CN116894836A - Yarn defect detection method and device based on machine vision - Google Patents

Abstract

The application discloses a yarn flaw detection method and device based on machine vision, wherein the method comprises the steps of obtaining a yarn image; performing binarization processing on the image of the yarn to obtain a binarized image of the yarn; performing dimension reduction treatment on the binarized image to obtain a one-dimensional sequence of the yarn width value; carrying out average value pooling on the one-dimensional sequence to obtain a characteristic sequence; constructing an optimal feature vector based on the feature sequence; inputting the optimal feature vector into a deep learning neural network, and outputting the feature attribute of the yarn. The application can realize automatic flaw detection in the yarn production process, solves the problems that the detection precision of the traditional yarn flaw detection method is easy to be influenced, has low automation level and is easy to be subjected to false detection, and has the advantages of low occupancy rate of computing resources, high detection speed and high recognition precision.

Description

Yarn defect detection method and device based on machine vision

Technical field

This application belongs to the field of textile computer detection technology, and specifically relates to a yarn defect detection method and device based on machine vision.

Background technique

During the yarn production process, affected by mechanical transmission equipment, spinning raw materials, etc., yarn defects such as neps and details will inevitably appear. Therefore, detecting defects in yarn during the production process is of great research significance for controlling yarn production quality and improving yarn production efficiency.

At present, the commonly used yarn defect detection methods mainly include: photoelectric detection method, capacitive detection method and manual visual inspection method.

The photoelectric detection method generally consists of three parts: a luminous body, an optical system, and a light receiver. After the infrared light generated by the luminous body passes through the optical system, a detection area with a uniform light field is formed, and the light receiver converts the light energy into an analog output. When the yarn runs in the detection area, it blocks part of the light, reducing the light energy received by the photoreceptor. The change in light energy reflects the diameter of the yarn in the detection area. The photoreceptor converts light energy into electrical signals, the amplitude of which corresponds to the diameter of the yarn. The detection accuracy of the photoelectric detection method is easily affected by the aging of the photoelectric device and the light transmission of yarn hairiness.

The capacitive detection method consists of two vertical capacitive plates. When the yarn passes through the detection area formed by the two plates in a non-contact manner, the dielectric constant between the two plates changes due to the addition of fiber medium, resulting in The capacitance value of the plate changes, which indirectly reflects the changes in yarn quality. This method can detect ordinary yarns, but the detection accuracy is easily affected by air humidity, yarn moisture content, spatial electric field unevenness, etc. in the environment.

The manual visual inspection method uses inspectors to randomly inspect the same batch of yarn and rely on experience and naked eye observation to obtain the type and quantity of yarn defects. This method has a low level of automation, the detection results rely on the subjectivity of the operator, and are prone to false detections.

Contents of the invention

The purpose of this application is to provide a yarn defect detection method and device based on machine vision to solve the problems of yarn defect detection methods in the prior art that are susceptible to detection accuracy, low automation level, and prone to misdetection. question.

In order to achieve the above purpose, a technical solution adopted in this application is:

A yarn defect detection method based on machine vision is provided, including:

Get an image of yarn;

Perform binarization processing on the image of the yarn to obtain a binary image of the yarn;

Perform dimensionality reduction processing on the binarized image to obtain a one-dimensional sequence of yarn width values;

Perform average pooling on the one-dimensional sequence to obtain a feature sequence, wherein the pooling size of the average pooling is P _i =f(a _i ), a _i =1, 2, 3...N;

Based on the feature sequence, construct an optimal feature vector;

The optimal feature vector is input into a deep learning neural network and the characteristic attributes of the yarn are output, where the characteristic attributes include normal or the type of yarn defect.

In one or more embodiments, the step of performing binarization processing on the image of the yarn to obtain the binarized image of the yarn includes:

Calculate the OSTU threshold T _OSTU of the image of the yarn based on the OSTU threshold method;

The OSTU threshold is corrected based on the following formula to obtain the corrected threshold T _{g-global-OSTU} :

In the formula, V _q is the gray value of the pixel with the largest number of the same gray value among all the pixels in the image of the yarn, and Z _x is the smallest gray value among all the pixels in the image of the yarn. The gray value of the pixel;

The image is binarized based on the correction threshold to obtain a binarized image of the yarn.

In one or more embodiments, the step of performing dimensionality reduction processing on the binarized image to obtain a one-dimensional sequence of yarn width values includes:

Taking the background pixel of the binary image as value 1 and the foreground pixel as value 0, the matrix X _n*m of the binary image is obtained, where n is the width direction of the yarn in the binary image The number of pixels on, m is the number of pixels of the binary image in the length direction of the yarn;

Based on the following formula, the one-dimensional sequence D of the yarn width value is calculated:

D＝[nn...n] _1*m -[1 1...1] _1*n *X _n*m .

