CN111650204B - Method and system for detecting defects of transmission line fittings based on cascade target detection - Google Patents

Abstract

The invention discloses a transmission line hardware defect detection method and system based on cascade target detection, comprising the following steps: using a trained first target detection model to detect a connection area of the power transmission line image, and cutting the detected connection area; taking n connecting areas with the area size meeting preset conditions as images to be identified; performing fine hardware defect detection on the image to be identified by using a trained second target detection model, and obtaining coordinates of the fine hardware defect on the image to be identified; according to the method, the defect of the fine hardware fitting is displayed in an original image according to the mapping relation between the coordinates of the image to be identified and the coordinates of the original image, and the method adopts a cascade target detection algorithm to deeply convolve the neural network for identifying the small target of the fine hardware fitting of the power transmission line, and then identifies the connection area in the power transmission line image, and then identifies the defect condition of the fine hardware fitting of the connection area, so that the defect detection precision of the fine hardware fitting is remarkably improved.

Description

Method and system for detecting defects of transmission line fittings based on cascade target detection

technical field

The invention relates to the field of transmission line defect detection, in particular to a method and system for detecting defects of transmission line fittings based on cascade target detection.

Background technique

Small fittings such as power bolts are used in our common power transmission and distribution lines. They need to withstand long-term field operation corrosion and strong collision friction. They have a huge number in the power grid and play a stable role in foundations and line equipment. However, due to the complex and harsh environment in which small hardware is located, it is also a component that is extremely prone to damage. Once damaged, it will cause power interruption and affect the safe operation of the entire power grid. Currently, deep learning image recognition technology is used to perform image recognition processing on defect pictures such as bolt missing parts and bolt missing nuts, so as to form defect diagnosis.

When using the deep learning network model to process image recognition, a large number of training samples are required, but for the defect detection of small fittings, it mainly includes defect categories such as missing pins and protruding pins, but no matter which category, there are too few samples that can be collected. It is easy to overfit when used in deep learning, resulting in unsatisfactory test results, and traditional image augmentation methods, such as random cropping, random translation, random rotation, etc., cannot provide more fine and small hardware defects. layered features.

Moreover, since the defect target of small fittings is too small, accounting for less than 5% of the original image, the traditional deep learning image recognition method has insufficient feature extraction capabilities, resulting in the detection effect of small fittings defects failing to meet actual business needs.

Contents of the invention

Aiming at the problems existing in the above-mentioned prior art, the present invention provides a method for detecting defects of transmission line fittings based on cascaded target detection, uses a generative adversarial network to augment the captured images, increases the number of training samples for the deep learning network, and proposes A two-stage detection strategy, which can significantly improve the detection accuracy of small fitting defects, has the value of promotion and application, including:

(11) Acquiring and preprocessing the images taken by inspection;

(12) Use the trained first target detection model to detect the connected area of the patrol shot image, and cut out the detected rectangular connected area;

(13) Obtain the area size of the connected area, and use n connected areas whose area size meets the preset condition as the image to be identified;

(14) Use the trained second target detection model to detect small metal defects on the image to be recognized, and obtain the coordinates of the small metal defect on the image to be recognized;

(15) According to the mapping relationship between the coordinates of the image to be recognized and the coordinates of the original image, small metal defects are displayed in the original image.

As a further optimization of the above scheme, before training the first target detection model and the second target detection model, data augmentation processing is performed on the collected transmission line images, and the images generated by the augmentation processing and the collected original transmission line images are used together as Datasets for training the first object detection model and the second object detection model.

As a further optimization of the above solution, the data augmentation process uses a trained generative adversarial network model.

