature vectors corresponding to each pixel position in the image are extracted from the feature maps as convolutional features and stored as a convolutional feature set. Read the red channel brightness value, green channel brightness value and blue channel brightness value of each pixel in the original image, and calculate the brightness gradient; The chromaticity difference is calculated based on the luminance values ​​of the red channel, green channel, and blue channel. Neighborhood consistency is calculated based on a fixed neighborhood window; The convolutional features corresponding to the current pixel in the convolutional feature set are combined with the red channel brightness value, green channel brightness value, blue channel brightness value, brightness gradient, chromaticity difference and neighborhood consistency in a fixed order to form a pixel feature vector.

3. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The formation of the initial topology includes: Two sets of pixel feature vectors are selected from the set of pixel feature vectors in a fixed index order, and the selected pixel feature vectors are used as the node feature vectors of the initial node. Set an initial node error value for each initial node, establish initial connections between initial nodes, and set an initial aging value for each initial connection. Set the node insertion interval and the number of iterations; The initial topology is formed by combining the initial nodes, node feature vectors, initial node error values, initial edges, initial aging values, node insertion intervals, and iteration counts.

4. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The generation of the node update results and the edge aging status includes: Input pixel feature vectors sequentially, and calculate the distance between the input pixel feature vector and the feature vectors of all nodes; The winning node is determined based on the principle of minimum distance, and the neighboring nodes are determined based on the edge relationships of the winning node in the topology. The node feature vector of the winning node is updated based on the brightness deviation between the input pixel feature vector and the winning node; Update the node feature vectors of neighboring nodes based on the neighborhood consistency between the input pixel feature vector and neighboring nodes; The node index of the winning node and the node index corresponding to the input pixel feature vector are combined into a pair of index groups. This index group is added to the edge set, and an aging value is set for the edge. The aging value is then reset to the initial aging value. Output the update results of the winning node, the update results of the neighboring nodes, and the aging status of the edges.

5. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The generation of the connection weakening region includes: The set of edges is traversed. For each edge traversed, the current aging value of the edge is read. The current aging value is added to a preset fixed increment. The aging value is then added to the fixed increment, and the result of the addition operation is used to overwrite the aging value of the edge to form an updated aging value. The updated aging value is compared with the set aging threshold. If the updated aging value is greater than or equal to the set aging threshold, the edge is deleted from the edge set. When the edge is deleted, the node indices of the two nodes corresponding to the edge are recorded and the node index pair is used as part of the connection weakening region. After completing the aging value increment and edge deletion for all connections, the node index pairs obtained from the edge deletion will form a weakened connection region.

6. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The generation of the weak connection subgraph includes: Read the node error of all nodes in the current iteration, add the node error of each node to the node error of the previous iteration, and use the result of the addition operation as the dynamic error value of the node. Perform a comparison operation on all dynamic error values, determine the node with the largest dynamic error value as the target node, and use the index of this node as the input index for generating new nodes and inserting subsequent edges; Read the node feature vector of the nearest node of the target node, perform addition operations on the node feature vector of the target node and the node feature vector of the nearest node respectively, take the average value, and use the average value as the node feature vector of the new node. Insert the new node into the growing neural gas network, establish an edge between the new node and the target node, set the aging value of the edge to the initial aging value, and use the edge formed between the new node and the target node as a component of the connection of the new growth region. The newly formed regions and the weakened regions are combined and merged into a group of regions, and the merged region is used as a weak connection subgraph.

7. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The generation of the stable topology includes: Read the node error of all nodes in the current iteration, multiply the node error of each node with the preset decay coefficient, and use the result of the multiplication operation as the updated node error of that node. Read the node feature vectors and edge sets of all nodes. For each edge connecting two nodes, read their node feature vectors. Perform addition on the two node feature vectors respectively, take the average value, and use the average value as the update reference value for the two nodes. The above updated reference values ​​are added to the corresponding node feature vectors in turn and divided by two, and the result is used as the iterative update value of the node feature vectors. After all node feature vectors have been updated, the updated set of node feature vectors and the updated set of edges are merged into a set of structures, and this set of structures is used as a temporary topological structure for topological convergence. Multiple rounds of iterative updates are performed on the temporary topology. The change in the node feature vector after each round of iteration is compared with a preset change threshold. The iteration result with a change less than the preset change threshold is taken as the stable topology.

8. The LED pixel defect detection method based on convolutional neural networks according to claim 1, characterized in that, The determination of bright spot defect pixels, dark spot defect pixels, or color shift defect pixels includes: Read the node errors of all nodes in the stable topology, compare the node errors with a preset error threshold, mark the nodes whose node errors are greater than the preset error threshold as error abnormal nodes, and use the error abnormal nodes as the initial content of the abnormal set. Read the set of edges in the stable topology, filter the nodes with zero edges, mark the nodes with zero edges as isolated nodes, and add the isolated nodes to the anomaly set; Read the number of node updates in the topology during the iteration process, compare the number of node updates with a preset growth density threshold, mark the nodes with an update count greater than the preset growth density threshold as abnormal growth density nodes, and add the abnormal growth density nodes to the anomaly set. Read the node index in the weak connection subgraph, mark the node corresponding to the node index in the weak connection subgraph as a weak connection exception node, and add the weak connection exception node to the exception set; Perform a correspondence matching between the node feature vectors of all nodes in the anomaly set and the node feature vectors of the stable topology. Classify the nodes according to their corresponding pixel positions and mark the corresponding pixels as bright spot defect pixels, dark spot defect pixels, or color shift defect pixels.

9. The LED pixel defect detection method based on convolutional neural network according to claim 1, characterized in that, The output of the final defect coordinate information includes: Read the node indices of pixels identified as bright spot defects, dark spot defects, and color cast defects, match the node indices with the pixel positions in the original image, and find and record the horizontal and vertical positions of the corresponding pixels as pixel coordinates based on the node indices. The recorded pixel coordinates are categorized into bright spot defect pixels, dark spot defect pixels, and color cast defect pixels, forming bright spot defect coordinate sets, dark spot defect coordinate sets, and color cast defect coordinate sets, respectively. Perform an integration operation on the set of bright spot defect coordinates, the set of dark spot defect coordinates, and the set of color cast defect coordinates, merge the three sets of coordinates into one output list, and output this output list as the final defect coordinate information.

Images