A method and apparatus for detecting end face defects of a fiber optic connector

Abstract

The application discloses a kind of detection methods and devices of fiber connector end face defect, the method includes: the image data of fiber connector end face is collected, image data is input into the convolutional neural network model of completion training, output whether the conclusion of fiber connector end face exists defect;Wherein, convolutional neural network model is configured to include: five convolutional layers, five pooling layers and three fully connected layers, five convolutional layers and five pooling layers are stacked in an alternating manner, three fully connected layers are located after five convolutional layers and five pooling layers, along the direction of data input to output of convolutional neural network model, after the first pooling layer experienced, there is a normalization layer.The method can improve the precision of fiber connector end face defect detection by improved convolutional neural network structure, i.e., by five convolutional layers and pooling layers stacked alternately, normalization layer after first pooling layer and three fully connected layers, to realize high sensitivity identification of micro defect.

Description

Technical Field

[0001] This application relates to the field of optical fiber testing, and in particular to a method and apparatus for detecting defects on the end face of an optical fiber connector. Background Technology

[0002] As optical fiber communication technology develops towards higher speeds and higher reliability, the requirements for the quality inspection of optical fiber connector end faces are increasing. However, the following technical problems are encountered in the actual inspection process: with the increasing requirements for defect detection accuracy, existing methods encounter insufficient sensitivity when identifying micro-scratches and small dents; with the diversification of defect types, traditional methods relying on manually designed features encounter bottlenecks in adaptability and generalization ability. Summary of the Invention

[0003] This application provides a method for detecting defects on the end face of an optical fiber connector, which can achieve high-precision detection and stronger identification of minute defects.

[0004] In a first aspect, embodiments of this application provide a method for detecting defects on the end face of an optical fiber connector, the method comprising:

[0005] Image data of the fiber optic connector end face is collected, and the image data is input into a trained convolutional neural network model to output a conclusion on whether there is a defect in the fiber optic connector end face.

[0006] The convolutional neural network model is configured to include five convolutional layers, five pooling layers, and three fully connected layers. The five convolutional layers and five pooling layers are stacked alternately, and the three fully connected layers are located after the five convolutional layers and five pooling layers. A normalization layer is set after the first pooling layer along the data input to output direction of the convolutional neural network model.

[0007] In one embodiment, the convolutional neural network is trained through the following steps:

[0008] Construct a sample set, which includes multiple image sample data covering positive and negative samples. Each image sample data includes video data acquired using an industrial camera and grayscale interferometric data obtained using white light vertical scanning interferometry.

[0009] The grayscale interferometric data of any image sample data in the sample set are arranged in time to generate an interferometric data curve. The interferometric data curve is transformed in two dimensions to obtain a two-dimensional feature map. The grayscale interferometric data in the sample set is replaced with the two-dimensional feature map to generate a preprocessed sample set.

[0010] The preprocessed sample set is classified and labeled to obtain the training set, validation set, and test set;

[0011] Based on the initialized convolutional neural network model, the model is trained using the prepared training and validation sets. The trained convolutional neural network model is then tested using the test set to obtain test results. The training of the convolutional neural network model is complete when the accuracy of the test results is greater than a preset value.

[0012] In one embodiment, a weighted fusion module is further provided between the input layer and the first convolutional layer of the convolutional neural network model. The weighted fusion module is configured to perform weighted fusion of the camera data and the two-dimensional feature map of any image sample data using the following formula to obtain fused features:

[0013] ;

[0014] ;

[0015] In the formula, Indicates fusion features, For camera data, It is a two-dimensional feature map. The dynamic weight matrix is ​​represented by W; W represents the weight matrix of the convolution kernel, and b represents the bias term. This indicates a channel splicing operation.

[0016] In one embodiment, an adaptive Laplacian edge enhancement module is further provided between the weighted fusion module and the first convolutional layer. The adaptive Laplacian edge enhancement module is configured to enhance the edges of the fused features through the following steps:

[0017] Convolutional calculations are performed on the fused features to obtain the edge feature map;

[0018] The mean gradient of each convolutional channel is calculated, and enhancement weights are dynamically generated based on the mean gradient. These enhancement weights are then used as weights for the edge feature map. The edge feature map is then superimposed with the fused feature to obtain the enhanced feature.