As a further optimization of the above scheme, the steps of generating the confrontation network model training are:

(31) Establish a generator and a discriminator, set the loss function of the discriminator and the objective function of generating an adversarial network;

(32) Input the random noise signal z into the generator to obtain generated samples, mark the collected transmission line images and generated samples, and then input them into the discriminator to distinguish between real samples and generated samples, and use the back propagation algorithm to adjust the discrimination The network parameters of the network maximize the objective function of the generative confrontation network to obtain an optimized discriminant network;

(33) Substituting the network parameters of the optimized discriminant network into the objective function of the generative confrontation network, using the backpropagation algorithm to adjust the network parameters of the generative confrontation network, so as to minimize the objective function of the generative confrontation network, and obtain the network parameters of the discriminant network after optimization, Then the optimized generation network is obtained;

(34) Determine whether the number of iterations reaches the preset maximum number of iterations, if it is less than the preset maximum number of iterations, repeat steps (32)-(34), otherwise stop training, and save the generated confrontation model that has been trained.

As a further optimization of the above solution, the first target detection model and the second target detection model use a deep neural network based on a target detection algorithm.

As a further optimization of the above scheme, the first target detection model and the second target detection model adopt a deep neural network based on the FasterRCNN algorithm, and the training process of the FasterRCNN network model includes:

(61) Carry out nested labeling for the data set used for training the model, first mark the connection area of the transmission line in the image, and then mark whether the small fittings have defects based on the connection area;

(62) Data division: divide the training set and verification set in proportion;

(63) Establish the first target detection model and the second target detection model, the forward network all adopts the basic FasterRCNN network, and set the corresponding loss function and weight update method;

(64) Use the training set with nested label data as input data to input the first target detection model and the second target detection model respectively, and calculate the output data and connection area of the forward network with the loss function preset by the first target detection model The loss function value of the labeled data is calculated by using the loss function preset by the second target detection model to calculate the output data of the forward network and the loss function value of the small fitting defect label data;

(65) When the loss function value converges or the number of iterations reaches the preset maximum number of iterations, stop the corresponding network model training and proceed to step (66), otherwise, update the weight of the FasterRCNN network with the corresponding weight update method, and repeat the steps (64) and (65);

(66) After stopping the training, input the verification set data into the FasterRCNN network to obtain the output results, and perform statistics on the recall rate and precision rate. If the recall rate and precision rate meet the preset conditions, the training ends and the trained The object detection model is stored, otherwise, it is retrained.

As a further optimization of the above solution, the loss function of the second target detection model includes a second category loss and a second position loss,

The second category loss uses a modified binary cross-entropy loss function as follows:

Among them, N represents the number of prediction frames output by the network, P _0i is the probability value that the prediction frame network output is a normal pin, P _1i is the probability value that the prediction frame network output is a defective pin, and y _0i = 1 indicates that the prediction frame is labeled is a normal pin, y _0i = 0 indicates an abnormal pin, y _1i = 1 indicates that the prediction box is marked as a defective pin, y _1i = 0 indicates a non-defective pin, g ⁰ and g ¹ indicate a normal pin and a defective pin, respectively Loss weight distribution;

The second position loss formula is as follows:

Among them, M represents the number of positive samples output by the network, that is, the number of prediction frames whose label value y _0i or y _1i is 1, ( _xi , y _i ) represents the position of the center point of the prediction frame, (w _i , h _i ) represents the prediction Frame size, (x _i ^gt , y _i ^gt ) represents the center point of the label frame, (w _i ^gt , h _i ^gt ) represents the size of the label frame; L ₁ (x) uses the SmoothL1 function, the formula is as follows:

g _i represents the weight distribution in the position loss, which is consistent with the weight distribution in the category loss, as follows:

As a further optimization of the above solution, the weight update methods of the first object detection model and the second object detection model both use the stochastic gradient descent method to optimize the loss function value.

Based on the above method for detecting defects of transmission line fittings based on cascade target detection, the present invention also provides a detection system for defects of transmission line fittings based on cascade target detection, including:

The data acquisition module is used to obtain and preprocess the images taken by inspection;

The connected area detection module uses the trained first target detection model to perform connected area detection on the inspection shot image, and cuts out the detected rectangular connected area;

A connected area screening module, configured to obtain the area size of the connected area, and use n connected areas whose area size satisfies a preset condition as images to be identified;

The fine fitting defect detection module is used to use the trained second target detection model to detect the fine fitting defect on the image to be recognized, and obtain the coordinates of the fine fitting defect on the image to be recognized;

The original image display module of small metal defects is used to display the small metal defects in the original image according to the mapping relationship between the coordinates of the image to be recognized and the coordinates of the original image.

