LED Strip Defect Detection Method and System Based on Deep Support Vector Description

Abstract

The present invention discloses a method and system for detecting defects in LED light strips based on deep support vector description, which relates to the field of computer vision technology. The specific steps are: obtaining a positive light image of a single LED lamp bead; performing threshold segmentation on the positive light image to divide it into normal samples and abnormal samples; performing model training on the positive light image through a Contractive‑Deep SVDD neural network to obtain a training model, obtain the minimum hypersphere representation of the original feature space, calculate the hypersphere radius and the abnormality score and divide the threshold; using the normal samples and abnormal samples to test the training model, evaluate the model performance, and determine whether the positive light image has defects. The present invention provides a deep single classification algorithm and implementation example, which can realize end-to-end simultaneous detection of whether LED light strip products have brightness and color difference defects.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more particularly to a method and system for detecting defects in LED light strips based on deep support vector description. Background Art

[0002] The production process of LED light strips has always been plagued by a mismatch between inspection speed and actual production speed. Due to limited testing conditions, brightness testing is performed using a photodiode to detect point light sources on the conveyed strips. Parameter thresholds are set, and positioning clamps are used to identify defects in the sample. Colorimetric testing of LEDs often relies on the human eye to determine if the target sample's brightness and color meet standards. Manufacturers typically use laboratory integrating sphere spectrometers or color temperature illuminometers to spot-sample the color temperature of batches. These samples are difficult to represent overall characteristics. In the high-speed environment of production lines, traditional optical inspection methods are simply unable to meet the production line's manufacturing speed requirements and meet the required yield rate. The developed light strip rapid inspection system uses machine vision technology to innovate traditional LED light strip inspection methods, replacing traditional optical inspection methods with non-contact inspection methods. It also incorporates artificial intelligence into the equipment production process, laying the foundation for intelligent manufacturing in industrialized production.

[0003] Existing machine vision defect detection technologies are mostly based on traditional pattern recognition algorithms and have been proven in multiple industries both domestically and internationally. These methods compare the defect and background histograms to determine the boundary thresholds between abnormal and normal samples and determine whether the sample under inspection is an abnormal defect. In real industrial environments, the ratio of normal to abnormal samples is extremely uneven, and abnormal samples also vary in distribution due to their non-identity, posing a significant challenge to traditional methods. Emerging deep learning-based detection methods can build classification models entirely based on normal test samples, while providing end-to-end detection without consuming large amounts of memory and achieving more effective detection results. Compared to traditional machine vision methods, they are more efficient and versatile. Therefore, how to detect defects in LED light strips based on the shrinking deep support vector data description algorithm is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for LED light strip defect detection based on deep support vector description, which can address the problems of low accuracy and high cost of traditional LED light strip defect detection methods, use a single class of normal sample data to train a deep single classification model, and improve detection efficiency.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting defects in LED light strips based on deep support vector description, the specific steps of which include the following:

[0006] Get the positive light image of a single LED lamp bead;

[0007] Performing threshold segmentation on the positive luminescence image to divide it into normal samples and abnormal samples;

[0008] Performing model training on the positive luminescence image using a Contractive-Deep SVDD neural network to obtain a training model, obtaining a minimum hypersphere representation of the original feature space, calculating the hypersphere radius and anomaly score, and dividing the threshold;

[0009] The training model is tested using the normal samples and the abnormal samples to evaluate the model performance and determine whether the positive luminescence image has defects.

[0010] Optionally, the positive light images of the LED multi-light strips are collected at equal intervals within a fixed time, the positive light images of the LED multi-light strips are segmented and the ROI area is extracted with a fixed window size, and the fixed area is divided into positive light images of single LED lamp beads.

[0011] Optionally, the threshold segmentation method for the positive luminescence image is: adjusting the size of the positive luminescence image, and performing the threshold segmentation through image color coordinates and histogram conversion, and the calculation formula is:

[0012]

[0013] Among them, L* represents the lightness of the color, a* and b* represent the chromaticity of the color, a* positive value represents reddish, negative value represents greenish; b* positive value represents yellowish, negative value represents bluish, X, Y, Z represent the dot product of the RGB three-channel components, f() represents the linear mapping function, X N 、Y N , Z N are the modification coefficients of X, Y, and Z respectively;

[0014] Convert the original RGB image to L*a*b* color space and apply the color difference formula to calculate:

[0015]

[0016] in, Indicates the calculated color difference value, ΔL*, Δa*, and Δb* represent the distribution difference between the current sample and the standard normal sample in the L*a*b* color space;

[0017] according to Divide the normal lamp bead samples into abnormal samples.