[0019] In one embodiment, the gradient mean is calculated using the following formula:

[0020] = ;

[0021] In the formula, m represents the horizontal width of the two-dimensional feature map, and n represents the vertical height of the two-dimensional feature map. x The pixel index representing the horizontal width of the two-dimensional feature map. y The pixel index representing the vertical height of the two-dimensional feature map. c It is the index of the channel. Let represent the absolute value of the gradient of the c-th channel of the edge feature map E at position (x,y).

[0022] In one embodiment, during training, network parameters are adjusted iteratively based on the stochastic gradient descent method until the maximum number of iterations is reached to complete one round of training. After the convolutional neural network model completes one round of training, it is tested using a test set.

[0023] In one embodiment, a 5px wide sliding window is used to convert the interference data curve into a two-dimensional feature map.

[0024] In one embodiment, the method further includes:

[0025] Before generating an interference curve by arranging the grayscale interferometric data of any image sample data in the sample set according to time, the grayscale interferometric data is compensated using the following calibration formula:

[0026] ;

[0027] In the formula, This represents the positional parameters of the fiber optic connector end face at a certain point in the original interferometric data. For local height difference, The camera tilt angle.

[0028] In one embodiment, when acquiring camera data, an LED light source coaxially positioned with the industrial camera is used for illumination.

[0029] Secondly, this application also provides a fiber optic connector end-face defect detection device, the device comprising:

[0030] Industrial cameras, which have coaxially arranged LED light sources, are used to acquire image data from the end face of fiber optic connectors;

[0031] A white light interferometer is used to perform vertical scanning interference on the end face of an optical fiber connector using white light to obtain grayscale interference data.

[0032] The processing module is configured to perform defect detection on the fiber optic connector end face based on data acquired by an industrial camera and a white light interferometer, using the fiber optic connector end face defect detection method described above, and output a conclusion on whether there is a defect on the fiber optic connector end face.

[0033] The aforementioned method for detecting defects on the end face of fiber optic connectors employs a specially designed convolutional neural network model structure: five convolutional layers and five pooling layers are alternately stacked to form a feature extraction network, followed by three fully connected layers to complete the classification decision, and a normalization layer is set after the first pooling layer. This structural design effectively improves the detection capability for minute defects on the end face of fiber optic connectors, achieving high detection accuracy. Attached Figure Description

[0034] Figure 1 This is a framework diagram of a convolutional neural network model in one embodiment;

[0035] Figure 2 This is a flowchart of the training steps for a convolutional neural network in one embodiment;

[0036] Figure 3 This is a flowchart illustrating edge enhancement using fused features in one embodiment;

[0037] Figure 4 This is a frame diagram of a fiber optic connector end-face defect detection device in one embodiment. Detailed Implementation

[0038] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0040] In fiber optic communication, it is often necessary to connect two fiber optic segments from different modules, devices, or systems. To improve the stability of fiber optic connections and reduce information transmission loss caused by the connection, a special type of connector has been developed for connecting two fiber optic segments: the fiber optic connector. A fiber optic connector is a passive component used to stably, but not permanently, connect two or more optical fibers. Its function is to achieve a non-permanent, precise connection between the two end faces of the optical fibers, enabling the optical path to transmit along a specified channel to achieve predetermined requirements.

[0041] The fiber optic connector end face refers to the flat part of the fiber optic connector used to connect two optical fibers. It is directly related to the transmission quality of optical signals, so its design and manufacturing have strict requirements.

[0042] In one embodiment, a method for detecting defects on the end face of an optical fiber connector is provided, the method comprising:

[0043] Image data of the fiber optic connector end face is collected, and the image data is input into a trained convolutional neural network model to output a conclusion on whether there is a defect in the fiber optic connector end face.

[0044] Among them, such as Figure 1 As shown, the convolutional neural network model is configured to include: five convolutional layers, five pooling layers, and three fully connected layers. The five convolutional layers and five pooling layers are stacked alternately, and the three fully connected layers are located after the five convolutional layers and five pooling layers. Along the direction from data input to output of the convolutional neural network model, a normalization layer is set after the first pooling layer.