As a further optimization of the above scheme, it also includes a data augmentation processing module, which is used to perform data augmentation processing on the collected transmission line images before training the first target detection model and the second target detection model, and the data generated by the augmentation processing The image and the collected original transmission line image are jointly used as a data set for training the first object detection model and the second object detection model.

As a further optimization of the above solution, the first target detection model and the second target detection model use a deep neural network based on a target detection algorithm.

The method for detecting defects of transmission line fittings based on cascade target detection of the present invention has the following beneficial effects:

1. The method for detecting defects of power transmission line fittings based on cascaded target detection of the present invention, for the identification and detection of small targets of fine and small fittings of power transmission lines, since there are too few samples that can be collected for the defect detection of small fittings, the method of first photographing is adopted. The image is generated through a confrontation network to increase the number of samples and avoid overfitting when directly using deep learning for recognition, resulting in unsatisfactory results during testing.

2. The method for detecting defects of transmission line fittings based on cascaded target detection of the present invention, for the identification and detection of small targets of small fittings on transmission lines, adopts the deep convolutional neural network of the cascaded target detection algorithm, and first detects the power transmission lines captured by the patrol inspection. The larger target in the line image, that is, the connection area, is identified by the first target detection model. Based on the recognition result of the first target detection model, the second target detection model is used to identify the defects of small fittings in the connection area, reducing a lot of invalid The feature extraction of image information allows the network to focus on key areas, greatly reducing the loss of effective image information, ensuring the quality of key pixel information, and significantly improving the detection accuracy of small metal defects.

Description of drawings

Fig. 1 is a block diagram of the overall process for detecting small hardware defects on the inspection image to be detected in the method for detecting defects of metal fittings of transmission lines based on cascaded target detection of the present invention;

Fig. 2 is a block diagram of the overall process of performing model parameter training on the sample data in the method for detecting defects in transmission line fittings based on cascaded target detection, the first target detection model and the second target detection model;

Fig. 3 is a block diagram of the procedure for training the parameters of the adversarial generation network model on the sample data in the method for detecting the defect of the transmission line fittings based on the cascaded target detection of the present invention;

Fig. 4 is a block diagram of the overall process of performing model parameter training on the first target detection model and the second target detection model on the sample data in the method for detecting defects of transmission line fittings based on cascaded target detection according to the present invention;

5 is a block diagram of the internal structure of the transmission line hardware defect detection system based on cascade target detection of the present invention;

Fig. 6 is an illustration of nesting and labeling images in the method for detecting defects in transmission line fittings based on cascade target detection according to the present invention;

Fig. 7 is an illustration of small fitting defects detected by the second target detection model in the method for detecting defects of fittings of transmission line based on cascade target detection of the present invention;

Fig. 8 is a diagram showing the small fitting defects identified in Fig. 7 in the original image in the method for detecting defects of fittings of transmission lines based on cascade target detection of the present invention.

Implementation

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

Referring to Figures 1-8, the present invention provides a method for detecting defects in transmission line fittings based on cascaded target detection. Considering that the defect targets of small fittings in transmission lines are too small, accounting for less than 5% of the original image, traditional deep learning images The feature extraction capability of the recognition method is insufficient, which leads to the fact that the detection effect of small fitting defects cannot meet the actual business needs. This solution proposes a two-stage detection strategy for the identification of small fitting fault defects, which can significantly improve the detection accuracy of small fitting defects, including:

Before using the cascaded target detection model to detect the defects of transmission line fittings, the training process of the target detection model used in the two-stage detection is carried out first. In addition, considering that when using the deep learning network model to process image recognition, a large number of Training samples, but for the defect detection of small fittings, it mainly includes defect categories such as missing pins and protruding pins, but no matter which category, there are too few samples that can be collected, and it is easy to overfit when used for deep learning, resulting in testing. The effect is not ideal. In this embodiment, before training the first target detection model and the second target detection model, data augmentation processing is performed on the collected transmission line image, and the image generated by the augmentation processing is combined with the original transmission line image. As a data set for training the first object detection model and the second object detection model.

