CN113569981A - Power inspection bird nest detection method based on single-stage target detection network - Google Patents
Power inspection bird nest detection method based on single-stage target detection network Download PDFInfo
- Publication number
- CN113569981A CN113569981A CN202110928645.1A CN202110928645A CN113569981A CN 113569981 A CN113569981 A CN 113569981A CN 202110928645 A CN202110928645 A CN 202110928645A CN 113569981 A CN113569981 A CN 113569981A
- Authority
- CN
- China
- Prior art keywords
- network
- bird nest
- detection
- method based
- picture
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 86
- 235000005770 birds nest Nutrition 0.000 title claims abstract description 48
- 235000005765 wild carrot Nutrition 0.000 title claims abstract description 48
- 238000007689 inspection Methods 0.000 title claims abstract description 21
- 238000000034 method Methods 0.000 claims abstract description 28
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000000605 extraction Methods 0.000 claims abstract description 9
- 230000008569 process Effects 0.000 claims description 22
- 238000010586 diagram Methods 0.000 claims description 7
- 230000001629 suppression Effects 0.000 claims description 7
- 244000000626 Daucus carota Species 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 230000002776 aggregation Effects 0.000 claims description 3
- 238000004220 aggregation Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 abstract description 4
- 238000012545 processing Methods 0.000 abstract description 3
- 238000005457 optimization Methods 0.000 abstract 1
- 230000006870 function Effects 0.000 description 11
- 230000004927 fusion Effects 0.000 description 6
- 239000012212 insulator Substances 0.000 description 6
- 230000004913 activation Effects 0.000 description 4
- 238000011176 pooling Methods 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 238000005520 cutting process Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000003708 edge detection Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- JUJWROOIHBZHMG-UHFFFAOYSA-N Pyridine Chemical group C1=CC=NC=C1 JUJWROOIHBZHMG-UHFFFAOYSA-N 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 210000003128 head Anatomy 0.000 description 2
- 230000005764 inhibitory process Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000003707 image sharpening Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
- UMJSCPRVCHMLSP-UHFFFAOYSA-N pyridine Natural products COC1=CC=CN=C1 UMJSCPRVCHMLSP-UHFFFAOYSA-N 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a power inspection bird nest detection method based on a single-stage target detection network, and relates to the technical field of computer vision and image detection. The method comprises the following steps: s1, training, S2 and detecting. According to the invention, a target detection algorithm based on deep learning is adopted, bird nest detection is carried out on an electric power tower picture, network model training is to input an electric power tower picture data set into a network, the network carries out feature extraction and network self weight optimization through data processing at an input end, and finally a network capable of accurately identifying bird nest features in the electric power tower picture is obtained.
Description
Technical Field
The invention relates to the technical field of computer vision and image detection, in particular to a power inspection bird nest detection method based on a single-stage target detection network.
Background
The fault detection is one of important applications of computer vision in the field of industrial manufacturing, can improve the production efficiency of factories, reduce or even replace human labor, and can effectively ensure the product quality. How to accurately and efficiently detect the bird nest of the power tower is a difficult point of current research, and at present, three research methods mainly exist, namely manual detection and image detection and classification by human eyes; one is to extract and classify the image features by using the traditional image detection method; one is to directly utilize a neural network to detect whether a power tower has a fault.
The manual detection method has the advantages of high concentration of human eyes for a long time, easy fatigue, low efficiency, high labor cost and high false detection rate. The main processes of the traditional image detection algorithm are data enhancement, edge detection, image segmentation, feature extraction and image classification. The data enhancement adopts the methods of gray level transformation enhancement, histogram enhancement, image sharpening and the like, the edge detection is divided into the detection of a space domain and a frequency domain, the edge detection operator of the space domain comprises a canny operator, a sobel operator, a Roberts operator and the like, and the detection method of the frequency domain comprises Fourier transformation, wavelet transformation, gabor transformation and the like. Also by the second moment, entropy, inverse moment, contrast, correlation, etc. in the grey histogram. However, the traditional image detection method is low in efficiency and low in precision, along with the development of a target detection algorithm of a deep learning technology, Girshick and the like propose an R-CNN algorithm, and deep learning is substituted into the application field of computer vision for the first time. He et al (HEKM, ZHANGXY, RENSQ, et al. spatial pyridine position in depth dependent network for visual retrieval [ J ]), proposed SPP-Net algorithm, solved the problem of object deformation caused by the candidate frame scaling to a uniform size, Girshick (GIRSHICK R. Fast R-CNN [ C// Proceedings of the IEEE International Conference on Computer Vision,2015: 176-. Such as SSD [ [6] Liu W, Angelelov D, Erhan D, et al.SSD: single shot multibox detector [ C ]// Proc of European Conference on computer vision.Amsterdam, Nederland: Springer,2016:21-37 ], etc., compared with the traditional image processing technology, the deep learning method has very high precision on the detection of large objects, but has the defects of easy false detection and missing detection on the detection of small-scale objects.