[0045] Specifically, automated defect detection is achieved by inputting the acquired image data of the fiber optic connector end face into a trained convolutional neural network model. This convolutional neural network model employs a hierarchical structure: five convolutional layers are responsible for extracting features from the fiber optic connector end face image layer by layer, capturing hierarchical feature representations from edges and textures to complex patterns through a kernel weight sharing mechanism; the pooling layers connected after each convolutional layer can downsample the feature map, reducing parameter computation while preserving features and enhancing the model's robustness to changes in feature location.

[0046] It should be noted that a normalization layer is introduced after the first pooling layer. By standardizing the activation values, the feature distribution shift caused by changes in illumination is effectively suppressed, accelerating training convergence and improving sensitivity to minor defects. Three fully connected layers map the extracted spatial features to decision functions in the sample space, and the presence of defects on the end face is determined through multi-layer nonlinear transformations.

[0047] In this embodiment, a normalization layer is added to the convolutional neural network model to improve the accuracy of deep learning training. Furthermore, this normalization layer is executed after the convolutional and pooling layers, enhancing the generalization ability of the convolutional neural network model. Specifically, in this embodiment, the normalization layer is placed after the first pooling layer. By normalizing the output of the first pooling layer, the original defect features can be preserved, preventing the low-order features (scratches / dents) of the fiber optic connector end-face defects from being diluted after multiple convolutional samplings through multiple pooling layers. The output of the first pooling layer contains 78% of the edge features. After normalization, the low-order features of the fiber optic connector end-face defects can be preserved to the greatest extent, thereby improving the sensitivity of fiber optic connector end-face defect detection. In addition, the normalization process of the normalization layer can also offset the local gray-scale fluctuations caused by the shift in illumination angle to the greatest extent. Normalization is performed after the first pooling layer to avoid the accumulation of gray-scale shifts with feature transmission and to prevent the signal of small defects from being submerged by noise, thereby preserving the core features of small defects and greatly improving the detection sensitivity of scratches smaller than 30 micrometers.

[0048] In this embodiment, the method employs a specially designed convolutional neural network model structure: five convolutional layers and five pooling layers are alternately stacked to form a feature extraction network, followed by three fully connected layers to complete the classification decision, and a normalization layer is set after the first pooling layer. This structural design effectively improves the detection capability of minute defects on the end face of fiber optic connectors, achieving high detection accuracy.

[0049] In one embodiment, such as Figure 2 As shown, a convolutional neural network is trained through the following steps:

[0050] Step 201: Construct a sample set, which includes multiple image sample data covering positive and negative samples. Each image sample data includes video data acquired using an industrial camera and grayscale interference data obtained using white light vertical scanning interference.

[0051] Specifically, the sample set includes negative samples (i.e., end-face samples with defects such as scratches, dents, and contamination) and positive samples (completely normal end-face samples) to ensure that the convolutional neural network model can learn the feature differences between defective and normal states. Each image sample data consists of two different modalities: end-face image data acquired by a high-precision industrial camera under specific coaxial light source illumination, which can clearly present the two-dimensional appearance morphology of the end face; and grayscale interference data acquired simultaneously using white light vertical scanning interferometry, which records the microscopic height features of the three-dimensional morphology of the end face by measuring the optical path difference (e.g., reflecting scratch depth).

[0052] Step 202: Arrange the grayscale interferometric data of any image sample data in the sample set according to time to generate an interferometric data curve, perform two-dimensional transformation on the interferometric data curve to obtain a two-dimensional feature map, replace the grayscale interferometric data in the sample set with the two-dimensional feature map, and generate a preprocessed sample set.

[0053] In this embodiment, the number of grayscale interference data points collected when constructing the sample set can be adjusted according to actual detection needs. The specific number is related to the fiber diameter and scanning accuracy. For example, when using higher precision scanning, the number of interference data points may increase to 2048 or more; while for smaller fiber diameters or lower precision requirements, the number of points can be reduced to 512. This embodiment uses 1024 data points as an example, but practical applications are not limited to this specific number.

[0054] For example, for the grayscale interferometric data (a one-dimensional discrete data sequence obtained by vertical scanning interference with white light) contained in each image sample data, an interferometric data curve containing 1024 interferometric data points is generated by arranging them in time series. This curve fully records the change in interference intensity at each interferometric data point during the scanning process.

