Cable appearance defect detection method

Abstract

The invention relates to a cable appearance defect detection method, and belongs to the technical field of image analysis. According to the method, a historical cable surface 3D point cloud is acquired and preprocessed, then the cable surface 3D point cloud is converted into a 2D depth image, the 2D depth image is segmented into a plurality of image segments, and each image segment is labeled with a defect type label. Secondly, equally dividing all the images into two types of sets of a training set and a verification set, inputting the two types of sets into a convolutional auto-encoder to train feature extraction capability, and inputting the two types of sets into four prediction models to train prediction capability; and finally, calculating the fusion entropy of the deep features without appearance defects of the cable and the fusion entropy of the deep features of the to-be-detected cable, determining which most suitable prediction model is input for prediction according to the difference value of the two fusion entropies, and taking the obtained appearance defect type with the maximum probability as the appearance defect type of the to-be-detected cable.

Description

Technical Field

[0001] The invention relates to a method for detecting appearance defects of a cable, and belongs to the technical field of image analysis (G06T7 / 00). Background Art

[0002] Cables play a key role in the construction of infrastructure such as electricity and communications. However, during the laying process, scratches, cracks, uneven surfaces, and uneven appearance defects often appear on the cable surface due to external factors such as mechanical stretching, friction, and pressure. Therefore, cable appearance defect detection has become an indispensable quality control method in the cable laying process.

[0003] Traditional cable appearance inspection mainly relies on artificial vision. However, this method is inefficient, inconsistent, and easily affected by operator fatigue and subjective judgment, resulting in unstable and unreliable inspection results. With the advancement of computer vision technology, cable appearance defect detection technology based on 2D images has gradually been widely used. This technology captures cable surface images through a CCD camera and identifies appearance defects through image processing such as edge detection and contrast enhancement. Although 2D image detection technology improves detection efficiency, it is still difficult to accurately identify millimeter-level defect types and lacks cable depth information. In addition, the technology has poor adaptability and needs to readjust threshold parameters when detecting cables of different types or colors.

[0004] In view of the limitations of existing technologies, there is an urgent need to develop a cable appearance defect detection technology to achieve high-precision detection of cable appearance defects. Summary of the invention

[0005] The technical problem to be solved by the present invention is: how to achieve high-precision detection of cable appearance defects.

[0006] The present invention proposes a technical solution to solve the above technical problems: a method for detecting cable appearance defects, comprising the following steps:

[0007] Step 1: Take a historical cable, scan the surface of the historical cable to obtain the 3D point cloud of the historical cable surface, and remove outliers and reduce noise from the 3D point cloud of the historical cable surface;

[0008] Step 2: converting the 3D point cloud of the historical cable surface into a 2D depth image of the historical cable surface;

[0009] Step 3: Sequentially segment the historical cable surface 2D depth image to obtain multiple image segments, and annotate each image segment with a label value belonging to a certain historical appearance defect type;

[0010] The appearance defect types include five types: no defect, crack, scratch, bulge and dent;

[0011] All image segments are equally divided into two sets: training set and validation set. Both the training set and validation set contain image segments of five historical appearance defect types.

[0012] Step 4: Input each image segment of the training set and the validation set into the convolutional autoencoder for training;

[0013] Step 4.1: The convolutional autoencoder extracts edge, texture, and shape features from each image segment in the training set according to its first initial hyperparameters to generate the first reconstructed image X of the same length and width for each image segment. rec ;

[0014] Step 4.2: Calculate the mean square error loss between each image segment in the training set and the first reconstructed image generated by each image segment according to formula (1): , edge preservation loss , similarity difference loss And the loss function L1

[0015] (1);

[0016] In formula (1), X ij is the depth value of each image fragment plane coordinate i, j in the training set; X rec，ij is the depth value at the coordinates i and j of each first reconstructed image plane; O is the height of each first reconstructed image and each image segment in the training set; P is the width of each first reconstructed image and each image segment in the training set; ∇ is the gradient operator;

[0017] The training set loss function L2 is calculated according to formula (2):

[0018] (2);

[0019] In formula (2), k1 is the number of image segments in the training set;