In summary, the development of a power inspection bird nest detection method based on a single-stage target detection network is still a key problem to be solved urgently in the technical field of computer vision and image detection.
Disclosure of Invention
The invention aims to provide a power inspection bird nest detection method based on a single-stage target detection network, which solves the problems in the background technology.
In order to achieve the purpose, the invention is realized by the following technical scheme: a power inspection bird nest detection method based on a single-stage target detection network comprises the following steps:
and S1, training.
And S2, detecting.
Further, in the operation step S1, the training process is to construct a network model for detecting the bird ' S nest, and input the electric tower picture data set with the bird ' S nest into the single-stage target detection network model for training, so as to finally obtain the neural network capable of accurately detecting the bird ' S nest.
Further, in operation S1, the training process includes three processes, i.e., collecting an electric tower image data set, preprocessing an image at a network input end, and extracting feature information.
Further, the collection of the electric tower image data set is that the unmanned aerial vehicle with the high-definition camera is used for fixed-point collection, the unmanned aerial vehicle sets a route and a shooting angle through a laser radar, the specific route and the shooting angle ensure the safety of the electric tower and the unmanned aerial vehicle, the shooting precision and quality of pictures are improved, and the shooting rule is adopted from far to near, from whole to local, from right to left, from top to bottom, and from the front to the back. The tower shooting content: the whole pole tower, the foundation, the tower head, the left phase insulator (including a ground wire support), the middle phase insulator string and the right phase insulator (including a ground wire support), the number of the collected pictures of the pole tower is not less than 6, and the defects are photographed independently.
Furthermore, the image preprocessing at the network input end is to perform data enhancement through a Mosaic data enhancement algorithm, the Mosaic data enhancement adopts 4 pictures, and the operations of random scaling, random cutting and random arrangement are performed and then splicing is performed, so that the Mosaic data enhancement algorithm enriches a data set, increases a plurality of small targets, and enables the network to have better robustness.
Further, the feature information extraction is the work of a Backbone and a Neck network structure, the network Backbone structure is Focus + CBL + CSP1_1+ CBL + CSP1_3+ CBL + CSP1_3+ CBL + SPP, the Focus has 32 convolution kernels, the purpose of the Focus is to perform slicing operation, so that an original 608 × 3 picture is finally changed into a 304 × 32 feature map, the CSPX _ Y network structure is composed of a convolution layer and Y residual error structures in X Resnet networks, the Resnet residual error structure increases the capability of network feature extraction and feature fusion, the CBL is composed of Conv + Bn + Leaky _ Relu activation functions, the SPP performs multi-scale feature fusion again by adopting the maximum pooling mode of 1 × 1, 5 × 5, 9 × 9 and 13 × 13, the obtained feature vector passes through the FPN + network, the Neck network structure performs multi-scale feature fusion downwards through the FPN + network structure, and the multi-layer FPN structure performs multi-scale feature fusion by self-sampling, and obtaining a feature map for prediction, wherein the PAN is a bottom-up feature pyramid, strong positioning features are transmitted by bottom-up transmission, feature aggregation is carried out from different detection layers, and feature maps with high electrical tower nest scores and corresponding weights are packaged into a weight file.
Further, in the operation step S2, the detection process is to perform data enhancement by performing histogram equalization on the picture to be detected, and then input the picture of the electric tower to be detected into the output layer of the network for detection, and in the detection process, it is necessary to determine whether the picture to be detected has the characteristics of the bird nest, and if the characteristics of the bird nest exist, the picture is marked as the picture with the bird nest, otherwise, the picture is marked as the picture without the bird nest.