[0055] Furthermore, the one-dimensional curve was transformed into a two-dimensional model using a 5-pixel wide sliding window (this size is set based on the optical characteristic that the minimum defect size of the fiber optic connector end face is 10μm). The curve was sampled in 1-pixel increments, generating a total of 1020 data windows. Zero-padding was used to maintain data integrity in the data boundary regions generated by the window sliding. The processed data was rearranged into a 224×224 pixel two-dimensional feature map according to row priority rules. This transformation method preserves the depth physical information (such as scratch depth, indentation height, and other morphological features) contained in the original interferometric data while resolving the dimensionality matching problem with the end-face camera data. Finally, the original grayscale interferometric data of all samples in the sample set were replaced with the transformed two-dimensional feature map, forming a preprocessed sample set with a unified structure.

[0056] Step 203: Classify and label the preprocessed sample set to obtain the training set, validation set, and test set.

[0057] Specifically, each sample (including the end face image and the converted interference feature map) is precisely labeled according to the defect judgment criteria, distinguishing between negative samples (with defects such as scratches, dents, and contamination) and positive samples (completely normal). A stratified random sampling strategy is further employed, dividing the entire sample set into a training set, a validation set, and a test set in a 7:1:2 ratio. 70% of the samples are used as the training set for learning and optimizing the neural network model parameters, 10% as the validation set for hyperparameter tuning and monitoring the training process (using an early stopping mechanism to prevent overfitting), and the remaining 20% ​​as the test set for the final evaluation of the model's generalization performance and detection accuracy.

[0058] This partitioning method ensures that the proportion of samples of each category is balanced in the subset, and also guarantees the effectiveness of model training and the objectivity of evaluation results.

[0059] Step 204: Based on the initialized convolutional neural network model, train it using the prepared training and validation sets, and test it using the test set based on the trained convolutional neural network model to obtain the test results. When the accuracy of the test results is greater than the preset value, the training of the convolutional neural network model is completed.

[0060] Specifically, during training, a stochastic gradient descent (SGD) optimizer is used. The model parameters are iteratively optimized by calculating cross-entropy loss through forward propagation and updating weights through backpropagation. A validation set is used to monitor the training process, and an early stopping mechanism prevents overfitting. After training is complete, the trained model is tested using an independent test set. Objective test results are obtained by calculating metrics such as sensitivity, specificity, and accuracy. If the test accuracy consistently reaches a preset threshold (e.g., 99.0%), the convolutional neural network model is considered to have completed training and is ready for practical deployment. Otherwise, hyperparameters or data need to be adjusted and retrained until the performance requirements are met.

[0061] In one embodiment, a weighted fusion module is further provided between the input layer and the first convolutional layer of the convolutional neural network model. The weighted fusion module is configured to perform weighted fusion of the camera data and the two-dimensional feature map of any image sample data using the following formula to obtain fused features:

[0062] ;

[0063] ;

[0064] In the formula, Indicates fusion features, For camera data, It is a two-dimensional feature map. The dynamic weight matrix is ​​represented by W; W represents the weight matrix of the convolution kernel, and b represents the bias term. This indicates a channel splicing operation.

[0065] Specifically, dynamic weights The weighted fusion is performed in the weighted fusion module, generated by a 3x3 convolutional layer. After sigmoid activation, the values ​​range from [0,1], and the dimension is the same as that of two-dimensional features. Figure 1 To. Features of the end face image (64-dimensional). This is the converted interference feature map. This indicates a channel concatenation operation. Channel concatenation can preserve spatial information and only superimpose modal dimensions, allowing dual-modal data to be stored together in a single feature tensor, avoiding reliance on features from a single modal data.

[0066] Furthermore, a weighted fusion output is performed, calculated using the following formula:

[0067] .

[0068] Based on dynamic weights By performing weighted summation of dual-modal data pixel by pixel, each pixel position is not affected by the information of surrounding pixels after fusion, and the network can eventually automatically focus on the depth anomalous region revealed by the interference curve (such as the bottom of the scratch).