In one or more embodiments, the yarn defects include detail defects and long gap defects, and the step of performing average pooling on the one-dimensional sequence to obtain a feature sequence includes:

Perform average pooling on the one-dimensional sequence, and the pooling size Obtain the characteristic sequence S ₁ (a ₁ ), where a ₁ =1, 2, 3...N ₁ , m is the length of the one-dimensional sequence, />

In one or more embodiments, the yarn defects include slub defects, and the step of performing average pooling on the one-dimensional sequence to obtain the feature sequence includes:

Perform average pooling on the one-dimensional sequence, and the pooling size Obtain the characteristic sequence S ₂ (a ₂ ), where a ₂ =1, 2, 3...N ₂ , m is the length of the one-dimensional sequence, />

In one or more embodiments, the yarn defects include missing knot defects and decorative yarn staggered defects. The step of performing average pooling on the one-dimensional sequence to obtain the feature sequence includes:

Perform average pooling twice on the one-dimensional sequence. The pooling sizes of the two average poolings are P _3-1 =4a ₃ +2 and P _3-2 =2a ₃ +1 respectively, and the sequence M is obtained. _short and M _long , where a ₃ =1, 2, 3...N ₃ , m is the length of the one-dimensional sequence.

Perform backward search and difference and forward search and difference on the sequences M _short and M _long respectively to obtain a backward difference sequence and a forward difference sequence;

Compare the backward difference sequence and the forward difference sequence, and take the larger value of each corresponding position of the backward difference sequence and the forward difference sequence to construct the feature sequence S ₃ (a ₃ ).

In one or more embodiments, the steps of performing backward search and difference and forward search and difference on the sequences M _short and M _long respectively, and obtaining the backward difference sequence and the forward difference sequence include:

The first a ₃ values at the head end of the sequence M _short are removed, and the sequences M _short and M _long are sequentially compared from the head end to the tail end to obtain the backward difference sequence S _3-1 =|M _short [a ₃ +1:m+1-4a ₃ ]-M _long [1:m-1-5a ₃ ]|, where m is the length of the one-dimensional sequence;

Remove the last ₃ values of the tail end of the sequence M _short , and compare the sequences M _short and M _long in sequence from the tail end to the head end, and obtain the forward difference sequence S _3-2 =|M _short [1 :m-1-5a ₃ ]-M _long [3a ₃ +2:m-2a ₃ ]|, where m is the length of the one-dimensional sequence.

In one or more embodiments, the step of constructing an optimal feature vector based on the feature sequence includes:

The stochastic leapfrog algorithm based on partial least squares calculates the optimal solution of a _i ;

Based on the optimal solution, the maximum value, minimum value and average value of the feature sequence are calculated, and the optimal feature vector is collected.

In one or more embodiments, in the step of inputting the optimal feature vector into a deep learning neural network and outputting the characteristic attributes of the yarn, the neural network is a two-layer ANN classifier, and the Training methods for deep learning neural networks include:

Obtain a sample training set, the sample training set includes the optimal feature vectors of several groups of yarn images marked with label values;

The sample training set is input, and based on the cross-entropy loss function and the label value, the weight of the deep learning neural network is updated in the direction of gradient descent until the cross-entropy loss function converges, and the deep learning neural network is obtained .

In order to achieve the above purpose, another technical solution adopted by this application is:

A yarn defect detection device based on machine vision is provided, including:

Get module, used to get the image of yarn;

A binarization processing module is used to perform binarization processing on the image of the yarn to obtain a binary image of the yarn;

A dimensionality reduction processing module, used to perform dimensionality reduction processing on the binary image to obtain a one-dimensional sequence of yarn width values;

A mean pooling module, used to perform mean pooling on the one-dimensional sequence to obtain a feature sequence, wherein the pooling size of the mean pooling is Pi ₌ f (a _i ), a _i = 1, 2 ,3...N;

A building module for constructing an optimal feature vector based on the feature sequence;

A classification output module, configured to input the optimal feature sequence into a deep learning neural network and output characteristic attributes of the yarn, where the characteristic attributes include normal or types of yarn defects.