In this solution, the data augmentation process adopts the trained generative confrontation network model, and the steps for the training of the generative confrontation network model are as follows:

(31) Establish a generator and a discriminator, set the loss function of the discriminator and the objective function of generating an adversarial network; the loss function of the discriminator is:

L(P)=-y*ln(p)+(y-1)*ln(1-p), wherein, p is the probability value output by the discriminant network; y represents the label value, and its value is 0 or 1;

The objective function of generating an adversarial network is:

V(D,G)＝E _X～pd(x) lgD(x)+E _X～pz(x) lg(1-D(x)), where,

x represents the discriminative network input, G represents the generating network, D represents the discriminative network, x ~ Pd(x) represents that x obeys the collected transmission line image data distribution Pd(x), that is, x comes from the real collected transmission line image, X ~ Pd(x) Pz(x) means that X obeys the distribution Pz(x) of the random noise signal z, that is, X is the generated sample, and E[·] represents the mathematical expectation.

(32) Input the random noise signal z into the generator to obtain generated samples, mark the collected transmission line images and generated samples, and then input them into the discriminator to distinguish between real samples and generated samples, and use the back propagation algorithm to adjust the discrimination The network parameters of the network maximize the objective function of the generative confrontation network to obtain an optimized discriminant network;

(33) Substituting the network parameters of the optimized discriminant network into the objective function of the generative confrontation network, using the backpropagation algorithm to adjust the network parameters of the generative confrontation network, so as to minimize the objective function of the generative confrontation network, and obtain the network parameters of the discriminant network after optimization, Then the optimized generation network is obtained;

(34) Determine whether the number of iterations reaches the preset maximum number of iterations, if it is less than the preset maximum number of iterations, repeat steps (32)-(34), otherwise stop training, and save the generated confrontation model that has been trained.

Then call the trained generative adversarial network, use the generator in it to perform data augmentation processing on the collected transmission line images, and obtain generated samples, and then use the generated samples generated by the augmentation processing and the collected original transmission line images together as the training first A data set of an object detection model and a second object detection model.

In this embodiment, the first target detection model and the second target detection model use a deep neural network based on a target detection algorithm. The target detection algorithm includes but is not limited to FasterRCNN, FPN, YoloV3, etc. In this solution, the FasterRCNN algorithm based on For the deep neural network, the training process for the first target detection model and the second target detection model includes:

(61) Nested labeling of the data set used for training the model, referring to Figure 6, first labeling the connection area of the transmission line in the image, and then labeling whether the small metal fittings have defects based on the connection area;

(62) Data division: divide the training set and verification set in proportion;

(63) Establish the first target detection model and the second target detection model, the forward network all adopts the basic FasterRCNN network, and set the corresponding loss function and weight update method;

(64) Use the training set with nested label data as input data to input the first target detection model and the second target detection model respectively, and calculate the output data and connection area of the forward network with the loss function preset by the first target detection model The loss function value of the labeled data is calculated by using the loss function preset by the second target detection model to calculate the output data of the forward network and the loss function value of the small fitting defect label data;

(65) When the loss function value converges or the number of iterations reaches the preset maximum number of iterations, stop the corresponding network model training and proceed to step (66), otherwise, update the weight of the FasterRCNN network with the corresponding weight update method, and repeat the steps (64) and (65);

(66) After stopping the training, input the verification set data into the FasterRCNN network to obtain the output results, and perform statistics on the recall rate and precision rate. If the recall rate and precision rate meet the preset conditions, the training ends and the trained The object detection model is stored, otherwise, it is retrained.

Among them, the loss function of the first target detection model includes the first category loss and the first position loss,

The first category loss uses a cross-entropy loss function as follows:

Among them, N represents the number of prediction frames output by the network, and the default is 2000, p _i represents the network output probability value of the connected area, y _i represents the label value, 1 indicates that the area is marked as a connected area, and 0 indicates the background area.