[0069] In one embodiment, such as Figure 3 As shown, an adaptive Laplacian edge enhancement module is also set between the weighted fusion module and the first convolutional layer. The adaptive Laplacian edge enhancement module is configured to enhance the edges of the fused features through the following steps:

[0070] Step 301: Perform convolution calculation on the fused features to obtain the edge feature map;

[0071] Specifically, a 3×3 Laplacian convolution kernel is used to process the weighted and fused feature map F. fused Perform convolution operations. This convolution kernel has the characteristic of a negative value at the center and a positive value at the periphery, and highlights the gray-level abrupt change region through second-order differential calculation.

[0072] During convolution, the kernel matrix slides across the feature map with a stride of 1, calculating the weighted sum of the local region and the convolution kernel pixel by pixel to generate the edge feature map E. In the edge feature map E, positive value regions correspond to the raised features of the defect edge (such as the bulges on both sides of a scratch), negative value regions correspond to the concave features (such as the bottom of a scratch), and zero value regions represent a flat surface with no significant changes.

[0073] The edge feature map is calculated by utilizing the sensitivity of the Laplacian operator to abrupt changes in grayscale, thereby enhancing the edge gradient intensity of minor defects in the fused features and suppressing the response of flat regions.

[0074] Step 302: Calculate the mean gradient of each convolutional channel, and dynamically generate enhancement weights based on the mean gradient. Use the enhancement weights as the weights of the edge feature map, and superimpose the edge feature map with the fused feature to obtain the enhancement feature.

[0075] Specifically, after obtaining the edge feature map, the average gradient of each convolutional channel is calculated, that is, the average of the absolute gradient values ​​of all pixels in each convolutional channel. The average gradient of the convolutional channel can reflect the overall intensity level of the edge response of that channel.

[0076] Furthermore, enhancement weights are dynamically generated based on the average gradient of each convolutional channel. This is achieved by normalizing the weights by dividing the average gradient of each convolutional channel by the maximum average gradient of all channels, thus controlling the weight coefficients to range between 0 and 1. Channels with significant edge feature maps therefore receive larger weights, while channels with weaker responses receive correspondingly smaller weights.

[0077] Furthermore, the fused features are superimposed with the weighted edge feature map. This operation enhances the real defect area due to its higher edge gradient value, while the flat background area remains basically unchanged due to its weak edge response. This effectively suppresses noise amplification while improving the contrast of defect edges, and outputs the enhanced features for use by subsequent network layers.

[0078] In one embodiment, the gradient mean is calculated using the following formula:

[0079] = ;

[0080] In the formula, m represents the horizontal width of the two-dimensional feature map, and n represents the vertical height of the two-dimensional feature map. x The pixel index representing the horizontal width of the two-dimensional feature map. y The pixel index representing the vertical height of the two-dimensional feature map. c It is the index of the channel. Let represent the absolute value of the gradient of the c-th channel of the edge feature map E at position (x,y).

[0081] Specifically, in |E(x,y,c)|, x can represent the pixel position index in the horizontal direction of the two-dimensional feature map (with a value range of 1≤x≤m); y can represent the pixel position index in the vertical direction of the two-dimensional feature map (with a value range of 1≤y≤n); and c is the channel index (with a value of an integer sequence, such as 1~64). This can represent the absolute value of the gradient in the c-th channel of the edge feature map E at position (x, y). This absolute value calculation eliminates the influence of gradient direction, retaining only gradient intensity information, and is used to quantify the strength of the edge response at that location.

[0082] For example, m and n can be 224, meaning the mean gradient can be calculated using the following formula:

[0083] = ;

[0084] Finally, the mean gradient μ is obtained through summation and averaging operations. c It can reflect the average intensity level of the edge response at all pixel locations in the c-th convolutional channel.

[0085] In one embodiment, during training, network parameters are iteratively adjusted based on the stochastic gradient descent method until the maximum number of iterations is reached to complete one round of training. After the convolutional neural network model completes one round of training, it is tested using a test set.

[0086] Specifically, in each iteration, a small batch of samples is randomly drawn from the training set and input into the network. The cross-entropy loss between the predicted output and the true label is calculated through forward propagation. Then, the gradient of the loss function with respect to the parameters of each network layer is calculated using the backpropagation algorithm, and the result is determined based on the learning rate (e.g., initially set to 10). -4 Hyperparameter settings such as momentum (e.g., 0.9) and weight decay (e.g., 0.0005) update the convolution kernel weights and bias terms.