Different from the existing technology, the beneficial effects of this application are:

The yarn defect detection method of this application is based on machine vision. It collects yarn images and performs binarization processing and then reduces the dimension to obtain a one-dimensional sequence of yarn width values. The one-dimensional sequence is processed with different pooling sizes. Average pooling is used to obtain a feature sequence that can highlight different defects. Based on the feature sequence, the optimal feature vector can be generated and brought into the deep learning neural network to automatically output the yarn defect type, which can realize automated defect detection in the yarn production process. Detection solves the problem that traditional yarn defect detection methods are susceptible to detection accuracy, low automation level, and prone to false detections. It also has the advantages of low computing resource usage, fast detection speed, and high recognition accuracy.

Description of the drawings

Figure 1 is a schematic flow chart of an embodiment of the yarn defect detection method based on machine vision of the present application;

Figure 2 is a diagram of the defect types of chenille yarn;

Figure 3 is a schematic diagram of an implementation of backward search in this application;

Figure 4 is a schematic diagram of an implementation of backward search in this application;

Figure 5 is a characteristic sequence diagram of yarns with different defect types in this application;

Figure 6 is a schematic structural diagram of an embodiment of the deep learning neural network of this application;

Figure 7 is a schematic structural diagram of an embodiment of the machine vision-based yarn defect detection device of the present application;

FIG. 8 is a hardware structure diagram of an embodiment of the electronic device of the present application.

Detailed ways

The present application will be described in detail below with reference to each embodiment shown in the accompanying drawings. However, these embodiments do not limit this application, and any structural, method, or functional changes made by those of ordinary skill in the art based on these embodiments are included in the protection scope of this application.

As mentioned in the background art, current yarn defect detection methods have major limitations, making it difficult to effectively control yarn production quality and efficiency.

To this end, the applicant has developed a yarn defect detection method based on machine vision. This method can process the image by collecting images of the yarn and automatically identify multiple defects in the yarn. It has fast detection speed and high recognition accuracy. , has low computing resource usage and can be widely used in defect detection of various yarns.

Specifically, please refer to Figure 1, which is a schematic flow chart of an embodiment of the machine vision-based yarn defect detection method of the present application.

Detection methods include:

S100. Obtain the image of yarn.

First, images of the yarn can be obtained during the production process. Among them, the yarn can be in a moving state or a static state.

In one embodiment, in order to ensure the collection effect of the yarn image, the collection can be carried out under backlight illumination conditions.

S200: Perform binarization processing on the yarn image to obtain a binary image of the yarn.

In order to accurately segment the range of yarns in the image, the image can be binarized. In one embodiment, the threshold method can be used to segment the yarn range. In other embodiments, other methods can also be used to segment the yarn range, such as using a deep learning neural network to segment the yarn range.

Specifically, in one embodiment, the OSTU threshold T _OSTU of the yarn image can be first calculated based on the OSTU threshold method, and then the OSTU threshold is corrected to improve the segmentation accuracy.

Among them, the OSTU threshold method is a conventional threshold segmentation method in this field, and the calculation process of the OSTU threshold will not be described in detail here.

The formula for correcting the OSTU threshold can be as follows:

In the formula, V _q is the gray value of the pixel with the largest number of the same gray value among all the pixels in the yarn image, and Z _x is the gray value of the pixel with the smallest gray value among all the pixels in the yarn image. degree value.

The collected yarn image can then be binarized based on the correction threshold T _{g-global-OSTU} , that is, the gray level of the pixels whose gray level is less than the correction threshold is set to 0, and the gray level of the pixels whose gray level is greater than the correction threshold is set to 0. The grayscale is set to 255, resulting in a binary image.

S300: Perform dimensionality reduction processing on the binary image to obtain a one-dimensional sequence of yarn width values.

It can be understood that the binary image includes the image of the yarn and the background. Based on the binary image, the width of any place in the length direction of the yarn can be calculated.

The specific calculation method can be as follows: taking the background pixel of the binary image as value 1 and the foreground pixel as value 0, thereby obtaining the matrix X _n*m of the binary image, where n is the value of the binary image in the yarn width direction. The number of pixels, m is the number of pixels of the binary image in the length direction of the yarn.

Based on the formula D1=[1 1 ... 1] _1*n *X _n*m , the number of background pixels in the yarn width direction of the yarn image can be calculated. Then the total number of pixels n in the yarn width direction of the yarn image can be calculated to obtain the yarn width value. Specifically, the one-dimensional sequence D of the yarn width value can be calculated by referring to the following formula:

D＝[nn...n] _1*m -[1 1...1] _1*n *X _n*m .

S400: Perform average pooling on the one-dimensional sequence to obtain the feature sequence.

Among them, the pooling size of average pooling is P _i =f(a _i ), a _i =1, 2, 3...N.

After obtaining the one-dimensional sequence D of the yarn width index, the one-dimensional sequence can be average pooled to reflect the changing trend of the yarn width in the length direction. Based on this changing trend, possible defects in the yarn can be reflected question.

