Crimping pipe detection method based on image denoising

Abstract

The invention discloses a crimping tube detection method based on image denoising, which comprises the following steps of: firstly, acquiring an aerial image of a crimping tube, and optimizing by taking an exponential decay time constant alpha T and an intrinsic voltage constant VT of a pulse coupling neural network as initial population positions of an improved aurora optimization algorithm; obtaining an optimal exponential decay time constant alpha T and an optimal natural voltage constant VT of the pulse coupling neural network; the method comprises the following steps of: firstly, training a pulse coupling neural network to obtain a crimping pipe detection model, secondly, inputting an original image obtained by aerial photography into the trained model so as to de-noise the original image, and finally, inputting the de-noised image into a yolov5 model for detection. According to the method provided by the invention, the de-noising capability of the image can be remarkably enhanced, so that the detection precision of the yov5 model on the crimping pipe is improved.

Description

Technical Field

[0001] The present invention relates to the field of crimp tube detection technology, and in particular to a crimp tube detection method based on image denoising. Background Art

[0002] The breakdown failure of power cables is mainly concentrated in the cable joints and terminal positions. As an important part of the cable joint, the cable conductor crimping tube is responsible for completing the connection task of the conductors at both ends of the cable joint. The installation work of the cable conductor crimping tube is basic but easy to be overlooked. In recent years, there have often been cases where poor crimping of cable conductors has caused the heating of the crimping tube position to accelerate the aging of the internal materials of the cable joint under long-term working conditions, thereby causing the joint to break down and cause failures. This seriously threatens the safe and stable operation of the transmission line. Therefore, regular quality inspection of typical hardware such as transmission line crimping tubes is one of the important contents of the operation and maintenance management of the transmission system.

[0003] At present, the quality inspection of power transmission line fittings mainly includes manual tower inspection, overhead maintenance robot inspection and drone inspection. Among them, drone inspection is the most convenient, but the image detection technology relied on by drone inspection requires a lot of calculations due to the complex aerial background, and the accuracy of the image quality is poor. Therefore, image denoising is of great significance for crimping tube inspection. Summary of the invention

[0004] In view of the above-mentioned problems, many researchers have proposed a variety of solutions, but the effects are not significant. In order to actually solve the current technical difficulties, the present invention discloses a compression tube detection method based on image denoising. First, an aerial image of the compression tube is obtained, and then the image is denoised using a pulse coupled neural network optimized by an improved Aurora algorithm. Finally, the denoised image is input into the YOLOv5 model for image detection.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: a method for detecting a crimping tube based on image denoising, comprising the following steps:

[0006] Step S1: Use a drone to take aerial photos of the compressed tube image.

[0007] Step S2: construct a pulse coupled neural network model, input the compression tube image into the pulse coupled neural network model for training, and during the training process, obtain the optimal exponential decay time constant α of the pulse coupled neural network model through the improved Aurora algorithm T and the inherent voltage constant V T ;

[0008] Step S3: inputting the compressed tube image taken by the drone in real time into the optimized pulse coupled neural network model for denoising;

[0009] Step S4: input the denoised image into the yolov5 model for crimping tube detection;

[0010] Step S2 is specifically as follows:

[0011] Step S21: setting the initialization population size and maximum number of iterations of the Aurora algorithm;

[0012] Step S22: Initialize population position;

[0013] Step S23A: simulating the sun's rotation motion;

[0014] Step S23B: simulating the aurora elliptical trail and introducing escape strategies and nonlinear factors;

[0015] Step S23C: simulating particle collision;

[0016] Step S24: Determine whether the current iteration has reached the maximum number of times. If not, continue the iteration; otherwise, stop the iteration and output the optimal individual, which is the optimal exponential decay time constant α of the pulse coupled neural network model. T and the inherent voltage constant V T The optimal value of .

[0017] Furthermore, the improved Aurora algorithm is used to obtain the optimal exponential decay time constant α of the pulse coupled neural network model. T and the inherent voltage constant V T The specific process is:

[0018] Step S21: setting the initialization population size and maximum number of iterations of the Aurora algorithm;

[0019] Step S22: The optimal exponential decay time constant and inherent voltage constant V of the pulse coupled neural network model are T As the initial position of the population, the original method of initializing the population is improved by using the good point set operator to improve the traversability of the population; it is expressed as:

[0020]

[0021] x i,j =P n {j}(ub-lb)+lb;

[0022] In the formula, X represents the initial population after the optimization of the good point set operator, x i,jRepresents the value of the i-th aurora individual in the j-th dimension after the optimization of the good point set, i∈1,2,…,n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; each aurora individual represents a set of optimal exponential decay time constants α of the pulse coupled neural network model T and the inherent voltage constant V T The parameter solution, P n {j} represents the good point set operator, ub and lb are the upper and lower bounds of the problem respectively.