[0087] This process is repeated until the preset maximum number of iterations (e.g., 10,000) is reached, completing one round of training. Immediately after training, performance is validated using an independent test set. Test samples are input into the trained model, and model performance is quantified by calculating metrics such as true positive rate (e.g., sensitivity), true negative rate (e.g., specificity), and overall accuracy. Confusion matrices are recorded to analyze the identification of various defects. Based on a preset performance threshold (e.g., accuracy ≥ 99.0%), it is determined whether the convolutional neural network model meets the deployment standard.

[0088] In one embodiment, the method for detecting defects on the end face of a fiber optic connector further includes:

[0089] Before generating an interference curve by arranging the grayscale interferometric data of any image sample data in the sample set according to time, the grayscale interferometric data is compensated using the following calibration formula:

[0090] ;

[0091] In the formula, This represents the positional parameters of the fiber optic connector end face at a certain point in the original interferometric data. For local height difference, The camera tilt angle.

[0092] Specifically, xThe position parameter representing a point on the fiber optic connector end face in the original interferometric data reflects the spatial coordinates of that point before compensation; Δ h The local height difference between this point and the reference plane can be obtained by measuring with a white light interferometer, reflecting the undulations and changes in the microstructure of the fiber optic connector end face; θ The angle between the optical axis of the industrial camera and the normal to the end face.

[0093] This compensation mechanism, by introducing correction terms for angle and height differences, can effectively offset the spatial distortion of interference data caused by camera tilt deviation and uneven end face, thereby improving the accuracy of defect detection.

[0094] In one embodiment, the image data includes: video data acquired using an industrial camera and grayscale interference data obtained using white light vertical scanning interference;

[0095] Before inputting the image data into the trained convolutional neural network model, the grayscale interference data is first arranged in time to generate an interference data curve. The interference data curve is then transformed in two dimensions to obtain a two-dimensional feature map.

[0096] For example, the image data includes image data of the fiber optic connector end face acquired by an industrial camera, and grayscale interference data obtained by white light vertical scanning interferometry. For instance, the image data of the fiber optic connector end face acquired by the industrial camera could be an image acquired at a resolution of 640×480 using a MindVision 300,000-pixel industrial camera equipped with a lens with an optical magnification of 0.13-2. The number of grayscale interference data points can be adjusted according to actual inspection needs, and the specific number is related to the fiber diameter and scanning accuracy. This embodiment uses 1024 data points as an example, but practical applications are not limited to this specific number.

[0097] Before inputting these two types of data into the trained convolutional neural network model, the grayscale interferometric data needs to be preprocessed. The collected discrete grayscale interferometric data are arranged according to a time series to generate an interferometric data curve containing 1024 grayscale interferometric data points. This one-dimensional sequence of the interferometric data curve completely records the change in optical path difference at each point of the fiber optic connector end face during vertical scanning, including the three-dimensional topographic depth information of the end face. Furthermore, the curve undergoes dimensionality transformation processing, using a sliding window with a width of 5px and a step size of 1px to slide sample the one-dimensional sequence. The width of this window is determined based on the optical imaging characteristics of the fiber optic connector end face with a minimum defect size of 10μm, ensuring that even the finest defect features can be captured.

[0098] During the sliding sampling process, a total of 1020 data windows were generated. For regions at the end of the sequence that could not be fully sampled, zero-padding was used to maintain data integrity. Following a row-major arrangement rule, these processed data windows were reconstructed into a 224×224 pixel two-dimensional feature map.

[0099] In this embodiment, the entire preprocessing process maintains the original physical characteristics of the interference data, including key morphological parameters such as scratch depth and indentation height, so that it is consistent with the end face camera data in terms of data dimension. This ensures that the subsequent neural network can use the appearance visual features and depth morphological features for collaborative analysis, providing data support for achieving high-precision defect detection.

[0100] In one embodiment, when acquiring camera data, an LED light source coaxially positioned with the industrial camera is used for illumination.