Furthermore, the output layer is a process of performing classified output after obtaining feature vectors of pictures through a network, matching the weights of the learned fault feature maps with the images of the test set, obtaining a plurality of preliminary prediction frames by adopting GIoU _ Loss as a Loss function of the bounding box, and identifying some shielded and overlapped targets by using an NMS non-maximum suppression algorithm on the basis, wherein the generated prediction frames are more accurate, the GIoU target is equivalent to a penalty of adding a ground route and a closed packet formed by the prediction frames in the Loss function, and a penalty term of the GIoU is that the smaller the proportion of the area obtained by subtracting the union of the two frames from the closed packet in the closed packet is, the better the area is.
Further, the GIoU algorithm includes the steps of:
inputting: any two target boxes:
a is the group Truth, B is the prediction box;
and (3) outputting: a GIoU;
step one, finding a rectangular frame C with a minimum range covering A and B according to the target frames A and B;
step two, calculating IoU:
step three, calculating the GIoU:
and C is a closure of the A and B areas, a prediction box is generated through a GIoU _ Loss and NMS non-maximum value inhibition algorithm, and the prediction box is the position of the bird nest of the detection power tower.
The invention provides a power inspection bird nest detection method based on a single-stage target detection network. The method has the following beneficial effects:
(1) the single-stage target detection network model can accurately detect the bird nest in the picture of the power tower.
(2) According to the invention, GIoU _ Loss is used as a Loss function of the bounding box, and GIoU not only pays attention to the overlapped region of the prediction frames, but also pays attention to other non-overlapped regions, so that the contact ratio of the two regions can be better reflected.
(3) The prediction frame NMS non-maximum value suppression algorithm can avoid the influence of difficult detection of small objects caused by random initialization of the prediction frame;
(4) the algorithm used by the invention has limited requirements on the computing performance, can be transplanted to various computer equipment, and is suitable for application scenes with multiple target types and more complex targets.
Drawings
FIG. 1 is a flow chart of a power patrol bird nest detection method based on a single-stage target detection network;
FIG. 2 is a schematic diagram of a network structure of a single-stage target detection network in a power inspection bird nest detection method based on the single-stage target detection network;
FIG. 3 is a schematic structural diagram of Focus in a power inspection bird nest detection method based on a single-stage target detection network;
fig. 4 is a schematic structural diagram of a heck in the power inspection bird nest detection method based on a single-stage target detection network.
Detailed Description
The invention is described in further detail below with reference to the following figures and specific examples, but the invention is not limited thereto.
Example (b):
referring to fig. 1 to 4, a power inspection bird nest detection method based on a single-stage target detection network includes the following steps:
step one, a training process.
And step two, a detection process.
The training process is to construct a network model for detecting the bird nest, input an electric tower picture data set with the bird nest into the single-stage target detection network model for training, and finally obtain a neural network capable of accurately detecting the bird nest.
The training process comprises three processes of electric tower image data set acquisition, image preprocessing at the network input end and characteristic information extraction.
The electric tower image data set is collected by using an unmanned aerial vehicle with a high-definition camera to carry out fixed-point collection, the unmanned aerial vehicle with the high-definition camera can shoot pictures according to route fixed points, the unmanned aerial vehicle sets routes and shooting angles through a laser radar, the electric tower and the unmanned aerial vehicle are safe while the specific routes and the shooting angles ensure the safety of the electric tower and the unmanned aerial vehicle, the shooting precision and the quality of the pictures are improved, and the shooting rules are adopted from far to near, from whole to local, from right to left, from top to bottom, and from the front to the back. The tower shooting content: the whole, basis, tower head, left phase insulator (including the ground wire support), well looks insulator chain, right phase insulator (including the ground wire support) of shaft tower, 6 are no less than to the shaft tower collection photo, and the defect is taken a picture alone to the data set that will collect and obtain is according to 7: the ratio of 3 is divided into a training set and a test set.
The image preprocessing at the network input end is to perform data enhancement through a Mosaic data enhancement algorithm, wherein the Mosaic data enhancement adopts 4 pictures, the operations of random scaling, random cutting and random arrangement are performed, then splicing is performed, namely, the four pictures are randomly cut and then spliced to one picture to serve as training data, and the Mosaic data enhancement algorithm comprises the following specific steps: randomly reading four pictures from the data set each time; respectively carrying out operations of turning (turning left and right on the original picture), zooming (zooming the size of the original picture), and color gamut change (changing the brightness, saturation and hue of the original picture) on the four pictures; and placing the four pictures, cutting out the fixed areas of the four pictures in a matrix mode after the placement of the four pictures is completed, and splicing the four pictures to form a new picture. The Mosaic data enhancement algorithm enriches the data set, increases a plurality of small targets and enables the network to have better robustness.