In one application scenario, taking chenille yarn as an example, chenille yarn may have five major types of defects during the production process: detail defects, long missing joint defects, thick joint defects, short joint defects and interlacing decorative yarn defects. Specifically, reference can be made to Figure 2, which is a diagram of defect types of chenille yarn.

As shown in Figure 2, the changing trends of width values in the length direction of chenille yarns with different defect types are different. Among them, detail defects and long missing joint defects are long-segment defects, and thick joint defects are short-segment defects. The missing knot defects and interlacing decorative yarn defects are mutation defects. Therefore, for different types of defects, the pooling size of the average pooling should be different. For example, for long-segment defects, the amount of data averaged each time should be larger, so as to highlight the changes in the width values of long and short yarns; For short-segment defects, the corresponding amount of data averaged each time should be smaller to highlight the characteristics of the defects.

In one embodiment, when yarn defects include detail defects and long gap defects, the pooling size of the average pooling can be Among them, a ₁ =1, 2, 3...N ₁ , m is the length of the one-dimensional sequence,

Based on the pooling size, the feature sequence S ₁ (a ₁ ) can be obtained, whose length is For example, when m=100 and a ₁ =1, the pooling size is 50, that is, 50 values are averaged each time and translated sequentially. Finally, the feature sequence S ₁ (a ₁ ) constructed from 51 values can be obtained. ).

In one embodiment, yarn defects include slub defects, and the pooling size of the average pooling can be Among them, a ₂ =1, 2, 3...N ₂ , m is the length of the one-dimensional sequence,/>

Based on the pooling size, the feature sequence S ₂ (a ₂ ) can be obtained, whose length is For example, when m = 100, a ₂ = 1, the pooling size is 2, that is, two values are averaged each time and translated sequentially, and finally the feature sequence S ₂ (a _{2 )} constructed with 99 values can be obtained. ).

In one embodiment, yarn defects include short knot defects and decorative yarn staggered defects. Due to the mutation characteristics of these two defects, when performing average pooling, average pooling with different pooling sizes can be used twice, and then the The difference value between the two is used to construct the feature sequence.

Specifically, two average pooling can be performed on the one-dimensional sequence respectively. The pooling sizes of the two average poolings are P _3-1 =4a ₃ +2 and P _3-2 =2a ₃ +1 respectively, and we get Sequences M _short and M _long , where a ₃ =1, 2, 3...N ₃ , m is the length of the one-dimensional sequence.

Among them, the length of the sequence M _short is m-1-a ₃ × 4, and the length of the sequence M _long is ma ₃ × 2. Afterwards, backward search and difference and forward search and difference can be performed on the sequences M _short and M _long respectively to obtain the backward difference sequence and the forward difference sequence.

Specifically, since the lengths of the two sequences are different, the method for backward search to find the difference is: remove the first ₃ values from the head end of the sequence M _short , and compare the sequences M _short and M _long sequentially from the head end to the tail end. The backward difference sequence S _3-1 =|M _short [a ₃ +1:m+1]| is obtained. Please refer to Figure 3, which is a schematic diagram of an implementation of backward search in this application.

The method of forward search and difference is: remove the last a ₃ values from the tail end of the sequence M _short , and sequentially compare the sequences M _short and M _long from the tail end to the head end, and obtain the forward difference sequence S _3-2 = |M _short [1:m-1-5a ₃ ]-M _long [3a ₃ +2:m-2a ₃ ]|. Please refer to Figure 4, which is a schematic diagram of an implementation of backward search in this application.

After obtaining the backward difference sequence and the forward difference sequence, the larger value of each corresponding position of the backward difference sequence and the forward difference sequence can be taken to construct the feature sequence S ₃ (a ₃ ).

It can be understood that the above-mentioned average pooling method can effectively highlight the characteristics of mutation defects, thereby facilitating subsequent classification.

Please refer to Figure 5. Figure 5 is a feature sequence diagram of yarns with different defect types in this application. As shown in Figure 5, for yarns with different defect types, different feature sequences can well characterize their characteristics, thereby providing It lays a reliable foundation for subsequent defect detection.

S500. Based on the feature sequence, construct the optimal feature vector.

Specifically, since the feature sequence is a sequence related to the variable a _i , when the value of a _i is different, the feature sequence is also different. Therefore, it is necessary to calculate the optimal solution for variable a _i .

In one implementation, the optimal solution of a _i can be calculated based on the stochastic leapfrog algorithm of partial least squares. The specific method is introduced in detail below.