[0101] Specifically, a beam splitter precisely guides the light emitted from the LED onto a path perfectly aligned with the optical axis of the camera lens, ensuring that the light illuminates the fiber optic connector end face perpendicularly. This coaxial illumination method ensures that the light beam aligns with the normal direction of the end face, fundamentally eliminating specular reflections caused by differences in end face curvature and material refractive index (e.g., core refractive index 1.467, cladding refractive index 1.447). Because the light is incident perpendicularly, minor scratches, dents, and other defects on the end face surface will cause diffuse reflection, while normal, flat areas will show regular reflections, thus clearly revealing defect features in the image. This illumination method effectively avoids the shadow effects and overexposure problems caused by side illumination, resulting in high-quality end face images with uniform illumination, clean backgrounds, and clear details.

[0102] In one embodiment, the steps for detecting defects on the end face of an optical fiber connector after the convolutional neural network has been trained are as follows:

[0103] Dual-mode data acquisition is performed on the end face of the fiber optic connector to be inspected. An industrial camera equipped with a coaxial LED light source acquires visible light images of the end face, and a white light vertical scanning interferometer obtains interference data for the same area. The acquired raw interference data needs to be preprocessed using the calibration formula: x'=x+0.38·Δh·sinθ. Compensation is applied to each data point to effectively eliminate measurement errors caused by installation deviations or end face tilt.

[0104] This embodiment uses 1024 interferometric data points as an example. After data preprocessing, the interferometric data are arranged in a time series to generate a one-dimensional interferometric curve. A two-dimensional transformation is then performed using a sliding window with a width of 5 pixels. The window slides in 1-pixel increments, generating 1020 data windows. After zero-value filling and row-first rule-based recombination, a 224×224 pixel two-dimensional feature map is finally formed.

[0105] Furthermore, the processed bimodal data (end-face image and interference feature map) is input into the pre-trained convolutional neural network model. The convolutional neural network model performs adaptive weighted fusion of the two types of data through a cross-modal fusion module: a dynamic weight matrix α is generated using a 3×3 convolution kernel, and the weight values ​​are limited to the range [0,1] using the sigmoid function, and then calculated according to the formula: F fused =α·F img +(1-α)·F intf Perform a pixel-by-pixel weighted summation.

[0106] The fused features are fed into an adaptive Laplacian edge enhancement module. This module uses a 3×3 Laplacian kernel for convolution operations to generate an edge feature map, where positive values ​​correspond to convex edges and negative values ​​correspond to concave edges. The mean gradient of each channel is calculated, and enhancement weights W are generated based on this. The original fused features are then superimposed with the weighted edge feature map to output the enhanced feature F. en .

[0107] The enhanced features are then incorporated into the network's backbone feature extraction section. After alternating processing through five convolutional layers and five pooling layers, batch normalization is performed after the first pooling layer. This batch normalization after the first pooling layer addresses the characteristics of fiber optic connector end-face defects: approximately 78% of defect feature responses are concentrated in the shallow network, and early normalization effectively preserves the core features of minor defects. Finally, three fully connected layers map the extracted features to the classification space, outputting a judgment result of whether the end face is normal or defective. For defect samples, further classification information on the defect type is output.

[0108] Based on the same concept, such as Figure 4 As shown, the application also provides a fiber optic connector end-face defect detection device, the device comprising:

[0109] Industrial camera 401, the industrial camera has a coaxially set LED light source, used to acquire camera data of the end face of the fiber optic connector;

[0110] The white light interferometer 402 is used to perform vertical scanning interference on the end face of an optical fiber connector using white light to obtain grayscale interference data.

[0111] The processing module 403 is configured to use the data acquired by the industrial camera and white light interferometer to perform defect detection on the fiber optic connector end face based on the above-mentioned fiber optic connector end face defect detection method, and output a conclusion on whether there is a defect on the fiber optic connector end face.

[0112] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a solid-state disk (SSD), etc.