[0018] Optionally, the method further includes pre-training the training model using a contraction convolutional autoencoder, extracting Encoder module parameters as Contractive-DeepSVDD neural network parameters, and inputting unlabeled data of normal light strip samples for model training.

[0019] Optionally, the performance of the training model is evaluated using an anomaly score evaluation mechanism, wherein the anomaly score evaluation mechanism is established to obtain a score threshold of the distance between the training sample and the center c of the hypersphere, and the score of the normal sample farthest from the center c of the hypersphere is used as the hypersphere radius to evaluate the abnormal sample;

[0020] The anomaly score evaluation mechanism is:

[0021] g(x i )=||φ(x i ;W * )-c|| 2 ;

[0022] Among them, x i represents the data sample, c is the center of the hypersphere, W * represents the parameters of the deep neural network, and φ represents the training model for the input x i Nonlinear mapping function.

[0023] Optionally, the normal samples and abnormal samples are input into the training model to obtain different training abnormality score values g i , the ROC receiver operating characteristic curve is drawn by the abnormal score and the AUC value is calculated to test the detection performance and distinguish normal luminous lamp beads from abnormal luminous lamp beads.

[0024] On the other hand, a LED light strip defect detection system based on deep support vector description is provided, including an image acquisition module, a sample division module, a model training module, and a testing module; wherein,

[0025] The image acquisition module is used to acquire the positive light emitting image of a single LED lamp bead;

[0026] The sample segmentation module is used to perform threshold segmentation on the positive luminescence image to divide it into normal samples and abnormal samples;

[0027] The model training module is used to perform model training on the positive luminescence image through a Contractive-Deep SVDD neural network to obtain a training model, obtain a minimum hypersphere representation of the original feature space, calculate the hypersphere radius and anomaly score, and divide the threshold;

[0028] The testing module is used to test the training model using the normal samples and abnormal samples, evaluate the model performance, and determine whether the positive luminescence image has defects.

[0029] Optionally, a pre-training module is also included, which is used to pre-train the training module using a contraction convolutional autoencoder.

[0030] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for detecting defects in LED light strips based on deep support vector description, which has the following beneficial technical effects:

[0031] (1) For the task of detecting defects in LED light strips, a complete set of machine vision solutions is provided to process light strip samples using image processing technology and deep single classification algorithms. Compared with traditional machine vision methods that require manual extraction of feature attributes, the present invention provides an end-to-end solution that can achieve real-time monitoring of assembly line products. It only needs to use normal samples as model training. After training, defect detection can be completed without secondary preprocessing of sample images.

[0032] (2) The photodiode method in the existing automated testing method can only test the brightness and darkness properties of the light strip, and cannot detect the color difference and color temperature problems existing in the colored light strip. The solution of the present invention can simultaneously process multiple optical properties of the lamp beads;

[0033] (3) The integrating sphere test method in the existing automated testing method can only perform sampling tests on light strip products, and the test time for a single lamp bead is more than five minutes. The solution of the present invention can process and detect defective lamp beads in parallel, which is more efficient and convenient than the integrating sphere method;

[0034] (4) Traditional support domain-based machine learning algorithms can map low-dimensional space samples to high-dimensional separable space through kernel techniques. However, this method requires a large amount of memory to store kernel plane support vectors. The solution of the present invention does not require caching the sample space and only needs to compare the anomaly score threshold to complete product detection. The model is more lightweight and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0036] Figure 1 is a flow chart of the method of the present invention;

[0037] Figure 2 This is a threshold segmentation histogram of the positive light emission image of the light strip in an embodiment of the present invention;

[0038] Figure 3 This is a structural diagram of the shrinkage convolutional autoencoder module in the embodiment of the present invention;

[0039] Figure 4 This is the Contractive-Deep SVDD network structure diagram of the present invention;

[0040] Figure 5 This is a ROC curve diagram of the LED light strip defect detection method in the present invention;

[0041] Figure 6 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Embodiment 1 of the present invention discloses a method for detecting defects in LED light strips based on deep support vector description, such as Figure 1 As shown, the specific steps include:

[0044] S1. Use a multi-channel LED light strip conveyor to collect positive-lighting raw image data of multiple LED light strips in blocks. Use a fixed window size for image segmentation. The window size for each acquisition is 18x3, indicating that there are 18x3 positive-lighting LED lamp bead images within the field of view of each lens. Collect data 50 times in total, and obtain 50x18x3 positive-lighting images of single LED lamp beads.