The characteristic information extraction is the work of a Backbone and a Neck network structure, the network Backbone structure is Focus + CBL + CSP1_1+ CBL + CSP1_3+ CBL + CSP1_3+ CBL + SPP, the Focus has 32 convolution kernels, the purpose is to perform slicing operation, an original 608 x 3 picture is finally changed into a 304 x 32 characteristic diagram, the CSPX _ Y network structure is a FPN + PAN structure, the FPN network is a top-down structure, and the multi-scale characteristic information of a high layer is transmitted and fused in an up-sampling mode to obtain a characteristic diagram for prediction. The network comprises convolutional layers and Y residual error structures in X Resnet networks, the Resnet residual error structures increase the capability of network feature extraction and feature fusion, and the Resnet residual error structure network is calculated as follows:
inputting: data a generated by the l-th networkl；
And (3) outputting: a isl+2；
S1: convolving data generated by the l layer network with the l +1 layer network:
zl+1＝wl+1al+bl+1；
s2: carrying out nonlinear operation on the convolution result of the l +1 layer network:
al+1＝Relu(zl+1)；
s3: convolving data generated by the (l + 1) th layer network with the (l + 2) th layer network:
zl+2＝wl+2al+1+bl+2；
s4: adding the convolution result of the l + 2-th network and the data generated by the l-th network and performing nonlinear operation:
al+2＝Relu(zl+2+al)；
y＝al+2；
wherein, wlWeight parameter representing the l-th layer of the network, blOffset parameter representing the l-th layer of the network, alData output representing the l-th layer of the network, zlRepresents the convolution result of the l-th layer of the network, Relu being the activation function used by this residual network.
Relu(x＜0)＝0
Relu(x≥0)＝x；
The CBL consists of Conv + BN + Leaky _ relu activation functions, Conv is a convolution layer, Conv has 128 convolution kernels with the size of 3 x 3 and the step length of 2, BN is data batch normalization and can accelerate the convolution speed in training, and the BN algorithm comprises the following steps:
inputting: dividing the input data set into m small data sets: b ═ x1, x 2.., xm };
and (3) outputting: { yi＝BNγ,β(xi)}；
S1: calculating a data mean value:
s2: calculating the data variance:
s3: data normalization:
s4: training parameters gamma, beta, output y
The Leaky _ relu activation function is as follows:
yi＝xi if xi≥0
wherein, aiIs a fixed parameter in the (1 +∞) interval, SPP is a spatial pyramid pooling operation, which is mainly composed of Conv, maxporoling and concat, and is downsampled by Conv extraction feature output and maximum pooling of Gaussian kernels with 3 sizes of 5 × 5, 9 × 9 and 13 × 13, and the 3 maximum pooling operations are performedAnd splicing and fusing the obtained data, adding the spliced and fused data to the initial characteristics of the data, and returning the output to be consistent with the original input through Conv. After being processed by the FPN + PAN network, the FPN network is sent to an output end, the FPN network is from Top to bottom, the FPN comprises three parts of bottom-upstream path, Top-downstream path and lateralconnections,
Pi、Pi+1、...、Pi+n＝f(Ci、Ci+1、...、Ci+n)
wherein, CiThe method comprises the steps of inputting FPN (multi-scale feature vector), namely generating multi-scale feature vector in the previous step, outputting Pi after fusion, transmitting and fusing feature information with different resolutions generated in the previous step in an up-sampling mode to obtain a feature map for prediction, transmitting and transmitting strong positioning features from bottom to top by PAN (PAN), performing feature aggregation from different detection layers, packaging feature maps with high bird nest scores of different towers and corresponding weights into weight files by FPN + PAN structure shown in figure 3, and obtaining a network model capable of detecting corresponding features, wherein the model structure is shown in figure 2.
The detection process includes three processes of collecting an image data set of the electric tower, enhancing data, detecting characteristics and outputting.