[0113] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method of detecting an endface defect of a fiber optic connector, the method comprising: The method comprises: Collecting image data of the end face of the fiber connector, inputting the image data into a trained convolutional neural network model, and outputting a conclusion on whether the end face of the fiber connector has defects; The convolutional neural network model is configured to include five convolutional layers, five pooling layers, and three fully connected layers, the five convolutional layers and the five pooling layers are stacked in an alternating manner, and the three fully connected layers are located after the five convolutional layers and the five pooling layers. A normalization layer is arranged after the first pooling layer in the direction of data input to output of the convolutional neural network model. The convolutional neural network is trained by the following steps: A sample set is constructed, which includes a plurality of image sample data covering positive samples and negative samples, any image sample data includes camera data collected using an industrial camera and gray value interference data obtained using white light vertical scanning interference; The gray value interference data of any image sample data in the sample set is arranged according to time to generate an interference data curve, the interference data curve is two-dimensionally converted to obtain a two-dimensional feature map, the gray value interference data in the sample set is replaced with the two-dimensional feature map to generate a preprocessed sample set; The preprocessed sample set is classified and labeled to obtain a training set, a validation set, and a test set; Based on the initialized convolutional neural network model, the training set and the validation set are used for training, and based on the trained convolutional neural network model, the test set is used for testing to obtain a test result. When the accuracy of the test result is greater than a preset value, the training of the convolutional neural network model is completed.

2. The method for detecting defects of the end face of the fiber connector according to claim 1, wherein a weighted fusion module is further arranged between the input layer and the first convolutional layer of the convolutional neural network model, the weighted fusion module is configured to perform weighted fusion on the camera data and the two-dimensional feature map of any image sample data according to the following formula to obtain a fusion feature: An adaptive Laplacian edge enhancement module is further arranged between the weighted fusion module and the first convolutional layer, the adaptive Laplacian edge enhancement module is configured to perform edge enhancement on the fusion feature by the following steps: ； ； wherein, denotes fusion features, is the camera data, is the two-dimensional feature map, denotes a dynamic weight matrix; W denotes a weight matrix of a convolution kernel, and b denotes a bias term, denotes a channel concatenation operation.

3. The method of detecting an endface defect of a fiber optic connector of claim 2, wherein, Convolution calculation is performed on the fusion feature to obtain an edge feature map; The gradient mean value of each convolution channel is calculated, and an enhancement weight is dynamically generated based on the gradient mean value, the enhancement weight is used as the weight of the edge feature map, the edge feature map is superimposed with the fusion feature to obtain an enhanced feature. The gradient mean value is calculated according to the following formula:

4. The method of detecting an endface defect of a fiber optic connector of claim 3, wherein, In the training process, the network parameters are adjusted in an iterative manner based on the stochastic gradient descent method until the maximum number of iterations is reached, one round of training is completed, and the convolutional neural network model is tested after one round of training. = ； where m represents a horizontal width of the two-dimensional feature map, n represents a vertical height of the two-dimensional feature map, x a pixel position index representing the horizontal width of the two-dimensional feature map, y a pixel position index representing the vertical height of the two-dimensional feature map, c is an index of a channel, represents an absolute value of a gradient of the edge feature map E at the position (x, y) in the cth channel.

5. The method of detecting an endface defect of a fiber optic connector of claim 1, wherein, 6. The method for detecting defects of the end face of the fiber connector according to claim 1, wherein a 5px wide sliding window is used to convert the interference data curve into the two-dimensional feature map. The method further comprises: ​ 7. The method of detecting an endface defect of a fiber optic connector of claim 1, wherein, ​ Before generating an interference data curve by arranging the gray value interference data of any image sample data in the sample set according to time, the gray value interference data is compensated by using the following calibration formula: ； wherein represents the position parameter of the fiber connector end face at a point in the original interference data, is the local height difference, is the camera tilt angle.

8. The method of detecting an endface defect of a fiber optic connector of claim 1, wherein, When the camera data is collected, an LED light source coaxially arranged with the industrial camera is used for illumination.

9. An optical fiber connector endface defect detection apparatus, characterized by, The device comprises: an industrial camera having an LED light source coaxially arranged therewith, for collecting camera data of an end face of a fiber connector; a white light interferometer for performing vertical scanning interference on the end face of the fiber connector using white light to obtain gray value interference data; a processing module configured to perform defect detection on the end face of the fiber connector based on the data collected by the industrial camera and the white light interferometer by applying the method for detecting defects of the end face of the fiber connector according to any one of claims 1 to 8, and output a conclusion on whether the end face of the fiber connector has defects.

Images