[0023] Further, step S23A is specifically: simulating the rotation of the sun, expressed as:

[0024]

[0025] Among them, ν(t) represents the Lorentz force at the tth iteration, t represents the current number of iterations, the integral constant C, the charge q, mass m and geomagnetic field strength B carried by the charged particle do not change; a represents the damping factor, which takes a random value of [1,1.5]. Time is simulated by calculating the fitness evaluation process.

[0026] Further, step S23B: simulating the aurora elliptical trail and introducing escape strategies and nonlinear factors to find the global optimal or better solution; expressed as:

[0027]

[0028] Ao=Levy×(X avg -X(t))+lb+r 1 ×(ub-lb) / 2;

[0029]

[0030]

[0031]

[0032] Among them, X(t+1) is the position of the aurora individual at the t+1 iteration, X(t) is the position of the aurora individual at the t iteration, represents the individual closest to individual t, randn represents a random variable that follows a normal distribution, rand represents a random value from 0 to 1, θ represents a constant that changes with the number of iterations, Ao represents the distribution of the complex changes of the auroral ellipse simulated by the discrete variable, Levy represents the Levy flight, and X avg is the average position of the population, r 1 and r 2 To take a random value in [0,1], W 1 and W 2is a constant value that changes with the number of iterations, and T is the maximum number of iterations.

[0033] Furthermore, the simulation of particle collision is as follows: The mathematical model is shown as follows:

[0034] X(t+1)=X(t)+sin(r 3 ×π)×(X(t)-X r ),r 4 <Kandr 5 <0.05;

[0035]

[0036] Among them, X r represents the position of any individual in the population. As the algorithm proceeds, collisions between particles become more and more frequent, and are therefore controlled by the collision probability K; r 3 、r 4 and r 5 is a random value in the range [0,1].

[0037] The denoised image is input into the Yolov5 model. First, the denoised input image is converted into a multi-layer feature map through the backbone network for subsequent target detection tasks. The ResNet backbone network is used. This network is relatively lightweight and can minimize the amount of calculation and memory usage while ensuring high detection accuracy. The main structures in the backbone network are Conv module, C3 module, and SPPF module. Secondly, the extracted feature maps of different scales are input into the Neck feature pyramid. In Yolov5, FPN is used to fuse feature maps of different scales together through upsampling and downsampling operations to generate a multi-scale feature pyramid. The top-down part mainly realizes the fusion of features at different levels by upsampling and fusion with coarser-grained feature maps, while the bottom-up part fuses feature maps from different levels by using a convolutional layer. Finally, the fused feature map is input into the head for the final compression tube detection.

[0038] Furthermore, the optimized pulse coupled neural network model is used to perform denoising on the aerial images, as follows:

[0039] Step S31: initializing parameters;

[0040] Step S32: The output of the pulse coupled neural network is a binary matrix. It is not difficult to determine the position coordinates of the noise data based on the binary matrix, and each noise data is used as the midpoint of the 3*3 filter window in turn;

[0041] Step S33: Determine the weight according to the number of noise data in the other eight points of the window except the midpoint. If the number of noise data = 0, it corresponds to filter 1; if the number of noise data = 1, it corresponds to filter 2; if the number of noise data = 1, it corresponds to filter 2; ≥ 2 corresponds to filter 3;

[0042] Step S34: Process the noise data and sort the depth values ​​according to the weights and the eight point data in the filter window except the midpoint. It is worth mentioning that the noise data of non-window center points such as 0 and 255 will be removed during sorting, and the median value will be selected from the removed data and assigned to the noise data in the midpoint of the filter window.

[0043] Step S35: Output the reconstructed image.