First, you can calculate the maximum, minimum and average statistics of the feature sequence when a _i takes any value.

For example, the maximum value, minimum value and sum of the three characteristic sequences S ₁ (a ₁ ), S ₂ (a ₂ ), and S ₃ (a ₃ ) calculated in the above step S400 when the variables are 1 to N can be calculated. The average value is calculated to construct the feature sequence combination V(L). Taking N=5 as an example, see the following formula for details.

V(L)＝[L ₁ (S ₁ )L ₂ (S ₂ )L ₃ (S ₃ )]

From the above formula, it can be seen that the size of sequence V(L) is: 1×45. Afterwards, the random leapfrog algorithm based on partial least squares (PLS) can be used to find the optimal values of a ₁ , a ₂ , and a ₃ in the sequence V(L), so as to reconstruct the feature vector L={l[S ₁ ( a ₁ )] l[S ₂ (a ₂ )] l[S ₃ (a ₃ )]}

Among them, the implementation steps of the partial least squares stochastic frog leaping algorithm are as follows:

Step1: Initialization, including a subset V0 containing Q variables. The number of iterations is 1000, the number of variables Q=5, η=0.1, and ω=3.

Step2: Randomly generate Q* from the normal distribution Norm(Q,θQ), including the candidate subset V*.

Step3: If Q*=Q, then V*=V0. If Q*<Q, then use V0 to build a PLS model, calculate the regression coefficients of each variable, delete the Q-Q* variables related to the minimum absolute regression coefficient, and the remaining Q* variables constitute V*. If Q*>Q, then randomly select ω (Q*-Q) variables from V-V0 to form a variable subset S, use V0 and S to build a PLS model, calculate the regression coefficients of each variable, and retain the largest value in the PLS model The Q* variables of absolute regression coefficients constitute V*.

Step4: Calculate the cross-validation root mean square errors RMSECV and RMSECV* of V0 and V*. If RMSECV*<=RMSECV, then use V* as V1. Otherwise, accept V* as V1 with probability ηRMSECV/RMSECV*. Update V0 with V1 and return to Step2 until the iteration ends.

Step5: Calculate the selection probability of each variable after the iteration is completed. The frequency of the jth variable selected in the variable subset is denoted as Nj. The selection probability of each variable is as follows:

According to the selection probability of each variable, calculate [S ₁ (a ₁ )](a ₁ =1,2,3,4,5) and [S ₂ (a ₂ )](a ₂ =1,2,3 ,4,5) and [S ₃ (a ₃ )] (a ₃ =1,2,3,4,5), select the highest probability [S ₁ (a _i )], [S ₂ ( a _i )], [S ₃ (a _i )], a ₁ , a ₂ , a ₃ in the final reconstructed feature vector L, which is the optimal feature vector.

It should be noted that the above embodiments only illustrate the method of simultaneously calculating the optimal solution for three feature sequences, and the optimal feature vector finally obtained is a 1×9 vector. In other implementations, the optimal feature vector of one feature sequence can also be calculated solely based on the random leapfrog algorithm of partial least squares, or the optimal feature vectors of two feature sequences can be calculated, both of which can achieve the effects of this implementation.

S600. Input the optimal feature vector into the deep learning neural network and output the characteristic attributes of the yarn.

Among them, the characteristic attributes include normal or yarn defect type.

In one embodiment, the deep learning neural network may be a two-layer ANN classifier. The two-layer ANN classifier may include an input layer and an output layer located at both ends, and two hidden layers located inside.

Specifically, please refer to FIG. 6 , which is a schematic structural diagram of an embodiment of the deep learning neural network of the present application. As shown in Figure 6, in one embodiment, for the 1×9 optimal feature vector obtained in the above step S500, the input layer may include 9 neurons, and the output layer may include 6 neurons. Corresponding to the normal state and five defects respectively, the hidden layer can be a fully connected layer. Each fully connected layer can include 10 hidden neurons. By inputting the optimal feature vector into the input layer, after two layers of fully connected layers, it can be passed One neuron in the output layer performs classification output to obtain the yarn defect type.

Specifically, the training methods of the above-mentioned deep learning neural network include:

Obtain a sample training set, which includes several groups of optimal feature vectors of yarn images marked with label values;

Input the sample training set, and based on the cross-entropy loss function and label value, update the weight of the deep learning neural network in the direction of gradient descent until the cross-entropy loss function converges, and obtain the deep learning neural network.

Specifically, in order to achieve classification, the above-mentioned deep learning neural network needs to perform forward propagation and back propagation calculations during the training process.