The first position loss formula is as follows:

Among them, M represents the number of positive samples output by the network, that is, the number of prediction frames whose label value y _i is 1, ( _xi , y _i ) represents the position of the center point of the prediction frame, (w _i , h _i ) represents the size of the prediction frame, (x _i ^gt , y _i ^gt ) represents the position of the center point of the label frame, (w _i ^gt , h _i ^gt ) represents the size of the label frame; L ₁ (x) uses the SmoothL1 function, and the formula is as follows:

The loss function of the second target detection model includes a second category loss and a second position loss,

The second category loss uses a modified binary cross-entropy loss function as follows:

Among them, N represents the number of prediction frames output by the network, which is 2000 by default, P _0i is the probability value that the prediction frame network output is a normal pin, P _1i is the probability value that the prediction frame network output is a defective pin, and y _0i = 1 is expressed as The prediction box is marked as a normal pin, y _0i = 0 means an abnormal pin, y _1i = 1 means that the prediction frame is marked as a defective pin, y _1i = 0 means a non-defective pin, g ⁰ and g ¹ mean normal Loss weight distribution of pins and defective pins, since the model focuses on defective pins, g ¹ is greater than g ⁰ , default g ⁰ =0.7, g ¹ =1.4.

The second position loss formula is as follows:

Among them, M represents the number of positive samples output by the network, that is, the number of prediction frames whose label value y _0i or y _1i is 1, ( _xi , y _i ) represents the position of the center point of the prediction frame, (w _i , h _i ) represents the prediction Frame size, (x _i ^gt , y _i ^gt ) represents the center point of the label frame, (w _i ^gt , h _i ^gt ) represents the size of the label frame; L ₁ (x) uses the SmoothL1 function, the formula is as follows:

g _i represents the weight distribution in the position loss, which is consistent with the weight distribution in the category loss, as follows:

The weight update methods of the first object detection model and the second object detection model both use the stochastic gradient descent method to optimize the value of the loss function.

When adopting the transmission line fittings defect detection method based on cascade target detection of the present invention for identifying and detecting small fittings defects of transmission lines, the following steps are adopted:

(11) First obtain the inspection shot image and perform preprocessing. Considering that the shot image needs to be input into the neural network model for identification, in this embodiment, the inspection shot image is normalized and denoised;

(12) Then use the trained first target detection model to detect the connected area of the patrol shot image, and cut out the detected rectangular connected area;

(13) Obtain the area size of the connected area, and use n connected areas whose area size meets the preset condition as the image to be identified;

Specifically, considering the research on the distribution characteristics of inspection image defects, the connection area with defects usually appears in the relatively large area of the foreground, and other areas can be ignored as the background. Therefore, the cut out rectangular connection area is obtained by obtaining The coordinates of the area are calculated, and the detected connected areas are sorted from large to small according to the area size, and the first n of the sorted connected areas are selected as images to be recognized. In this embodiment, n=3 is adopted, that is, the area The largest 3 connection areas are used to identify small hardware defects.

(14) Use the trained second target detection model to detect small metal defects on the image to be recognized, and obtain the coordinates of the small metal defects on the image to be recognized, see Figure 7;

(15) According to the mapping relationship between the coordinates of the image to be recognized and the coordinates of the original image, display the small hardware defects in the original image, see FIG. 8 .

Referring to Fig. 5, based on the above-mentioned detection method for metal fitting defects of transmission lines, the present invention also provides a detection system for detection of metal fitting defects of transmission lines based on cascade target detection, including:

The data acquisition module is used to obtain and preprocess the images taken by inspection;

The connected area detection module uses the trained first target detection model to detect the connected area of the patrol shot image, and cuts out the detected rectangular connected area; the training data of the first target detection model and the second target detection model The set is a data set after data augmentation processing based on the generative confrontation network, which increases the number of network model training samples, reduces the phenomenon of over-fitting in the deep learning of the target detection model, and improves the accuracy of the target detection of the network model. A target detection model and a second target detection model adopt a deep neural network based on a target detection algorithm.