[0045] S2. Adjust the size of the LED light-emitting images in S1. The original image size exceeds 3MB. Crop the positive light-emitting images of each lamp bead to 50x50 pixels as the input image for the training model. Perform threshold segmentation through image color coordinates and histogram conversion. Manually divide the LED positive light-emitting images into normal samples and abnormal samples. There are 2646 normal lamp bead light-emitting images and 54 defective lamp bead light-emitting images.

[0046] In S2, the specific process of image color coordinate conversion is as follows:

[0047] The original RGB image is converted to the L*a*b* color space using the following color coordinate transformation formula:

[0048]

[0049] Among them, the L*a*b* color space is a uniform color space recommended by the International Commission on Illumination (CIE) in 1976. This space is a three-dimensional rectangular coordinate system. It is currently the most widely used color measurement system. The position of a color in the color space is represented by lightness L* and chromaticity coordinates a* and b*. L* represents the lightness of the color, a* positive value represents reddish, negative value represents greenish; b* positive value represents yellowish, negative value represents bluish. RGB cannot be directly converted to Lab color space, and it is necessary to use XYZ space transition. X, Y, and Z represent the dot product of the three RGB channel components respectively. In order to ensure equal mapping and make their sum equal to 1, it is necessary to multiply the modification coefficient in equal proportion, that is, X N 、Y N , Z N , f() represents the linear mapping function.

[0050] like Figure 2 As shown in the figure, after the histogram conversion, the luminous range of the normal luminous image is determined, where L_ave is 54.78, a_ave is 12.16, and b_ave is -6.96. L_ave, a_ave, and b_ave are the average values of the three color coordinates after conversion. The following color difference formula is used to divide the normal lamp bead samples and abnormal samples:

[0051]

[0052] in, The calculated color difference value is represented by ΔL*, Δa*, and Δb*, respectively, representing the distribution differences between the current sample and the standard normal sample in the L*a*b* color space. Some normal samples are used to input the deep network model for unsupervised clustering, while the remaining samples, including the abnormal light strip images, are used for model testing and effect verification.

[0053] S3: Pre-train S2 using a contraction convolutional autoencoder. Use a Contractive-Deep SVDD neural network to train a model on images of normal light strips to obtain the minimum hypersphere representation of the original feature space. Calculate the hypersphere radius and anomaly score and assign thresholds.

[0054] The convolutional neural network model used in S3 is:

[0055] The convolutional contraction autoencoder used in pre-training uses the LeNet-5 lightweight model, and the model structure is as follows Figure 3As shown in the figure, the Encoder layer is divided into two convolutional layers and one fully connected layer. In order to further reduce the computational parameters, the first convolutional layer uses 8 4x4 convolution kernels, the second convolution kernel is 4x4, the number is 4, the convolution step is 1, and the convolution layer is connected to a 2x2 maximum pooling layer and a Leaky Relu activation function as the output of the next layer. The Decoder layer and the Encoder layer have a completely symmetrical network structure. The unbiased representation of the input image and the output is obtained by minimizing the reconstruction error. The middle layer is used as the Encoder output, and the middle layer outputs 32-dimensional vector parameters. The pre-trained Encoder model parameters are copied as the Contractive-Deep SVDD neural network parameters.

[0056] The Contractive-Deep SVDD network model structure used in S3 is as follows Figure 4 , the training process is:

[0057] Contractive-Deep SVDD is based on Deep SVDD. By introducing a contraction penalty term, the input dimension is reduced to a small area, forcing the deep model to maintain the identity mapping relationship of the input samples. Under the condition of minimizing the Jacobi matrix loss, most of the reduced-dimensional data is forced to be insensitive to changes in the input gradient, capturing the structure of the samples in the low-dimensional popular space, enhancing the robustness of the model, and projecting the data into the minimum space hypersphere. The optimization function is:

[0058]

[0059] The optimization goal is to train the neural network parameters W, and obtain the identity mapping of all normal sample points to the minimum contraction hypersphere by minimizing the loss function J(θ), where W l Indicates the network layer parameters using regular constraints. In this embodiment, it is equivalent to W, n is the number of normal lamp bead luminescence samples, x i is a data sample, which is related to the image format and size; c is the center of the hypersphere, which is obtained after pre-training the shrinkage autoencoder, φ(x i ; W) represents the feature vector output by the original sample after dimensionality reduction by the autoencoder, is the Jacobian matrix of the penalty term, μ and λ are the penalty factors of the two regularization parameters, respectively. During training, it is assumed that all samples are correctly divided, that is, there is no interference from abnormal light strip samples, and the hard margin minimum hypersphere model is obtained.