The forward propagation of a double hidden layer neural network refers to the process of data being transferred from the input layer through each hidden layer to the output layer. Input x as a neuron into the first hidden layer and calculate it with the following formula,

h ₁ =σ(w ₁ x+b ₁ )

Among them, w ₁ and b ₁ are the weight and bias of the first hidden layer respectively, and σ is the ReLU (RectifiedLinearUnit) activation function, which is σ=max(0,x).

Pass the result h ₁ to the second hidden layer to be calculated with the following formula,

h ₂ =σ(w ₂ h ₁ +b ₂ )

Among them, w ₂ and b ₂ are the weight and bias of the second hidden layer respectively.

Pass the result h ₂ to the output layer for calculation,

y＝softmax(w ₃ h ₂ +b ₃ )

Among them, w ₃ and b ₃ are the weight and bias of the output layer respectively, softmax is the activation function,

The output of the neural network can be changed into a probability distribution between 0 and 1. The output y is a vector with a dimension of 6, including normal yarn, thick sections, details, long and missing sections, short sections and decorative yarn interleaving, and finally achieves Yarn classification based on feature vector L.

The back propagation of the two-layer neural network calculates the error between the label value and the result value, propagates the error along the reverse direction of the network according to the error, and updates the weight w and bias value b in each layer.

① Calculate the error of the output layer

Calculate the deviation E between the classification result and the expectation based on the cross-entropy loss function of the classification result and the actual label:

Among them, m is the number of training samples, C is the number of categories, t _ij and y _ij are respectively the true label of the j-th category of the i-th training sample and the probability that the model predicts the j-th category. The goal of the cross-entropy loss function is to minimize the gap between the predicted value and the true value. When the predicted probability is closer to the true label, the value of the loss function is closer to 0, and vice versa. Therefore, the cross-entropy loss function can help the model fit the training data better and improve the accuracy of the classification task.

② Calculate the error of the second hidden layer

The error can be passed along the network to the second hidden layer. Through the chain rule, the error can be expressed as:

Among them, δ ₂ is the error of the second hidden layer.

③Calculate the error of the first hidden layer

The second hidden layer error δ ₂ can also be passed along the network to the first hidden layer to calculate the error of the first hidden layer:

Among them, δ ₁ represents the error of the second hidden layer, (δ ₂ >0) is the indicator function, when δ ₂ >0, the value is 1, when δ ₂ <0, the value is 0.

④Update weights and biases

Use the error to update the weights and biases. The update of weights and biases can be achieved using the gradient descent algorithm, that is, by adjusting the weights and bias values in the direction of decreasing error. The weights and biases are updated as follows:

Among them, lr is the learning rate, and the amplitude can be adjusted according to the error each time it is updated.

After updating the weights and biases using the backpropagation algorithm, calculate the forward propagation again, and repeat the above process until the bias values in each round converge. By recording all values and deviations for forward propagation, a chenille yarn defect classification model can be established, and the model can be used to detect chenille yarns.

This application also provides a machine vision-based yarn defect detection device. Please refer to Figure 7 . Figure 7 is a structural schematic diagram of an embodiment of the machine vision-based yarn defect detection device of this application.

The device includes an acquisition module 21, a binarization processing module 22, a dimensionality reduction processing module 23, a mean pooling module 24, a construction module 25 and a classification output module 26.

Among them, the acquisition module 21 is used to acquire the image of the yarn;

The binarization processing module 22 is used to perform binarization processing on the image of the yarn to obtain a binary image of the yarn;

The dimensionality reduction processing module 23 is used to perform dimensionality reduction processing on the binary image to obtain a one-dimensional sequence of yarn width values;

The mean pooling module 24 is used to perform mean pooling on the one-dimensional sequence to obtain the feature sequence, where the pooling size of the mean pooling is P _i =f(a _i ), a _i =1, 2, 3.. .N;

The construction module 25 is used to construct the optimal feature vector based on the feature sequence;

The classification output module 26 is used to input the optimal feature sequence into the deep learning neural network and output the characteristic attributes of the yarn, where the characteristic attributes include normal or yarn defect types.

As above, with reference to Figures 1 to 6, the yarn defect detection method based on machine vision according to the embodiment of this specification is described. The details mentioned in the above description of the method embodiments are also applicable to the machine vision-based yarn defect detection device in the embodiments of this specification. The above yarn defect detection device based on machine vision can be implemented by hardware, software or a combination of hardware and software.

FIG. 8 is a hardware structure diagram of an embodiment of the electronic device of the present application. As shown in FIG. 8 , the electronic device 30 may include at least one processor 31 , a memory 32 (eg, a non-volatile memory), a memory 33 , and a communication interface 34 , and at least one processor 31 , a memory 32 , a memory 33 , and a communication interface. 34 are connected together via bus 35 . At least one processor 31 executes at least one computer readable instruction stored or encoded in memory 32 .