[0020] Step 4.3: The convolutional autoencoder extracts edge, texture and shape features from each image segment in the validation set according to its own first initial hyperparameters to generate a second reconstructed image of each image segment; the validation set loss function L3 is calculated according to equations (1) and (2);

[0021] The convolutional autoencoder continuously iterates and updates its first initial hyperparameters until the obtained training set loss function L2 value and validation set loss function L3 value no longer decrease, and the convolutional autoencoder training is completed;

[0022] Step 5: input each image segment of the training set and the validation set into an artificial neural network ANN for training;

[0023] Step 5.1: The artificial neural network ANN predicts the probability of each appearance defect type for each image segment in the training set according to its second initial hyperparameter;

[0024] The loss function L4 of the historical appearance defect type and the predicted probability of the appearance defect type for each image segment in the training set is calculated according to formula (3):

[0025] (3);

[0026] In formula (3), k2 is the number of appearance defect types; U t is the label value of the t-th historical appearance defect type; P t is the predicted probability of the tth historical appearance defect type;

[0027] The training set prediction loss function L5 is calculated according to formula (4):

[0028] (4);

[0029] Step 5.3: The artificial neural network ANN predicts the probability of each appearance defect type for each image segment in the validation set based on its own second initial hyperparameters, and calculates the validation set prediction loss function L6 according to equations (3) and (4);

[0030] The artificial neural network ANN continuously iterates and updates its second initial hyperparameter until the obtained training set prediction loss function L5 value and validation set prediction loss function L6 value no longer decrease, and the artificial neural network ANN training is completed;

[0031] Step 6: According to step 5, each image segment of the training set and the validation set is input into the support vector machine SVM, the extreme learning machine ELM, and the long short-term memory network LSTM to complete the training;

[0032] Step 7: Select an image segment without appearance defects in step 3 and input it into the convolutional autoencoder trained in step 3 to extract the deep features without appearance defects and obtain the deep feature sequence without appearance defects. ;

[0033] h t It is the deep feature without appearance defects at time point t;

[0034] Step 7.1: Construct the phase space vector for each time point t according to the deep feature sequence without appearance defects ;

[0035] τ is the delay time; m is the embedding dimension;

[0036] Step 7.1.1: According to each phase space vector H t1The order of arrangement of deep features without appearance defects in each phase space vector H t1 The sequential pattern of

[0037] Step 7.1.2: Count the probability P(π) of each sequential pattern appearing in the π sequential patterns, and calculate the permutation entropy P of the deep features without appearance defects according to the following formula (5);

[0038] (5);

[0039] Step 7.2: Construct the phase space vector for each time point t according to the deep feature sequence without appearance defects ;

[0040] z is the pattern dimension;

[0041] Step 7.2.1: Set the distance threshold r, the value range of which is 15%~20% of the standard deviation of the deep feature sequence data without appearance defects;

[0042] According to each phase space vector H t2 The distance to all other phase space vectors, counting the proportion of the number of distances within the distance threshold r to the total number of distances ;

[0043] The expanded mode dimension is z+1, according to each phase space vector H t2 The distance to all other phase space vectors, counting the proportion of the number of distances within the distance threshold r to the total number of distances ;

[0044] Step 7.2.2: Calculate the sample entropy S of the deep features without appearance defects according to formula (6);

[0045] (6);

[0046] Step 7.3: Calculate the fusion entropy F1 of the deep features without appearance defects according to formula (7)

[0047] (7);

[0048] In formula (7), a and b are weight coefficients, a=0.4, b=0.6;

[0049] Step 8: Take a cable to be tested, scan the surface of the cable to be tested to obtain a 3D point cloud of the cable surface to be tested, and remove outliers and reduce noise from the 3D point cloud of the cable surface to be tested;

[0050] Convert the 3D point cloud of the cable surface to be tested into a 2D depth image of the cable surface to be tested and input the image into the convolutional autoencoder trained in step 3 to extract the deep features to be tested and obtain the deep feature sequence to be tested;

[0051] Calculate the fusion entropy F2 of the deep feature to be measured according to steps 6.1 to 6.3;

[0052] like , input the deep features to be tested into the artificial neural network ANN trained in step 6, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested;

[0053] like , input the deep features to be tested into the support vector machine SVM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested;

[0054] like , input the deep features to be tested into the extreme learning machine ELM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested;

[0055] like , input the deep features to be tested into the long short-term memory network LSTM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested.

[0056] Furthermore, the step 1 uses the Z-score statistical method to remove outliers and uses Gaussian filtering to eliminate noise data.

[0057] The beneficial effects of the present invention are as follows: since the 3D point cloud of the historical cable surface is converted into a 2D depth image of the historical cable surface, the 2D depth image of the historical cable surface is segmented into multiple image segments, and the historical appearance defect type is labeled for each image segment, all image segments are equally divided into a training set and a validation set to input the convolutional autoencoder to train the feature extraction capability, it is helpful to identify and classify various subtle defects on the cable surface and improve the accuracy of the subsequent prediction model appearance defect type. In addition, since the permutation entropy and the sample are integrated into a fusion entropy, the structure and dynamic characteristics of the data are more carefully characterized, and the complexity and changes that cannot be described by a single entropy are captured, thereby improving the comprehensiveness of the feature description. By calculating the fusion entropy of the deep features of the cable without appearance defects and the fusion entropy of the deep features of the cable to be tested, the difference between the two fusion entropies is used to determine which is the most appropriate prediction model to input to achieve high-precision detection of cable appearance defects, while improving the robustness and adaptability of cable appearance defect detection. DETAILED DESCRIPTION

[0058] Example

[0059] A method for detecting cable appearance defects in this embodiment includes the following steps:

[0060] Step 1: Take a historical cable, scan the surface of the historical cable to obtain the 3D point cloud of the historical cable surface, and remove outliers and reduce noise from the 3D point cloud of the historical cable surface;

[0061] This embodiment uses the existing Z-score statistical method to remove outliers and uses Gaussian filtering to eliminate noise data.

[0062] Step 2: Convert the 3D point cloud of the historical cable surface into a 2D depth image of the historical cable surface;

[0063] This embodiment orthogonally projects the 3D point cloud of the historical cable surface onto a set plane to generate a 2D depth image of the historical cable surface. The grayscale value of each point on the 2D depth image of the historical cable surface is the distance from each point in the 3D point cloud of the historical cable surface to the plane.

[0064] Step 3: Sequence segmentation is performed on the historical cable surface 2D depth image to obtain multiple image segments, and each image segment is labeled with a label value belonging to a certain historical appearance defect type;

[0065] The appearance defect types include five types: no defect, crack, scratch, bulge and dent;

[0066] All image segments are equally divided into two sets: training set and validation set. Both the training set and validation set contain image segments of five historical appearance defect types.

[0067] Step 4: Input each image segment of the training set and the validation set into the convolutional autoencoder for training;

[0068] Convolutional autoencoder is a model that combines convolutional neural network (CNN) and autoencoder. It belongs to the existing technology and is used for feature extraction and data dimensionality reduction in unsupervised learning. It is mainly composed of encoder and decoder. The encoder gradually compresses the input data (image fragments) into low-dimensional feature representation through convolutional layers and pooling layers. The convolutional layer can effectively extract features from the input data, and the pooling layer reduces the spatial dimension of the data, thereby reducing the amount of calculation. The decoder restores the low-dimensional feature representation to the dimension of the original input data through the deconvolution layer. The goal of the decoder is to reconstruct the input data as accurately as possible;

[0069] Step 4.1: The convolutional autoencoder extracts edge, texture, and shape features from each image segment in the training set according to its first initial hyperparameters to generate the first reconstructed image X of the same length and width for each image segment. rec ;

[0070] Step 4.2: Calculate the mean square error loss between each image segment in the training set and the first reconstructed image generated by each image segment according to formula (1): , edge preservation loss , similarity difference loss And the loss function L1

[0071] (1);

[0072] In formula (1), X ij is the depth value of each image fragment plane coordinate i, j in the training set; X rec，ij is the depth value at the coordinates i and j of each first reconstructed image plane; O is the height of each first reconstructed image and each image segment in the training set; P is the width of each first reconstructed image and each image segment in the training set; ∇ is the gradient operator;