A connected area screening module, configured to obtain the area size of the connected area, and use n connected areas whose area size satisfies a preset condition as images to be identified;

The fine fitting defect detection module is used to use the trained second target detection model to detect the fine fitting defect on the image to be recognized, and obtain the coordinates of the fine fitting defect on the image to be recognized;

The original image display module of small metal defects is used to display the small metal defects in the original image according to the mapping relationship between the coordinates of the image to be recognized and the coordinates of the original image.

The data augmentation processing module adopts the generative confrontation network, which is used to perform data augmentation processing on the collected transmission line images before training the first target detection model and the second target detection model, and combines the images generated by the augmentation processing with the collected The original transmission line images are jointly used as a data set for training the first object detection model and the second object detection model.

The present invention is not limited to the above-mentioned specific implementation manners, and various transformations made by those skilled in the art starting from the above-mentioned ideas without creative work all fall within the scope of protection of the present invention.

Claims (10)

1. The power transmission line hardware defect detection method based on cascade target detection is characterized by comprising the following steps of: comprising the following steps:

(11) Acquiring a patrol shooting image and preprocessing;

(12) Using a trained first target detection model to detect a connection area of the inspection shooting image, and cutting out the detected rectangular connection area;

(13) Acquiring the area size of the connecting areas, and taking n connecting areas with the area size meeting preset conditions as images to be identified;

(14) Performing fine hardware defect detection on the image to be identified by using a trained second target detection model, and obtaining coordinates of the fine hardware defect on the image to be identified;

(15) And displaying the fine hardware defects in the original image according to the mapping relation between the coordinates of the image to be identified and the coordinates of the original image.

2. The method for detecting the defects of the power transmission line hardware based on the cascading target detection according to claim 1, wherein the method comprises the following steps: before the first target detection model and the second target detection model are trained, data augmentation processing is carried out on the acquired power transmission line images, and the image generated by the augmentation processing and the acquired original power transmission line images are used as a data set for training the first target detection model and the second target detection model.

3. The method for detecting the defects of the power transmission line hardware based on the cascading target detection according to claim 2, wherein the method comprises the following steps: the data augmentation processing adopts a trained generation countermeasure network model, and the step of generating the countermeasure network model training is as follows:

(31) Establishing a generator and a discriminator, setting a loss function of the discriminator and generating an objective function of an countermeasure network;

(32) Inputting a random noise signal z into a generator to obtain a generated sample, marking the acquired power transmission line image and the generated sample, inputting the marked power transmission line image and the generated sample into a discriminator, discriminating a real sample and the generated sample, regulating network parameters of a discrimination network by using a back propagation algorithm, and maximizing a target function of a generated countermeasure network to obtain an optimized discrimination network;

(33) Substituting the obtained network parameters of the optimized discrimination network into a generated countermeasure network objective function, and adjusting the network parameters of the generated network by using a back propagation algorithm to minimize the generated countermeasure network objective function, so as to obtain the network parameters of the optimized discrimination network, and further obtain the optimized generated network;

(34) And (3) judging whether the iteration times reach the preset maximum iteration times, if so, repeating the steps (32) - (34), otherwise, stopping training, and storing the generated countermeasure model after training is completed.

4. The method for detecting the defects of the power transmission line hardware based on the cascading target detection according to claim 2, wherein the method comprises the following steps: the first target detection model and the second target detection model adopt a deep neural network based on a target detection algorithm.