It should be understood that the computer-executable instructions stored in memory 32, when executed, cause at least one processor 31 to perform the various operations and functions described above in connection with FIGS. 1-4 in various embodiments of this specification.

In the embodiment of the present description, the electronic device 30 may include, but is not limited to: a personal computer, a server computer, a workstation, a desktop computer, a laptop computer, a notebook computer, a mobile electronic device, a smart phone, a tablet computer, a cellular phone, Personal digital assistants (PDAs), handheld devices, messaging devices, wearable electronic devices, consumer electronics devices, etc.

According to one embodiment, a program product, such as a machine-readable medium, is provided. The machine-readable medium may have instructions (i.e., the above-mentioned elements implemented in the form of software) that, when executed by a machine, cause the machine to perform various operations described above in connection with FIGS. 1-4 in various embodiments of this specification and Function. Specifically, a system or device equipped with a readable storage medium may be provided, on which the software program code that implements the functions of any of the above embodiments is stored, and the computer or device of the system or device may The processor reads and executes the instructions stored in the readable storage medium.

In this case, the program code itself read from the readable medium can implement the functions of any one of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute this specification a part of.

Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tape, non- Volatile memory cards and ROM. Alternatively, the program code can be downloaded from the server computer or the cloud by the communication network.

Those skilled in the art will understand that various variations and modifications can be made to the various embodiments disclosed above without departing from the essence of the invention. Therefore, the protection scope of this specification should be defined by the appended claims.

It should be noted that not all steps and units in the above-mentioned processes and system structure diagrams are necessary, and some steps or units can be ignored according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above embodiments may be a physical structure or a logical structure, that is, some units may be implemented by the same physical client, or some units may be implemented by multiple physical clients, or may be implemented by multiple physical clients. Some components in separate devices are implemented together.

In each of the above embodiments, the hardware unit or module can be implemented mechanically or electrically. For example, a hardware unit, module or processor may include permanent dedicated circuitry or logic (such as a dedicated processor, FPGA or ASIC) to complete the corresponding operation. The hardware unit or processor may also include programmable logic or circuits (such as a general-purpose processor or other programmable processors), which may be temporarily set by software to complete corresponding operations. The specific implementation method (mechanical method, or dedicated permanent circuit, or temporarily installed circuit) can be determined based on cost and time considerations.

The detailed description set forth above in conjunction with the drawings describes exemplary embodiments and does not represent all embodiments that can be implemented or that fall within the scope of the claims. The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not mean "preferred" or "advantageous" over other embodiments. The detailed description includes specific details for the purpose of providing an understanding of the described technology. However, these techniques can be implemented without these specific details. In some instances, well-known structures and devices are shown in block diagram form in order to avoid obscuring the concepts of the described embodiments.

The above description of the disclosure is provided to enable any person of ordinary skill in the art to make or use the disclosure. Various modifications to the present disclosure will be obvious to those of ordinary skill in the art, and the general principles corresponding to this article may also be applied to other modifications without departing from the scope of the disclosure. . Thus, the present disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. A yarn defect detection method based on machine vision, which is characterized by including: Get an image of yarn; Perform binarization processing on the image of the yarn to obtain a binary image of the yarn; Perform dimensionality reduction processing on the binarized image to obtain a one-dimensional sequence of yarn width values; Perform average pooling on the one-dimensional sequence to obtain a feature sequence, wherein the pooling size of the average pooling is P _i =f(a _i ), a _i =1, 2, 3...N; Based on the feature sequence, construct an optimal feature vector; The optimal feature vector is input into a deep learning neural network and the characteristic attributes of the yarn are output, where the characteristic attributes include normal or the type of yarn defect. 2. The yarn defect detection method according to claim 1, wherein the step of performing binarization processing on the image of the yarn to obtain the binary image of the yarn includes: Calculate the OSTU threshold T _OSTU of the image of the yarn based on the OSTU threshold method; The OSTU threshold is corrected based on the following formula to obtain the corrected threshold T _{g-global-OSTU} : In the formula, V _q is the gray value of the pixel with the largest number of the same gray value among all the pixels in the image of the yarn, and Z _x is the smallest gray value among all the pixels in the image of the yarn. The gray value of the pixel; The image is binarized based on the correction threshold to obtain a binarized image of the yarn. 3. The yarn defect detection method according to claim 1, wherein the step of performing dimensionality reduction processing on the binary image to obtain a one-dimensional sequence of yarn width values includes: Taking the background pixel of the binary image as value 1 and the foreground pixel as value 0, the matrix X _n*m of the binary image is obtained, where n is the width direction of the yarn in the binary image The number of pixels on, m is the number of pixels of the binary image in the length direction of the yarn; Based on the following formula, the one-dimensional sequence D of the yarn width value is calculated: D＝[n n...n] _1*m -[1 1...1] _1*n *X _n*m . 4. The yarn defect detection method according to claim 1, wherein the yarn defects include detail defects and long missing joint defects, and the one-dimensional sequence is averaged and pooled to obtain a feature sequence. The steps include: Perform average pooling on the one-dimensional sequence, and the pooling size Obtain the characteristic sequence S ₁ (a ₁ ), where a ₁ =1, 2, 3...N ₁ , m is the length of the one-dimensional sequence, /> 5. The yarn defect detection method according to claim 1, characterized in that the yarn defects include slub defects, and the step of performing average pooling on the one-dimensional sequence to obtain the feature sequence includes: Perform average pooling on the one-dimensional sequence, and the pooling size Obtain the characteristic sequence S ₂ (a ₂ ), where a ₂ =1, 2, 3...N ₂ , m is the length of the one-dimensional sequence, /> 6. The yarn defect detection method according to claim 1, characterized in that the yarn defects include missing knot defects and decorative yarn staggered defects, and the one-dimensional sequence is averaged and pooled to obtain features. The steps of the sequence include: Perform average pooling twice on the one-dimensional sequence. The pooling sizes of the two average poolings are P _3-1 =4a ₃ +2 and P _3-2 =2a ₃ +1 respectively, and the sequence M is obtained. _short and M _long , where a ₃ =1, 2, 3...N ₃ , m is the length of the one-dimensional sequence; Perform backward search and difference and forward search and difference on the sequences M _short and M _long respectively to obtain a backward difference sequence and a forward difference sequence; Compare the backward difference sequence and the forward difference sequence, and take the larger value of each corresponding position of the backward difference sequence and the forward difference sequence to construct the feature sequence S ₃ (a ₃ ). 7. The yarn defect detection method according to claim 6, wherein the backward search difference and the forward search difference are performed on the sequences M _short and M _long respectively to obtain a backward difference sequence. The steps for summing up the forward difference sequence include: The first a ₃ values at the head end of the sequence M _short are removed, and the sequences M _short and M _long are sequentially compared from the head end to the tail end to obtain the backward difference sequence S _3-1 =|M _short [a ₃ +1：m+1-4a ₃ ]-M _long [1：m-1-5a ₃ ]|, where m is the length of the one-dimensional sequence; Remove the last ₃ values of the tail end of the sequence M _short , and compare the sequences M _short and M _long in sequence from the tail end to the head end, and obtain the forward difference sequence S _3-2 =|M _short [1 :m-1-5a ₃ ]-M _long [3a ₃ +2: m-2a ₃ ]|, where m is the length of the one-dimensional sequence. 8. The yarn defect detection method according to claim 1, wherein the step of constructing an optimal feature vector based on the feature sequence includes: The stochastic leapfrog algorithm based on partial least squares calculates the optimal solution of a _i ; Based on the optimal solution, the maximum value, minimum value and average value of the feature sequence are calculated, and the optimal feature vector is collected. 9. The yarn defect detection method according to claim 1, characterized in that, in the step of inputting the optimal feature vector into a deep learning neural network and outputting the characteristic attributes of the yarn, the neural network It is a two-layer ANN classifier, and the training method of the deep learning neural network includes: Obtain a sample training set, the sample training set includes the optimal feature vectors of several groups of yarn images marked with label values; The sample training set is input, and based on the cross-entropy loss function and the label value, the weight of the deep learning neural network is updated in the direction of gradient descent until the cross-entropy loss function converges, and the deep learning neural network is obtained . 10. A yarn defect detection device based on machine vision, characterized by including: Get module, used to get the image of yarn; A binarization processing module is used to perform binarization processing on the image of the yarn to obtain a binary image of the yarn; A dimensionality reduction processing module, used to perform dimensionality reduction processing on the binary image to obtain a one-dimensional sequence of yarn width values; A mean pooling module, used to perform mean pooling on the one-dimensional sequence to obtain a feature sequence, wherein the pooling size of the mean pooling is Pi ₌ f (a _i ), a _i = 1, 2 ,3...N; A building module for constructing an optimal feature vector based on the feature sequence; A classification output module, configured to input the optimal feature sequence into a deep learning neural network and output characteristic attributes of the yarn, where the characteristic attributes include normal or types of yarn defects.