[0073] The training set loss function L2 is calculated according to formula (2):

[0074] (2);

[0075] In formula (2), k1 is the number of image segments in the training set;

[0076] Step 4.3: The convolutional autoencoder extracts edge, texture and shape features from each image segment in the validation set according to its own first initial hyperparameters to generate a second reconstructed image of each image segment; the validation set loss function L3 is calculated according to equations (1) and (2);

[0077] The convolutional autoencoder continuously iterates and updates its first initial hyperparameters until the obtained training set loss function L2 value and validation set loss function L3 value no longer decrease, and the convolutional autoencoder training is completed;

[0078] Step 5: Input each image segment of the training set and the validation set into the artificial neural network ANN for training;

[0079] Step 5.1: The artificial neural network ANN predicts the probability of each appearance defect type for each image segment in the training set according to its second initial hyperparameter;

[0080] Artificial neural network (ANN) is a computational model that simulates the connection mode of human brain neurons and belongs to the existing technology. ANN can learn complex feature representation by recognizing features and patterns through multiple layers of neural networks. Its structure usually includes an input layer, several hidden layers and an output layer. In the process of processing image fragments, ANN can integrate the feature information of the image and map these features to the probability distribution of each type of appearance defect through the output layer, thereby providing the predicted probability of each defect type. This probability prediction mechanism also exists in other existing technology models, such as support vector machine (SVM), extreme learning machine (ELM) and long short-term memory network (LSTM). SVM is good at processing high-dimensional features, achieving classification by maximizing the classification interval, and can be converted into probability output through post-processing; ELM has the characteristics of fast learning, generating probability distribution by randomly initializing weights and quickly learning the weights of the output layer; and LSTM is good at capturing the dependency of time series, memorizing time series features through a gating mechanism, and finally outputting the predicted probability distribution. These characteristics make them have their own advantages and disadvantages in the image fragment processing task.

[0081] The loss function L4 of the historical appearance defect type and the predicted probability of the appearance defect type for each image segment in the training set is calculated according to formula (3):

[0082] (3);

[0083] In formula (3), k2 is the number of appearance defect types; U t is the label value of the t-th historical appearance defect type; P t is the predicted probability of the tth historical appearance defect type;

[0084] The training set prediction loss function L5 is calculated according to formula (4):

[0085] (4);

[0086] Step 5.3: The artificial neural network ANN predicts the probability of each appearance defect type for each image segment in the validation set based on its own second initial hyperparameters, and calculates the validation set prediction loss function L6 according to equations (3) and (4);

[0087] The artificial neural network ANN continuously iterates and updates its second initial hyperparameter until the obtained training set prediction loss function L5 value and validation set prediction loss function L6 value no longer decrease, and the artificial neural network ANN training is completed;

[0088] Step 6: According to step 5, each image segment of the training set and the validation set is input into the support vector machine SVM, the extreme learning machine ELM, and the long short-term memory network LSTM to complete the training;

[0089] Step 7: Select an image segment without appearance defects in step 3 and input it into the convolutional autoencoder trained in step 3 to extract the deep features without appearance defects and obtain the deep feature sequence without appearance defects. ;

[0090] h t It is the deep feature without appearance defects at time point t;

[0091] Step 7.1: Construct the phase space vector for each time point t according to the deep feature sequence without appearance defects ;

[0092] τ is the delay time; m is the embedding dimension;

[0093] Step 7.1.1: According to each phase space vector H t1 The order of arrangement of deep features without appearance defects in each phase space vector H t1 The sequential pattern of

[0094] Step 7.1.2: Count the probability P(π) of each sequential pattern appearing in the π sequential patterns, and calculate the permutation entropy P of the deep features without appearance defects according to the following formula (5);

[0095] (5);

[0096] In this embodiment, there are 5 sequential modes, and the probability P(π) of each sequential mode appearing is 0.2, 0.15, 0.25, 0.2, 0.2 respectively;

[0097] According to formula (5), P = 0.1662;

[0098] Step 7.2: Construct the phase space vector for each time point t according to the deep feature sequence without appearance defects ;

[0099] z is the pattern dimension;

[0100] Step 7.2.1: Set the distance threshold r, the value range of which is 15%~20% of the standard deviation of the deep feature sequence data without appearance defects;

[0101] In this embodiment, the distance threshold r=15;

[0102] According to each phase space vector H t2 The distance to all other phase space vectors, counting the proportion of the number of distances within the distance threshold r to the total number of distances =0.3;

[0103] The expanded mode dimension is z+1, according to each phase space vector H t2The distance to all other phase space vectors, counting the proportion of the number of distances within the distance threshold r to the total number of distances =0.25;

[0104] Step 7.2.2: Calculate the sample entropy S of the deep features without appearance defects according to formula (6);

[0105] (6);

[0106] In this embodiment, S=0.34 is calculated according to formula (6);

[0107] Step 7.3: Calculate the fusion entropy F1 of the deep features without appearance defects according to formula (7)

[0108] (7);

[0109] In formula (7), a and b are weight coefficients, a=0.4, b=0.6;

[0110] In this embodiment, F1=0.83 is calculated according to formula (7);

[0111] Step 8: Take a cable to be tested, scan the surface of the cable to be tested to obtain a 3D point cloud of the cable surface to be tested, and remove outliers and reduce noise from the 3D point cloud of the cable surface to be tested;

[0112] Convert the 3D point cloud of the cable surface to be tested into a 2D depth image of the cable surface to be tested and input the image into the convolutional autoencoder trained in step 3 to extract the deep features to be tested and obtain the deep feature sequence to be tested;

[0113] According to steps 6.1 to 6.3, calculate the fusion entropy F2=0.9 of the deep features to be tested;

[0114] like , input the deep features to be tested into the artificial neural network ANN trained in step 6, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested;

[0115] like , input the deep features to be tested into the support vector machine SVM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested;

[0116] like , input the deep features to be tested into the extreme learning machine ELM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested;

[0117] like , input the deep features to be tested into the long short-term memory network LSTM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested.

[0118] This embodiment 0.07, input the deep features to be tested into the support vector machine SVM, and the appearance defect type with the highest probability is "crack", so the appearance defect type of the cable to be tested is "crack";

[0119] The above description is only a preferred embodiment of the present invention, but the present invention is not limited thereto. All equivalent replacements or equivalent changes of the concepts and technical solutions according to the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting cable appearance defects, characterized in that The steps include: Step 1: Take a historical cable, scan the surface of the historical cable to obtain the 3D point cloud of the historical cable surface, and remove outliers and reduce noise from the 3D point cloud of the historical cable surface; Step 2: converting the 3D point cloud of the historical cable surface into a 2D depth image of the historical cable surface; Step 3: Sequentially segment the historical cable surface 2D depth image to obtain multiple image segments, and annotate each image segment with a label value belonging to a certain historical appearance defect type; The appearance defect types include five types: no defect, crack, scratch, bulge and dent; All image segments are equally divided into two sets: training set and validation set. Both the training set and validation set contain image segments of five historical appearance defect types. Step 4: Input each image segment of the training set and the validation set into the convolutional autoencoder for training; Step 4.1: The convolutional autoencoder extracts edge, texture, and shape features from each image segment in the training set according to its first initial hyperparameters to generate the first reconstructed image X of the same length and width for each image segment. rec ; Step 4.2: Calculate the mean square error loss between each image segment in the training set and the first reconstructed image generated by each image segment according to formula (1): , edge preservation loss , similarity difference loss And the loss function L1 （1）； In formula (1), X ij is the depth value of each image fragment plane coordinate i, j in the training set; X rec，ij is the depth value at the coordinates i and j of each first reconstructed image plane; O is the height of each first reconstructed image and each image segment in the training set; P is the width of each first reconstructed image and each image segment in the training set; ∇ is the gradient operator; The training set loss function L2 is calculated according to formula (2): （2）； In formula (2), k1 is the number of image segments in the training set; Step 4.3: The convolutional autoencoder extracts edge, texture and shape features from each image segment in the validation set according to its own first initial hyperparameters to generate a second reconstructed image of each image segment; the validation set loss function L3 is calculated according to equations (1) and (2); The convolutional autoencoder continuously iterates and updates its first initial hyperparameters until the obtained training set loss function L2 value and validation set loss function L3 value no longer decrease, and the convolutional autoencoder training is completed; Step 5: input each image segment of the training set and the validation set into an artificial neural network ANN for training; Step 5.1: The artificial neural network ANN predicts the probability of each appearance defect type for each image segment in the training set according to its second initial hyperparameter; The loss function L4 of the historical appearance defect type and the predicted probability of the appearance defect type for each image segment in the training set is calculated according to formula (3): （3）； In formula (3), k2 is the number of appearance defect types; U t is the label value of the t-th historical appearance defect type; P t is the predicted probability of the tth historical appearance defect type; The training set prediction loss function L5 is calculated according to formula (4): （4）； Step 5.3: The artificial neural network ANN predicts the probability of each appearance defect type for each image segment in the validation set based on its own second initial hyperparameters, and calculates the validation set prediction loss function L6 according to equations (3) and (4); The artificial neural network ANN continuously iterates and updates its second initial hyperparameter until the obtained training set prediction loss function L5 value and validation set prediction loss function L6 value no longer decrease, and the artificial neural network ANN training is completed; Step 6: According to step 5, each image segment of the training set and the validation set is input into the support vector machine SVM, the extreme learning machine ELM, and the long short-term memory network LSTM to complete the training; Step 7: Select an image segment without appearance defects in step 3 and input it into the convolutional autoencoder trained in step 3 to extract the deep features without appearance defects and obtain the deep feature sequence without appearance defects. ; h t It is the deep feature without appearance defects at time point t; Step 7.1: Construct the phase space vector for each time point t according to the deep feature sequence without appearance defects ; τ is the delay time; m is the embedding dimension; Step 7.1.1: According to each phase space vector H t1 The order of arrangement of deep features without appearance defects in each phase space vector H t1 The sequential pattern of Step 7.1.2: Count the probability P(π) of each sequential pattern appearing in the π sequential patterns, and calculate the permutation entropy P of the deep features without appearance defects according to the following formula (5); （5）； Step 7.2: Construct the phase space vector for each time point t according to the deep feature sequence without appearance defects ; z is the pattern dimension; Step 7.2.1: Set the distance threshold r, the value range of which is 15%~20% of the standard deviation of the deep feature sequence data without appearance defects; According to each phase space vector H t2 The distance to all other phase space vectors, counting the proportion of the number of distances within the distance threshold r to the total number of distances ; The expanded mode dimension is z+1, according to each phase space vector H t2 The distance to all other phase space vectors, counting the proportion of the number of distances within the distance threshold r to the total number of distances ; Step 7.2.2: Calculate the sample entropy S of the deep features without appearance defects according to formula (6); （6）； Step 7.3: Calculate the fusion entropy F1 of the deep features without appearance defects according to formula (7) （7）； In formula (7), a and b are weight coefficients, a=0.4, b=0.6; Step 8: Take a cable to be tested, scan the surface of the cable to be tested to obtain a 3D point cloud of the cable surface to be tested, and remove outliers and reduce noise from the 3D point cloud of the cable surface to be tested; Convert the 3D point cloud of the cable surface to be tested into a 2D depth image of the cable surface to be tested and input the image into the convolutional autoencoder trained in step 3 to extract the deep features to be tested and obtain the deep feature sequence to be tested; Calculate the fusion entropy F2 of the deep feature to be measured according to steps 6.1 to 6.3; like , input the deep features to be tested into the artificial neural network ANN trained in step 6, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested; like , input the deep features to be tested into the support vector machine SVM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested; like , input the deep features to be tested into the extreme learning machine ELM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested; like , input the deep features to be tested into the long short-term memory network LSTM trained in step 7, and use the appearance defect type with the highest probability as the appearance defect type of the cable to be tested.

2. The method for detecting cable appearance defects according to claim 1, characterized in that: The step 1 uses the Z-score statistical method to remove outliers and uses Gaussian filtering to eliminate noise data.

Images