5. The method for detecting the defects of the power transmission line hardware based on the cascading target detection according to claim 4, wherein the method comprises the following steps: the first target detection model and the second target detection model adopt a deep neural network based on a FaterRCNN algorithm, and the training process of the FaterRCNN network model comprises the following steps:

(51) Nesting and labeling a data set for training a model, firstly labeling a power transmission line connection region in an image, and then labeling whether a tiny hardware fitting has a defect or not based on the connection region;

(52) Dividing data: dividing a training set and a verification set in proportion;

(53) Establishing a first target detection model and a second target detection model, wherein a forward network adopts a basic FasterRCNN network, and a corresponding loss function and a weight updating method are set;

(54) Respectively inputting a first target detection model and a second target detection model by taking a training set carrying nested marking data as input data, calculating the loss function values of the output data of the forward network and the marking data of the connecting area by using a loss function preset by the first target detection model, and calculating the loss function values of the output data of the forward network and the marking data of the small hardware defects by using a loss function preset by the second target detection model;

(55) Stopping the corresponding network model training when the loss function value converges or the iteration number reaches the preset maximum iteration number, and performing step (56), otherwise, performing weight update on the FasterRCNN network by using a corresponding weight update method, and repeating steps (54) and (55);

(56) After training is stopped, inputting verification set data into a FaterRCNN network to obtain an output result, counting the recall ratio and the precision ratio, if the recall ratio and the precision ratio meet preset conditions, finishing training, storing a trained target detection model, and otherwise, retraining.

6. The method for detecting the defects of the power transmission line hardware based on the cascading target detection according to claim 5, wherein the method comprises the following steps: the loss function of the second object detection model includes a second class loss and a second location loss,

the second class loss employs a modified binary cross entropy loss function as follows:

wherein N represents the number of predicted frames output by the network, and P _0i To predict the probability value of the frame network output as a normal pin, P _1i To predict the probability value of the frame network output as a defect pin, y _0i =1 indicates that the prediction box is marked as a normal pin, y _0i =0 as abnormal pin, y _1i =1 indicates that the prediction box is marked as a defect pin, y _1i =0, denoted as non-defective pin, g ⁰ And g ¹ Loss weight distribution of the normal pin and the defect pin is respectively shown;

the second position loss formula is as follows:

where M represents the number of positive samples of the network output, i.e. the tag value y _0i Or y _1i Number of prediction frames of 1, (x) _i ,y _i ) Representing the predicted box center position, (w) _i ,h _i ) Representing prediction block size, (x) _i ^gt ,y _i ^gt ) Representing the position of the center point of the marking frame, (w) _i ^gt ,h _i ^gt ) Representing the size of the annotation frame; l (L) ₁ (x) The SmoothL1 function is used, with the following formula:

g _i the weight assignment in the representation location loss is consistent with the weight assignment in the category loss as follows:

7. the method for detecting the defects of the power transmission line hardware based on the cascading target detection according to claim 5, wherein the method comprises the following steps: the weight updating methods of the first target detection model and the second target detection model both adopt a random gradient descent method to optimize the loss function value.

8. The utility model provides a transmission line gold utensil defect detection system based on cascade target detection which characterized in that: comprising the following steps:

the data acquisition module is used for acquiring the inspection shooting image and preprocessing the inspection shooting image;

the connection area detection module is used for detecting the connection area of the inspection shooting image by using the trained first target detection model, and cutting the detected rectangular connection area;

the connection region screening module is used for acquiring the area size of the connection regions and taking n connection regions with the area size meeting preset conditions as images to be identified;

the fine hardware defect detection module is used for detecting the fine hardware defect of the image to be identified by using the trained second target detection model, and acquiring coordinates of the fine hardware defect on the image to be identified;

and the fine hardware defects original image display module is used for displaying the fine hardware defects on the original image according to the mapping relation between the coordinates of the image to be identified and the original image coordinates.

9. The cascading object detection-based transmission line hardware defect detection system as claimed in claim 8, wherein: the system further comprises a data augmentation processing module, wherein the data augmentation processing module is used for carrying out data augmentation processing on the acquired power transmission line image before the first target detection model and the second target detection model are trained, and the image generated by the augmentation processing and the acquired original power transmission line image are used as a data set for training the first target detection model and the second target detection model together.

10. The cascading object detection-based transmission line hardware defect detection system as claimed in claim 7, wherein: the first target detection model and the second target detection model adopt a deep neural network based on a target detection algorithm.

Images