[0044] The mathematical model of pulse coupled neural network is as follows:

[0045]

[0046]

[0047] U ij (n) = F ij (n)[1+βL ij (n)];

[0048]

[0049]

[0050] Among them, F ij (n) is the input item of neuron (i, j) at time n, I ij (n) is the external input stimulus of neuron (i, j) at time n, L ij (n) is the connection input item of neuron (i, j) at time n, Y ij (n) is the output of neuron (i, j) at time n, and its value is 0 or 1. U ij (n) is the internal activity item of neuron (i, j) at time n, E ij (n) is the dynamic threshold of neuron (i, j) at time n. When the internal activity item is greater than the dynamic threshold, it will ignite and generate a pulse, that is, the output of the neuron Y ij =1, M ijkl W ijkl are the link weight matrices of the input items and the connecting input items, respectively. Their values ​​represent the influence of the surrounding neurons on the intermediate neurons. F , α L , α E Represents F ij ,L ij , Eij The decay time constant, V F , V L ,V E is the corresponding amplitude coefficient, β is the connection strength coefficient, which is used to adjust the degree of influence between a single neuron and nearby neurons, and e is the natural base.

[0051] The beneficial effects of the present invention are as follows: the pulse coupled neural network optimized by the improved Aurora optimization algorithm is used for image denoising, which solves the exponential decay time constant α of the pulse coupled neural network. T and the inherent voltage constant V T The problem of difficult to accurately select is optimized by improving the Aurora optimization algorithm to optimize the exponential decay time constant α T and the inherent voltage constant V T The subsequent pulse coupled neural network has enhanced the denoising ability of images and improved the detection accuracy of the yolov5 model for pressure pipes, which is of great significance for the safe dispatching and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Flowchart of the crimped tube detection method based on image denoising.

[0053] Figure 2 Flowchart of the Aurora algorithm.

[0054] Figure 3 Aurora algorithm improvement comparison chart. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0056] Reference Figure 1 , a crimping tube detection method based on image denoising, comprising the following steps:

[0057] Step S1: Use a drone to take aerial photos of the compressed tube image.

[0058] Step S2: construct a pulse coupled neural network model, input the compression tube image into the pulse coupled neural network model for training, and during the training process, obtain the optimal exponential decay time constant α of the pulse coupled neural network model through the improved Aurora algorithm T and the inherent voltage constant V T ;

[0059] Step S3: inputting the compressed tube image taken by the drone in real time into the optimized pulse coupled neural network model for denoising;

[0060] Step S4: input the denoised image into the yolov5 model for crimping tube detection;

[0061] Step S2 is specifically as follows:

[0062] Step S21: setting the initialization population size and maximum number of iterations of the Aurora algorithm;

[0063] Step S22: Initialize population position;

[0064] Step S23A: simulating the sun's rotation motion;

[0065] Step S23B: simulating the aurora elliptical trail and introducing escape strategies and nonlinear factors;

[0066] Step S23C: simulating particle collision;

[0067] Step S24: Determine whether the current iteration has reached the maximum number of times. If not, continue the iteration; otherwise, stop the iteration and output the optimal individual, which is the optimal exponential decay time constant α of the pulse coupled neural network model. T and the inherent voltage constant V T The optimal value of .

[0068] Furthermore, the improved Aurora algorithm is used to obtain the optimal exponential decay time constant α of the pulse coupled neural network model. T and the inherent voltage constant V T The specific process is:

[0069] Step S21: setting the initialization population size and maximum number of iterations of the Aurora algorithm;

[0070] Step S22: The optimal exponential decay time constant and inherent voltage constant V of the pulse coupled neural network model are T As the initial position of the population, the original method of initializing the population is improved by using the good point set operator to improve the traversability of the population; it is expressed as:

[0071]

[0072] x i,j =P n {j}(ub-lb)+lb;

[0073] In the formula, X represents the initial population after the optimization of the good point set operator, x i,j Represents the value of the i-th aurora individual in the j-th dimension after the optimization of the good point set, i∈1,2,…,n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; each aurora individual represents a set of optimal exponential decay time constants α of the pulse coupled neural network model Tand the inherent voltage constant V T The parameter solution, P n {j} represents the good point set operator, ub and lb are the upper and lower bounds of the problem respectively.

[0074] Further, step S23A is specifically: simulating the rotation of the sun, expressed as:

[0075]

[0076] Among them, ν(t) represents the Lorentz force at the tth iteration, t represents the current number of iterations, the integral constant C, the charge q, mass m and geomagnetic field strength B carried by the charged particle do not change; a represents the damping factor, which takes a random value of [1,1.5]. Time is simulated by calculating the fitness evaluation process.