[0060] S4: Test the training model obtained in S3 to obtain the test luminous image detection score, evaluate the model performance, and determine whether there are defects in the positive luminous image.

[0061] In S4, the anomaly score evaluation mechanism is established as g(x i )=||φ(x i ;W * )-c|| 2 , obtain the score threshold of the training sample from the center of the hypersphere c. i represents the data sample, c is the center of the hypersphere, W * represents the parameters of the deep neural network, and φ represents the training model for the input x i Nonlinear mapping function.

[0062] According to the model definition, the normal sample score farthest from the center c is used as the radius of the hypersphere to evaluate abnormal samples. Samples with detection results less than the threshold are normal luminous samples, otherwise they are abnormal samples. By calculating and sorting, the normal score threshold for this batch of training LED light strips is 0.0173.

[0063] The specific steps for testing the trained model in S4 are as follows:

[0064] The model is tested by using the normal and abnormal lamp bead luminous image samples obtained by S2, and the 50x50 labeled data is input into the training model to obtain different training anomaly score values g i , the ROC receiver operating characteristic curve is drawn by the anomaly score and the AUC (area under the curve) is calculated to test the detection performance, such as Figure 5 The results show that the normal and abnormal LEDs can be distinguished. The final result of the defect detection effect of LED positive light image is AUC value of 98.7, standard deviation of 0.1, and average accuracy of more than 99%.

[0065] Embodiment 2 of the present invention discloses a method for detecting defects in LED light strips based on deep support vector description, such as Figure 6 As shown, it includes image acquisition module, sample division module, model training module and testing module; wherein,

[0066] Image acquisition module, used to obtain the positive light image of a single LED lamp bead;

[0067] The sample segmentation module is used to perform threshold segmentation on the positive luminescence image and divide it into normal samples and abnormal samples;

[0068] The model training module is used to train the positive luminous image using the Contractive-Deep SVDD neural network, obtain the training model, obtain the minimum hypersphere representation of the original feature space, calculate the hypersphere radius and anomaly score, and divide the threshold;

[0069] The testing module is used to test the training model using normal samples and abnormal samples, evaluate the model performance, and determine whether there are defects in the positive luminous image.

[0070] Furthermore, a pre-training module is included for pre-training the training module using a contraction convolutional autoencoder.

[0071] Compared with the existing technology, the present invention introduces the deep single classification concept in machine vision into the process of LED light strip defect detection. It uses the characteristics of convolutional neural network that can automatically extract features to provide an end-to-end multi-channel LED light strip optical property detection solution. It can simultaneously detect the chromaticity and brightness optical property defects of a single LED lamp bead in the power-on state, and has higher accuracy and robustness.

[0072] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0073] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting defects in LED light strips based on deep support vector description, characterized in that: The specific steps include the following: Get the positive light image of a single LED lamp bead; Performing threshold segmentation on the positive luminescence image to divide it into normal samples and abnormal samples; Performing model training on the positive luminescence image using a Contractive-Deep SVDD neural network to obtain a training model, obtaining a minimum hypersphere representation of the original feature space, calculating the hypersphere radius and anomaly score, and dividing the threshold; Using the normal samples and the abnormal samples to test the training model, evaluate the model performance, and determine whether the positive luminescence image has defects; The convolutional neural network model used is: The convolutional contraction autoencoder used in pre-training uses the LeNet-5 lightweight model. The model structure is that the encoder layer is divided into two convolutional layers and one fully connected layer. The first convolutional layer uses 8 4x4 convolution kernels, and the second convolution kernel uses 4x4 convolution kernels with a convolution stride of 1. The convolution layer is connected to a 2x2 maximum pooling layer and a Leaky ReLU activation function as the output of the next layer. The decoder layer and the encoder layer have a completely symmetrical network structure. By minimizing the reconstruction error, an unbiased representation of the input image and the fully connected output is obtained. The middle layer serves as the encoder output, and the middle layer outputs 32-dimensional vector parameters. The pre-trained encoder model parameters are copied as the parameters of the Contractive-Deep SVDD neural network. The Contractive-Deep SVDD network model used has the following training process: Contractive-Deep SVDD is based on Deep SVDD. By introducing a contraction penalty term, the input dimension is reduced, forcing the deep model to maintain an identical mapping relationship with the input samples. Under the condition of minimizing the Jacobi matrix loss, the reduced-dimensional data is forced to be insensitive to changes in the input gradient, capturing the structure of the sample in the low-dimensional popular space, enhancing the robustness of the model, and projecting the data into the minimum space hypersphere. The optimization function is: The optimization goal is to train the neural network parameters W, and obtain the identity mapping of all normal sample points to the minimum contraction hypersphere by minimizing the loss function J(θ), where W l Indicates the network layer parameters using regular constraints, which is equivalent to W, n is the number of normal lamp bead luminescence samples, x i is a data sample, which is related to the image format and size; c is the center of the hypersphere, which is obtained after pre-training the shrinkage autoencoder, φ(x i ; W) represents the feature vector output by the original sample after dimensionality reduction by the autoencoder, is the Jacobian matrix of the penalty term, μ and λ are the penalty factors of the two regularization parameters respectively; during training, it is assumed that all samples are correctly divided, that is, there is no interference from abnormal light strip samples, and a hard margin minimum hypersphere model is obtained.