The acquisition of the electric tower image data set is that the unmanned aerial vehicle with a high-definition camera is used for fixed-point acquisition, the unmanned aerial vehicle sets a route and a shooting angle through a laser radar, the shooting precision and the shooting quality of pictures are improved by a specific route and the shooting angle, the image to be detected is subjected to histogram equalization, and as the image is a color image, the color image needs to be divided into R, G, B three channels, the histogram equalization is respectively carried out, and then the results of the 3 channels are combined, and the algorithm is as follows:
inputting: each channel is an image to be transformed with r gray level;
and (3) outputting: fusing the images converted by each channel to obtain a new image DN;
s1, calculating the probability density with the gray value r:
s2, calculating r gray value mapping S after algorithm processing:
s3, fusing the S new images of the 3 channels:
DN＝0.2989*R+0.5870*G+0.1140*B；
where r is the gray level of the image to be transformed, L represents the number of gray levels of the entire image, T (r) is the gray level mapping function, s is the gray level mapping of the transformed image pixels with gray level r, Pr(r) is the probability density with a gray level of r, and w is the pseudo-variable of the integral.
The output layer is used for realizing the process of classified output after the characteristic vectors of the pictures are obtained by the network, the weight of the learned fault characteristic picture is matched with the images of the test set, the GIoU _ Loss is used as a Loss function of the bounding box to obtain a plurality of preliminary prediction frames, on the basis, some shielded and overlapped targets are identified by using an NMS non-maximum suppression algorithm, and the generated prediction frames are more accurate.
The specific implementation steps of the NMS non-maximum suppression algorithm are as follows: setting a confidence threshold of the target frame; arranging the candidate frame list according to the confidence degree in a descending order; selecting a frame with the highest confidence coefficient, adding the frame into an output list, and deleting the frame from the candidate frame list; calculating IoU values of the box with the highest confidence coefficient and all boxes in the candidate box list, and deleting the candidate boxes larger than the threshold value; and repeating the process until the candidate box list is empty, and returning to the output list.
The goal of GIoU is to add a penalty of a closure formed by a group truth and a prediction box to the loss function, and its penalty term is that the smaller the proportion of the area of the closure minus the union of the two boxes in the closure is, the better.
Further, the GIoU algorithm includes the steps of:
inputting: any two target boxes:
a is the group Truth, B is the prediction box;
and (3) outputting: a GIoU;
s1, finding a rectangular frame C with the minimum range covering A and B according to the target frames A and B;
s2, calculation IoU:
s3, calculating GIoU:
and C is a closure of the A and B areas, a prediction box is generated through a GIoU _ Loss and NMS non-maximum value inhibition algorithm, and the prediction box is the position of the bird nest of the detection power tower.
The single-stage target detection network model can accurately detect the bird nest in the picture of the power tower, GIoU _ Loss is used as a Loss function of the bounding box, GIoU not only pays attention to the overlapping area of the prediction frame, but also pays attention to other non-overlapping areas, the coincidence degree of the GIoU and the bounding box can be better reflected, the NMS non-maximum suppression algorithm of the prediction frame can avoid the influence of difficult detection of small objects caused by random initialization of the prediction frame, the algorithm used has limited requirements on the calculation performance, can be transplanted to various computer equipment, and is suitable for application scenes with multiple target types and complex targets.
The above is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, many variations and modifications can be made without departing from the inventive concept of the present invention, which falls into the protection scope of the present invention.
Claims (9)
1. A power inspection bird nest detection method based on a single-stage target detection network is characterized by comprising the following steps:
s1, training;
and S2, detecting.
2. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 1, characterized in that: in the operation step S1, the training process is to construct a network model for detecting the bird ' S nest, and input the electric tower picture data set with the bird ' S nest into the single-stage target detection network model for training, so as to finally obtain a neural network capable of accurately detecting the bird ' S nest.
3. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 1, characterized in that: in the operation step S1, the training process includes three processes, i.e., collecting an electric tower image data set, preprocessing images at the network input end, and extracting feature information.
4. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 3, characterized in that: the electric tower image data set collection is that fixed point collection is carried out with the unmanned aerial vehicle of taking the high-definition camera, and unmanned aerial vehicle sets for the airline and shoots the angle through laser radar.
5. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 3, characterized in that: the image preprocessing at the network input end is to perform data enhancement through a Mosaic data enhancement algorithm.
6. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 3, characterized in that: the characteristic information extraction is the work of a Backbone and a Neck network structure, the network Backbone structure is Focus + CBL + CSP1_1+ CBL + CSP1_3+ CBL + CSP1_3+ CBL + SPP, the Neck structure is an FPN + PAN structure, the FPN is a top-down structure, multi-scale characteristic information of a high layer is transmitted and fused in an up-sampling mode to obtain a characteristic diagram for prediction, the PAN is a characteristic pyramid from bottom to top, strong positioning characteristics are transmitted and transmitted in the bottom to top, characteristic aggregation is performed from different detection layers, and the characteristic diagram with high bird nest scores of different towers and corresponding weights are packaged into a weight file.
7. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 1, characterized in that: in the operation step S2, the detection process is to perform data enhancement by performing histogram equalization on the picture to be detected, and then input the picture of the electric tower to be detected into an output layer of the network for detection, and in the detection process, it is necessary to determine whether the picture to be detected has characteristics of a bird nest, if so, the picture is marked as a picture with a bird nest, otherwise, the picture is marked as a picture without a bird nest.
8. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 7, characterized in that: the output layer is used for realizing the process of classified output after the characteristic vectors of the pictures are obtained by a network, matching the weights of the learned fault characteristic pictures with the images of the test set, obtaining a plurality of preliminary prediction frames by adopting GIoU _ Loss as a Loss function of the bounding box, and identifying some shielded and overlapped targets by using an NMS non-maximum suppression algorithm on the basis.
9. The electric power inspection bird nest detection method based on the single-stage target detection network according to claim 8, characterized in that: the GIoU algorithm includes the steps of:
inputting: renTwo target boxes are meant:
a is the group Truth, B is the prediction box;
and (3) outputting: a GIoU;
step one, finding a rectangular frame C with a minimum range covering A and B according to the target frames A and B;
step two, calculating IoU:
step three, calculating the GIoU:
and C is a closure of the two areas A and B, and a prediction box is generated through a GIoU _ Loss and nms non-maximum suppression algorithm, wherein the prediction box is the position of the bird nest of the detection power tower.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110928645.1A CN113569981A (en) | 2021-08-13 | 2021-08-13 | Power inspection bird nest detection method based on single-stage target detection network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110928645.1A CN113569981A (en) | 2021-08-13 | 2021-08-13 | Power inspection bird nest detection method based on single-stage target detection network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113569981A true CN113569981A (en) | 2021-10-29 |
Family
ID=78171525
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110928645.1A Pending CN113569981A (en) | 2021-08-13 | 2021-08-13 | Power inspection bird nest detection method based on single-stage target detection network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113569981A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114333040A (en) * | 2022-03-08 | 2022-04-12 | 中国科学院自动化研究所 | Multi-level target detection method and system |
| CN115272850A (en) * | 2022-07-20 | 2022-11-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon BAB type brake adjuster pull rod head breaking fault identification method |
| CN117392572A (en) * | 2023-12-11 | 2024-01-12 | 四川能投发展股份有限公司 | Transmission tower bird nest detection method based on unmanned aerial vehicle inspection |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111104906A (en) * | 2019-12-19 | 2020-05-05 | 南京工程学院 | Transmission tower bird nest fault detection method based on YOLO |
| CN111582072A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation picture bird nest detection method combining ResNet50+ FPN + DCN |
| CN112464883A (en) * | 2020-12-11 | 2021-03-09 | 武汉工程大学 | Automatic detection and identification method and system for ship target in natural scene |
| CN112560675A (en) * | 2020-12-15 | 2021-03-26 | 三峡大学 | Bird visual target detection method combining YOLO and rotation-fusion strategy |
| CN113052834A (en) * | 2021-04-20 | 2021-06-29 | 河南大学 | Pipeline defect detection method based on convolution neural network multi-scale features |
| WO2021139069A1 (en) * | 2020-01-09 | 2021-07-15 | 南京信息工程大学 | General target detection method for adaptive attention guidance mechanism |
| CN113205103A (en) * | 2021-04-19 | 2021-08-03 | 金科智融科技(珠海)有限公司 | Lightweight tattoo detection method |