[0077] Further, step S23B: simulating the aurora elliptical trail and introducing escape strategies and nonlinear factors to find the global optimal or better solution; expressed as:

[0078]

[0079] Ao=Levy×(X avg -X(t))+lb+r 1 ×(ub-lb) / 2;

[0080]

[0081]

[0082]

[0083] Among them, X(t+1) is the position of the aurora individual at the t+1 iteration, X(t) is the position of the aurora individual at the t iteration, represents the individual closest to individual t, randn represents a random variable that follows a normal distribution, rand represents a random value from 0 to 1, θ represents a constant that changes with the number of iterations, Ao represents the distribution of the complex changes of the auroral ellipse simulated by the discrete variable, Levy represents the Levy flight, and X avg is the average position of the population, r 1 and r 2 To take a random value in [0,1], W 1 and W 2 is a constant value that changes with the number of iterations, and T is the maximum number of iterations.

[0084] Furthermore, the simulation of particle collision is as follows: The mathematical model is shown as follows:

[0085] X(t+1)=X(t)+sin(r3 ×π)×(X(t)-X r ),r 4 <Kandr 5 <0.05;

[0086]

[0087] Among them, X r represents the position of any individual in the population. As the algorithm proceeds, collisions between particles become more and more frequent, and are therefore controlled by the collision probability K; r 3 、r 4 and r 5 is a random value in the range [0,1].

[0088] The denoised image is input into the Yolov5 model. First, the denoised input image is converted into a multi-layer feature map through the backbone network for subsequent target detection tasks. The ResNet backbone network is used. This network is relatively lightweight and can minimize the amount of calculation and memory usage while ensuring high detection accuracy. The main structures in the backbone network are Conv module, C3 module, and SPPF module. Secondly, the extracted feature maps of different scales are input into the Neck feature pyramid. In Yolov5, FPN is used to fuse feature maps of different scales together through upsampling and downsampling operations to generate a multi-scale feature pyramid. The top-down part mainly realizes the fusion of features at different levels by upsampling and fusion with coarser-grained feature maps, while the bottom-up part fuses feature maps from different levels by using a convolutional layer. Finally, the fused feature map is input into the head for the final compression tube detection.

[0089] Furthermore, the optimized pulse coupled neural network model is used to perform denoising on the aerial images, as follows:

[0090] Step S31: initializing parameters;

[0091] Step S32: The output of the pulse coupled neural network is a binary matrix. It is not difficult to determine the position coordinates of the noise data based on the binary matrix, and each noise data is used as the midpoint of the 3*3 filter window in turn;

[0092] Step S33: Determine the weight according to the number of noise data in the other eight points of the window except the midpoint. If the number of noise data = 0, it corresponds to filter 1; if the number of noise data = 1, it corresponds to filter 2; if the number of noise data = 1, it corresponds to filter 2; ≥ 2 corresponds to filter 3;

[0093] Step S34: Process the noise data and sort the depth values ​​according to the weights and the eight point data in the filter window except the midpoint. It is worth mentioning that the noise data of non-window center points such as 0 and 255 will be removed during sorting, and the median value will be selected from the removed data and assigned to the noise data in the midpoint of the filter window.

[0094] Step S35: Output the reconstructed image.

[0095] The mathematical model of pulse coupled neural network is as follows:

[0096]

[0097]

[0098] U ij (n) = F ij (n)[1+βL ij (n)];

[0099]

[0100]

[0101] Among them, F ij (n) is the input item of neuron (i, j) at time n, I ij (n) is the external input stimulus of neuron (i, j) at time n, L ij (n) is the connection input item of neuron (i, j) at time n, Y ij (n) is the output of neuron (i, j) at time n, and its value is 0 or 1. U ij (n) is the internal activity item of neuron (i, j) at time n, E ij (n) is the dynamic threshold of neuron (i, j) at time n. When the internal activity item is greater than the dynamic threshold, it will ignite and generate a pulse, that is, the output of the neuron Y ij =1, M ijkl W ijkl are the link weight matrices of the input items and the connecting input items, respectively. Their values ​​represent the influence of the surrounding neurons on the intermediate neurons. F , α L , α E Represents F ij ,L ij , E ij The decay time constant, V F , V L ,V E is the corresponding amplitude coefficient, β is the connection strength coefficient, which is used to adjust the degree of influence between a single neuron and nearby neurons, and e is the natural base.

[0102] Furthermore, the exponential decay time constant α of the pulse coupled neural network is T and the inherent voltage constant V T As the initial population position of the Aurora optimization algorithm, we can find the optimal exponential decay time constant α of the pulse coupled neural network. T and the inherent voltage constant V T ; The Aurora optimization algorithm speeds up the convergence speed and optimization accuracy of the algorithm while increasing its ability to jump out of the local optimal solution. The convergence curve of the Aurora optimization algorithm is shown in Figure 3 As shown in Figure 2, the convergence speed and accuracy have been improved. The optimization process of improving the information acquisition optimization algorithm is as follows: Figure 2 shown.

Claims

1. A method for detecting a crimped tube based on image denoising, characterized in that: The following steps are involved: Step S1: Use a drone to take aerial photos of the compressed tube image; Step S2: construct a pulse coupled neural network model, input the compression tube image into the pulse coupled neural network model for training, and during the training process, obtain the optimal exponential decay time constant α of the pulse coupled neural network model through the improved Aurora algorithm T and the inherent voltage constant V T ; Step S3: inputting the compressed tube image taken by the drone in real time into the optimized pulse coupled neural network model for denoising; Step S4: input the denoised image into the yolov5 model for crimping tube detection; Step S2 is specifically as follows: Step S21: setting the initialization population size and maximum number of iterations of the Aurora algorithm; Step S22: Initialize population position; Step S23A: simulating the sun's rotation motion; Step S23B: simulating the aurora elliptical trail and introducing escape strategies and nonlinear factors; Step S23C: simulating particle collision; Step S24: determine whether the current iteration has reached the maximum number of times, if not, continue iterating; Otherwise, the iteration stops and the optimal individual is output, which is the optimal exponential decay time constant α of the pulse coupled neural network model. T and the inherent voltage constant V T The optimal value of .

2. The method for detecting a crimped tube based on image denoising according to claim 1, characterized in that: The improved Aurora algorithm is used to obtain the optimal exponential decay time constant α of the pulse coupled neural network model. T and the inherent voltage constant V T The specific process is: Step S21: setting the initialization population size and maximum number of iterations of the Aurora algorithm; Step S22: The optimal exponential decay time constant and inherent voltage constant V of the pulse coupled neural network model are T As the initial position of the population, the original method of initializing the population is improved by using the good point set operator to improve the traversability of the population; it is expressed as: x i,j =P n {j}(ub-lb)+lb; In the formula, X represents the initial population after the optimization of the good point set operator, x i,j Represents the value of the i-th aurora individual in the j-th dimension after the optimization of the good point set, i∈1,2,…,n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; each aurora individual represents a set of optimal exponential decay time constants α of the pulse coupled neural network model T and the inherent voltage constant V T The parameter solution, P n {j} represents the good point set operator, ub and lb are the upper and lower bounds of the problem respectively.

3. The method for detecting a crimped tube based on image denoising according to claim 1, characterized in that: Step S23A is specifically: simulating the sun's rotation motion, expressed as: Among them, ν(t) represents the Lorentz force at the tth iteration, t represents the current number of iterations, the integral constant C, the charge q, mass m and geomagnetic field strength B carried by the charged particle do not change; a represents the damping factor, which takes a random value of [1,1.5]. Time is simulated by calculating the fitness evaluation process.

4. The method for detecting a crimped tube based on image denoising according to claim 1, characterized in that: Step S23B: simulate the aurora ellipse trail and introduce escape strategies and nonlinear factors to find the global optimal or better solution; expressed as: Ao=Levy×(X avg -X(t))+lb+r1×(ub-lb) / 2; Among them, X(t+1) is the position of the aurora individual at the t+1 iteration, X(t) is the position of the aurora individual at the t iteration, represents the individual closest to individual t, randn represents a random variable that follows a normal distribution, rand represents a random value from 0 to 1, θ represents a constant that changes with the number of iterations, Ao represents the distribution of the complex changes of the auroral ellipse simulated by the discrete variable, Levy represents the Levy flight, and X avg is the average position of the population, r1 and r2 are random values ​​in the range [0,1], W1 and W2 are constant values ​​that change with the number of iterations, and T is the maximum number of iterations.

5. The method for detecting a crimped tube based on image denoising according to claim 1, characterized in that: The specific simulation of particle collision is as follows: X(t+1)=X(t)+sin(r3×π)×(X(t)-X r ),r4<Kandr5<0.05; Among them, X r Represents the position of any individual in the population. As the algorithm proceeds, collisions between particles become more and more frequent, so it is controlled by the collision probability K; r3, r4, and r5 are random values ​​ranging from [0,1].

Images