2. The LED light strip defect detection method based on deep support vector description according to claim 1 is characterized in that: The positive light images of the LED multi-light strips are collected at equal intervals within a fixed time, and the positive light images of the LED multi-light strips are segmented and the ROI area is extracted with a fixed window size, and the fixed area is divided into positive light images of single LED lamp beads.

3. The LED light strip defect detection method based on deep support vector description according to claim 1 is characterized in that: The method of performing threshold segmentation on the positive luminescence image is as follows: adjusting the size of the positive luminescence image and performing threshold segmentation by image color coordinates and histogram conversion. The calculation formula is: Among them, L* represents the lightness of the color, a* and b* represent the chromaticity of the color, a* positive value represents reddish, negative value represents greenish; b* positive value represents yellowish, negative value represents bluish, X, Y, Z represent the dot product of the RGB three-channel components, f() represents the linear mapping function, X N 、Y N , Z N are the modification coefficients of X, Y, and Z respectively; Convert the original RGB image to L*a*b* color space and apply the color difference formula to calculate: in, Indicates the calculated color difference value, ΔL*, Δa*, and Δb* represent the distribution difference between the current sample and the standard normal sample in the L*a*b* color space; according to Divide the normal lamp bead samples into abnormal samples.

4. The LED light strip defect detection method based on deep support vector description according to claim 1 is characterized in that: The method also includes pre-training the training model using a contraction convolutional autoencoder, extracting encoder module parameters as contractive-deep SVDD neural network parameters, and inputting unlabeled data of normal light strip samples for model training.

5. The LED light strip defect detection method based on deep support vector description according to claim 1 is characterized in that: The performance of the training model is evaluated using an anomaly score evaluation mechanism. The anomaly score evaluation mechanism is established to obtain a score threshold of the distance between the training sample and the center c of the hypersphere. The score of the normal sample farthest from the center c of the hypersphere is used as the radius of the hypersphere to evaluate the abnormal sample. The anomaly score evaluation mechanism is: g(x i )=||φ(x i ;W * )-c|| 2 ; Among them, x i represents the data sample, c is the center of the hypersphere, W * represents the parameters of the deep neural network, and φ represents the training model for the input x i Nonlinear mapping function.

6. The LED light strip defect detection method based on deep support vector description according to claim 1 is characterized in that: The normal samples and abnormal samples are input into the training model to obtain different training abnormality score values g i , the ROC receiver operating characteristic curve is drawn by the abnormal score and the AUC value is calculated to test the detection performance and distinguish normal luminous lamp beads from abnormal luminous lamp beads.

7. A LED light strip defect detection system based on deep support vector description, characterized in that: The system is used to implement a method for detecting defects in LED light strips based on deep support vector description as described in any one of claims 1 to 6, comprising an image acquisition module, a sample division module, a model training module, and a testing module; wherein, The image acquisition module is used to acquire the positive light emitting image of a single LED lamp bead; The sample segmentation module is used to perform threshold segmentation on the positive luminescence image to divide it into normal samples and abnormal samples; The model training module is used to perform model training on the positive luminescence image through a Contractive-Deep SVDD neural network to obtain a training model, obtain a minimum hypersphere representation of the original feature space, calculate the hypersphere radius and anomaly score, and divide the threshold; The testing module is used to test the training model using the normal samples and abnormal samples, evaluate the model performance, and determine whether the positive luminescence image has defects.

8. The LED light strip defect detection system based on deep support vector description according to claim 7 is characterized in that: It also includes a pre-training module for pre-training the training module using a contraction convolutional autoencoder.

Images