| CN113205063A (en) * | 2021-05-19 | 2021-08-03 | 云南电网有限责任公司电力科学研究院 | Visual identification and positioning method for defects of power transmission conductor |
-
2021
- 2021-08-13 CN CN202110928645.1A patent/CN113569981A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111104906A (en) * | 2019-12-19 | 2020-05-05 | 南京工程学院 | Transmission tower bird nest fault detection method based on YOLO |
| WO2021139069A1 (en) * | 2020-01-09 | 2021-07-15 | 南京信息工程大学 | General target detection method for adaptive attention guidance mechanism |
| CN111582072A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation picture bird nest detection method combining ResNet50+ FPN + DCN |
| CN112464883A (en) * | 2020-12-11 | 2021-03-09 | 武汉工程大学 | Automatic detection and identification method and system for ship target in natural scene |
| CN112560675A (en) * | 2020-12-15 | 2021-03-26 | 三峡大学 | Bird visual target detection method combining YOLO and rotation-fusion strategy |
| CN113205103A (en) * | 2021-04-19 | 2021-08-03 | 金科智融科技(珠海)有限公司 | Lightweight tattoo detection method |
| CN113052834A (en) * | 2021-04-20 | 2021-06-29 | 河南大学 | Pipeline defect detection method based on convolution neural network multi-scale features |
| CN113205063A (en) * | 2021-05-19 | 2021-08-03 | 云南电网有限责任公司电力科学研究院 | Visual identification and positioning method for defects of power transmission conductor |
Non-Patent Citations (2)
| Title |
|---|
| 梁志芳，王迎娜: "计算机在材料加工中的应用", 31 March 2012, 媒炭工业出版社, pages: 29 - 38 * |
| 欧进永;杨渊;时磊;周振峰;邱实;: "基于深度学习的输电线路杆塔鸟窝识别方法研究", 机电信息, no. 24, pages 27 - 28 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114333040A (en) * | 2022-03-08 | 2022-04-12 | 中国科学院自动化研究所 | Multi-level target detection method and system |
| CN115272850A (en) * | 2022-07-20 | 2022-11-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon BAB type brake adjuster pull rod head breaking fault identification method |
| CN117392572A (en) * | 2023-12-11 | 2024-01-12 | 四川能投发展股份有限公司 | Transmission tower bird nest detection method based on unmanned aerial vehicle inspection |
| CN117392572B (en) * | 2023-12-11 | 2024-02-27 | 四川能投发展股份有限公司 | Transmission tower bird nest detection method based on unmanned aerial vehicle inspection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113065558B (en) | Lightweight small target detection method combined with attention mechanism | |
| CN112884064B (en) | Target detection and identification method based on neural network | |
| CN112183788B (en) | Domain adaptive equipment operation detection system and method | |
| CN113569981A (en) | Power inspection bird nest detection method based on single-stage target detection network | |
| CN111046880A (en) | Infrared target image segmentation method and system, electronic device and storage medium | |
| CN107133943A (en) | A kind of visible detection method of stockbridge damper defects detection | |
| CN109034184B (en) | Grading ring detection and identification method based on deep learning | |
| CN109961398B (en) | Fan blade image segmentation and grid optimization splicing method | |
| CN110298281B (en) | Video structuring method and device, electronic equipment and storage medium | |
| CN109544501A (en) | A kind of transmission facility defect inspection method based on unmanned plane multi-source image characteristic matching | |
| CN110390308B (en) | Video behavior identification method based on space-time confrontation generation network | |
| CN112686304A (en) | Target detection method and device based on attention mechanism and multi-scale feature fusion and storage medium | |
| CN110111346B (en) | Remote sensing image semantic segmentation method based on parallax information | |
| CN110992378B (en) | Dynamic updating vision tracking aerial photographing method and system based on rotor flying robot | |
| CN112560619B (en) | Multi-focus image fusion-based multi-distance bird accurate identification method | |
| CN116385958A (en) | Edge intelligent detection method for power grid inspection and monitoring | |
| Zhu et al. | Towards automatic wild animal detection in low quality camera-trap images using two-channeled perceiving residual pyramid networks | |
| CN113962973A (en) | Power transmission line unmanned aerial vehicle intelligent inspection system and method based on satellite technology | |
| CN115861799A (en) | Light-weight air-to-ground target detection method based on attention gradient | |
| CN116452966A (en) | Target detection method, device and equipment for underwater image and storage medium | |
| Tsutsui et al. | Distantly supervised road segmentation | |
| CN114359167A (en) | Insulator defect detection method based on lightweight YOLOv4 in complex scene | |
| CN116309270A (en) | Binocular image-based transmission line typical defect identification method | |
| CN115810123A (en) | Small target pest detection method based on attention mechanism and improved feature fusion | |
| CN111160255B (en) | Fishing behavior identification method and system based on three-dimensional convolution network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |