CN115661117A - Contact net insulator visible light image detection method - Google Patents
Contact net insulator visible light image detection method Download PDFInfo
- Publication number
- CN115661117A CN115661117A CN202211406265.2A CN202211406265A CN115661117A CN 115661117 A CN115661117 A CN 115661117A CN 202211406265 A CN202211406265 A CN 202211406265A CN 115661117 A CN115661117 A CN 115661117A
- Authority
- CN
- China
- Prior art keywords
- insulator
- network
- feature
- image
- visible light
- 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
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a method for detecting visible light images of insulators of a contact network, belonging to the field of detection of insulators of the contact network, and the method comprises the steps of obtaining an image set of the insulators of the contact network according to the visible light images of the insulators of the contact network, and constructing a data set of the insulators of the contact network; according to the contact network insulator data set, constructing and training an insulator target detection model based on the improved RetinaNet to obtain an optimal insulator detection model; detecting the image to be detected by using the optimal insulator detection model to obtain a type prediction result and an anchor frame position parameter adjustment result; and adjusting the preset anchor frame containing the insulator sub-target of the contact network according to the type prediction result and the anchor frame position parameter adjustment result, outputting a contact network insulator target detection result image with the prediction frame, and completing detection of the visible light image of the contact network insulator. The invention solves the problems of poor detection effect and low positioning accuracy of the insulator sub-targets of the contact network.
Description
Technical Field
The invention belongs to the field of contact net insulator detection, and particularly relates to a contact net insulator visible light image detection method.
Background
The contact network system is an important component of the electrified railway, the quality of the state of the contact network system directly influences the traction power supply quality and the railway operation, and the reliability and the stability of the contact network system are guaranteed, so that the important part of the railway maintenance and operation is provided. The insulator is used as a key part in a contact net traction power supply system, plays the roles of electrical insulation and structural support, works in a natural environment for a long time, is easily influenced by external factors to cause various faults such as damage, dirt, aging and the like, and influences the normal operation of a railway system. Therefore, the insulator needs to be periodically overhauled and maintained, the abnormal state of the insulator is timely detected, the probability of accidents can be effectively reduced, and the economic loss caused by the accidents is reduced.
The manual detection method has been widely replaced by the automatic identification detection method due to the characteristics of low working efficiency, high risk and the like. At present, the detection method of the insulator is mainly divided into two types:
1. the conventional image processing method comprises the following steps: and extracting the characteristics of the image such as texture, edge, color and the like by using the artificially designed characteristic descriptors, and then matching and positioning the characteristics by template matching and the like.
2. The deep learning method comprises the following steps: the deep learning method based on the convolutional neural network replaces manual low-dimensional features by automatically capturing relatively abstract high-dimensional features of images, so that the learned features are more representative and contain more semantic information, and the precision and robustness of insulator target positioning are greatly improved.
At present, insulator detection research based on deep learning is less, a mainstream idea is that a classical deep learning detection model is directly used, the size and parameters of a preset detection frame are adjusted to carry out optimization, the characteristics of large length and width of an insulator data set and the like and an actual detection scene are mostly not considered, and the actual detection effect is general.
Disclosure of Invention
Aiming at the defects in the prior art, the contact net insulator visible light image detection method provided by the invention solves the problems of poor contact net insulator sub-target detection effect and low positioning accuracy.
In order to achieve the purpose of the invention, the invention adopts the technical scheme that: a contact net insulator visible light image detection method comprises the following steps:
s1, acquiring a contact net insulator visible light image, and preprocessing the contact net insulator visible light image to obtain a contact net insulator image set;
s2, constructing a contact net insulator data set according to the contact net insulator image set;
s3, constructing an insulator target detection model based on the improved RetinaNet;
s4, training an insulator target detection model based on the improved RetinaNet according to the contact network insulator data set to obtain an optimal insulator detection model;
s5, detecting an image to be detected according to the optimal insulator detection model to obtain a type prediction result and an anchor frame position parameter adjustment result;
s6, adjusting the preset anchor frame containing the contact net insulator sub-target according to the type prediction result and the anchor frame position parameter adjustment result, outputting a contact net insulator target detection result image with the prediction frame, and completing detection of the contact net insulator visible light image.
The invention has the beneficial effects that: according to the method, a large number of contact network insulator visible light image samples are collected, and aiming at the problems of complex and changeable shooting angles, illumination and background in the contact network insulator detection process, the contact network insulator image set is preprocessed and data is enhanced to obtain a contact network insulator data set, so that the generalization capability and the detection effect of a model are improved; in the improved feature extraction network G _ ResNet, a Ghost module is used for replacing conventional convolution operation, so that the feature extraction capability of the network is improved, and the network parameter quantity and the calculated quantity are reduced; by adding a CBAM attention mechanism, irrelevant target characteristics are inhibited, and the detection capability of the key target is improved; according to the shape characteristics of the contact net insulator, the original network is improved in a mode of increasing the angle parameters, so that the prediction frame is more accurately matched with the contact net insulator target, the background noise is reduced, and the detection effect is more accurate.
Further, the method for constructing the catenary insulation sub data set in step S2 includes the following steps:
s201, performing data enhancement operation on the contact network insulator image set to obtain an initial data set;
s202, labeling the catenary insulator sub-targets in the initial data set, and storing the type information and the position information of the catenary insulator sub-targets in a VOC data set format to obtain a catenary insulator data set.
The beneficial effects of the above further scheme are: through gathering a large amount of contact net insulator visible light image samples, to complicated changeable shooting angle, illumination and background problem among the contact net insulator testing process, carry out preliminary treatment and data enhancement to contact net insulator image set, obtain contact net insulator data set, improved the generalization ability and the detection effect of model.
Further, the data enhancement operation in step S201 includes rotating, flipping, brightness transforming, and contrast transforming the image in the contact line insulator image set.
The beneficial effects of the above further scheme are: the data enhancement operation is carried out on the contact net insulator image set to obtain a contact net insulator data set, and the generalization capability and the detection effect of the model are improved.
Further, the insulator target detection model based on the improved RetinaNet in the step S3 includes an improved feature extraction network G _ ResNet, a CBAM attention mechanism, a feature pyramid network FPN and a prediction network which are connected in sequence;
the improved feature extraction network G _ ResNet is used for stacking a residual bottleneck structure by using a Ghost module to obtain a stacking result, obtaining an initial feature map of the image to be detected based on the stacking result, and inputting the initial feature map into a CBAM attention mechanism;
the CBAM attention mechanism is configured to generate an attention weight suppression irrelevant target according to the initial feature map, obtain a feature map with attention weight, and input the feature map with attention weight to a feature pyramid network FPN, where a calculation process expression of the CBAM attention mechanism is as follows:
wherein F is an initial characteristic diagram; m c Operate for channel attention;is element-by-element multiplication; f' is the channel attention module output; m s For spatial attention operations; f' is a feature map with attention weight;
the feature pyramid network FPN is used for performing multi-scale feature fusion by utilizing upsampling and side edge connection according to the feature graph with the attention weight to obtain a fused feature graph, and inputting the fused feature graph into a prediction network;
and the prediction network is used for obtaining a type prediction result and an anchor frame position parameter adjustment result by utilizing multiple convolutions according to the fused feature map.
The beneficial effects of the above further scheme are: by replacing conventional convolution operation with a Ghost module in the improved feature extraction network G _ ResNet, the feature extraction capability of the network is improved, and the network parameter quantity and the calculated quantity are reduced; by adding a CBAM attention mechanism, irrelevant target features are inhibited, and the detection capability of a key target is improved; and performing multi-scale fusion in the feature pyramid network FPN, and acquiring a type prediction result and an anchor frame position parameter adjustment result through a prediction network to provide data support for accurate positioning of the contact net insulator sub-target.
Further, the improved feature extraction network G _ ResNet uses a Ghost module to perform residual bottleneck structure stacking, where a calculation process expression of the Ghost module is as follows:
wherein X is an image to be detected; f' is a convolution kernel; y' is an intrinsic characteristic diagram of the image to be detected; y is i ' is the ith intrinsic mapping of the intrinsic feature map of the image to be measured; phi (phi) of i,j To generate y i ' a linear transformation of the jth similar feature map; y is ij Is y i ' th and j-th similar feature maps.
The beneficial effects of the above further scheme are: in the improved feature extraction network G _ ResNet, a Ghost module is used for replacing simple convolution operation to obtain a feature map, stacking is carried out according to a residual bottleneck structure, and the number of parameters of the network is greatly reduced while the feature extraction capability is not reduced.
Further, the CBAM attention mechanism includes a channel attention module and a space attention module connected in sequence;
the channel attention module is used for endowing the initial feature map with attention weight on channel dimension to obtain a channel feature map, and inputting the channel feature map into the space attention module;
the spatial attention module is used for giving attention weight on spatial dimension to the channel feature map to obtain a feature map with attention weight.
The beneficial effects of the above further scheme are: the input initial characteristic diagram generates attention weights on the dimensions of a channel and a space respectively to obtain characteristic information needing to be strengthened or inhibited, and the detection capability of the insulator sub-target is improved.
Further, the prediction network comprises a classification sub-network and a location regression sub-network;
the classification sub-network is used for predicting the existence probability of the target object at each spatial position on the fused feature map by utilizing multiple convolutions to obtain a category prediction result;
and the position regression subnetwork is used for regressing the offset of each anchor frame on the fused feature map to a nearby real target by utilizing multiple convolutions to obtain an anchor frame position parameter adjustment result.
The beneficial effects of the above further scheme are: and respectively obtaining a type prediction result of the contact net insulation sub-target and an anchor frame position parameter adjustment result in the image to be detected through the classification sub-network and the position regression sub-network, and providing support for accurate positioning of the contact net insulation sub-target.
Further, the classification sub-network comprises 4 layers of convolution layers with 256 convolution kernels and 1 layer of convolution layers with K multiplied by A which are sequentially connected, wherein K is the number of types of catenary insulator sub-targets detected by the network, and A is the number of anchor frames generated on each feature point by the feature layer;
the position regression subnetwork comprises 4 convolution layers of 256 convolution kernels and 1 convolution layer of 5 multiplied by B which are connected in sequence, wherein 5 is 5 parameter adjustment states of an anchor frame, and B is the number of prior frames generated on each feature point by the feature layer.
The beneficial effects of the above further scheme are: the design of multilayer convolution enables the classification sub-network and the position regression sub-network to extract semantic information of a higher layer of the initial feature map, and enables a category prediction result and an anchor frame position parameter adjustment result to be more accurate.
Further, the loss function expression of the insulator target detection model based on the improved RetinaNet in step S3 is as follows:
wherein L is a loss function of the insulator target detection model based on the improved RetinaNet; n is the total number of the preset anchor frames; t is t n ' is foreground/background, t when anchor frame is foreground n Taking 1 as background t n ' take 0; (x, y, w, h, theta) are prediction box coordinates; (x, y) are coordinates of the center point of the prediction frame; (w, h) is the prediction frame length and short side; theta is an angle which is passed by the positive direction of the x axis clockwise rotating to the long edge of the prediction frame, and the value range is [0,180 ]; l is a radical of an alcohol reg Position regression loss; l is cls Is a classification loss; t is t ny ' is the difference between the y coordinate parameter of the n prediction frame and the corresponding parameter of the preset anchor frame; t is t ny Representing the difference between the y coordinate parameter of the nth preset anchor frame and the corresponding parameter of the real frame; p n Indicates the nth preset anchor frameA confidence probability distribution of; t is t n A category label of a contact net insulating sub-target corresponding to the nth preset anchor frame; lambda [ alpha ] 1 And λ 2 Are parameters that balance each loss weight.
The beneficial effects of the above further scheme are: the loss function of the insulator target detection model based on the improved RetinaNet comprises position regression loss and classification loss, the position regression loss is calculated through a smooth L1 function, the classification loss is calculated through a focusing loss function, and the insulator detection model with the optimal solution can be obtained.
Further, the anchor box in step S5 is represented by a rotating rectangular box, and the expression of the rotating rectangular box is:
B=(x,y,w,h,θ)
wherein, B is a rotating rectangular frame.
The beneficial effects of the above further scheme are: according to the shape characteristics of the contact net insulator, the original network is improved in a mode of increasing the angle parameters, so that the prediction frame is more accurately matched with the contact net insulator target, the background noise is reduced, and the detection effect is more accurate.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is an example of a visible light image of an insulator according to the present invention.
Fig. 3 is a schematic structural diagram of an insulator target detection model based on improved retinaNet in the invention.
Fig. 4 is a schematic structural diagram of a Ghost module in the present invention.
Fig. 5 is an explanatory view of a rotating rectangular frame in the present invention.
Fig. 6 is a schematic diagram of a prediction network structure according to the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.
As shown in fig. 1, in an embodiment of the present invention, the present invention provides a contact line insulator visible light image detection method, including the following steps:
s1, acquiring a contact net insulator visible light image, and preprocessing the contact net insulator visible light image to obtain a contact net insulator image set;
in this embodiment, as shown in fig. 2, a contact network insulator visible light image is obtained by automatically shooting at a site through a contact network detection device equipped with a camera, and the collected image is manually screened to remove a blurred sample with poor quality, so as to obtain a contact network insulator image set, wherein the contact network insulator image set contains a plurality of illumination conditions and contact network insulators under a plurality of contact network backgrounds (a forward positioning type, a reverse positioning type, a contact network insulator with a compensation device type, and a double positioning type).
S2, constructing a contact net insulator data set according to the contact net insulator image set, wherein the construction method comprises the following steps:
s201, performing data enhancement operation on the contact network insulator image set to obtain an initial data set;
s202, labeling the catenary insulator sub-targets in the initial data set, and storing the type information and the position information of the catenary insulator sub-targets in a VOC data set format to obtain a catenary insulator data set.
The data enhancement operation in step S201 includes rotating, flipping, brightness transforming, and contrast transforming the image in the contact network insulator image set.
In this embodiment, the data enhancement operation includes one or more of rotation, flipping, brightness conversion, and contrast conversion of the image in the contact grid insulator image set at random.
In the embodiment, roLabelImg software is used for accurately marking the insulator sub-targets of the contact network in the initial data set, and the target types and the position information of the insulators of the contact network are stored in a VOC data set format to obtain an insulator data set of the contact network; and (3) the obtained contact network insulator data set is as follows: the ratio of 2 is randomly divided into a training set and a test set.
S3, constructing an insulator target detection model based on the improved RetinaNet, wherein the loss function expression of the insulator target detection model based on the improved RetinaNet is as follows:
wherein L is a loss function of the insulator target detection model based on the improved RetinaNet; n is the total number of the preset anchor frames; t is t n ' is foreground/background, t when anchor frame is foreground n When' takes 1, as background, t n ' take 0; (x, y, w, h, θ) are the predicted box coordinates; (x, y) are coordinates of the center point of the prediction frame; (w, h) are the predicted frame length and short sides; theta is an angle which is passed by clockwise rotation of the positive direction of the x axis to the long edge of the prediction frame, and the value range is [0,180 ]; l is reg Is the position regression loss; l is cls Is a classification loss; t is t ny ' is the difference between the y-th coordinate parameter of the n-th prediction frame and the corresponding parameter of the preset anchor frame; t is t ny Representing the difference between the y coordinate parameter of the nth preset anchor frame and the corresponding parameter of the real frame; p n Representing the probability distribution of the confidence coefficient of the nth preset anchor frame; t is t n A category label of a contact net insulating sub-target corresponding to the nth preset anchor frame; lambda 1 And λ 2 Are parameters that balance the respective loss weights.
As shown in fig. 3, the insulator target detection model based on the improved retinet includes an improved feature extraction network G _ ResNet, a CBAM attention mechanism, a feature pyramid network FPN and a prediction network, which are connected in sequence;
the improved feature extraction network G _ ResNet is used for stacking a residual bottleneck structure by using a Ghost module to obtain a stacking result, obtaining an initial feature map of the image to be detected based on the stacking result, and inputting the initial feature map into a CBAM attention mechanism;
the CBAM attention mechanism is configured to generate an attention weight suppression irrelevant target according to the initial feature map, obtain a feature map with attention weight, and input the feature map with attention weight to a feature pyramid network FPN, where a calculation process expression of the CBAM attention mechanism is as follows:
wherein, F is an initial characteristic diagram; m c Operate for channel attention;are multiplied element by element; f' is the channel attention module output; m s For spatial attention operations; f' is a feature map with attention weight;
the feature pyramid network FPN is used for carrying out multi-scale feature fusion by utilizing upsampling and side edge connection according to the feature graph with the attention weight to obtain a fused feature graph, and inputting the fused feature graph into a prediction network;
and the prediction network is used for obtaining a type prediction result and an anchor frame position parameter adjustment result by utilizing multiple convolutions according to the fused feature map.
The improved feature extraction network G _ ResNet uses a Ghost module to stack a residual bottleneck structure, and the expression of the computing process of the Ghost module is as follows:
wherein X is an image to be detected; f' is a convolution kernel; y' is an intrinsic characteristic diagram of the image to be detected; y is i ' is the ith intrinsic mapping of the intrinsic feature map of the image to be measured; phi i,j To generate y i ' linear transformation of the jth similar feature map; y is ij Is y i ' th and j-th similar feature maps.
The CBAM attention mechanism comprises a channel attention module and a space attention module which are connected in sequence;
the channel attention module is used for giving attention weight on channel dimension to the initial feature map to obtain a channel feature map, and inputting the channel feature map into the space attention module;
the spatial attention module is used for giving attention weight on spatial dimension to the channel feature map to obtain a feature map with attention weight.
The prediction network comprises a classification sub-network and a position regression sub-network;
the classification sub-network is used for predicting the existence probability of the target object at each space position on the fused feature map by utilizing multiple convolutions to obtain a category prediction result;
and the position regression subnetwork is used for regressing the offset of each anchor frame on the fused feature map to a nearby real target by utilizing multiple convolutions to obtain an anchor frame position parameter adjustment result.
The classification sub-network comprises 4 layers of convolution layers with 256 convolution kernels and 1 layer of convolution layer with K multiplied by A which are sequentially connected, wherein K is the number of types of contact net insulating sub-targets detected by the network, and A is the number of anchor frames generated on each feature point by the feature layer;
the position regression subnetwork comprises 4 convolution layers of 256 convolution kernels and 1 convolution layer of 5 multiplied by B which are connected in sequence, wherein 5 is 5 parameter adjustment states of an anchor frame, and B is the number of prior frames generated on each feature point by the feature layer.
In this embodiment, the loss function of the insulator target detection model based on the improved RetinaNet includes position regression loss and classification loss, the position regression loss is calculated through a smooth L1 function, the classification loss is calculated through a focus loss function, and the weight with stable convergence of the loss curve and the minimum final convergence value is selected based on the optimal weight of the insulator target detection model based on the improved RetinaNet.
In this embodiment, as shown in fig. 3, the insulator target detection model based on the improved retinet includes an improved feature extraction network G _ ResNet, a CBAM attention mechanism, a feature pyramid network FPN, and a prediction network, which are connected in sequence; in the improved feature extraction network G _ ResNet, a Ghost module is used for replacing simple convolution operation to obtain a similar feature map, stacking is carried out according to a residual bottleneck structure, and the number of parameters of the network is greatly reduced while the feature extraction capability is not reduced.
In this embodiment, the CBAM attention mechanism includes a channel attention module and a space attention module connected in sequence, and the initial feature map generates attention weights in channel and space dimensions, respectively, to obtain feature information that needs to be enhanced or suppressed.
In this embodiment, the prediction network structure is shown in fig. 6, and includes a classification subnetwork and a location regression subnetwork; the classification sub-networks sequentially comprise 4 convolutions of 256 channels and 1 convolution of K multiplied by A, wherein K refers to the number of categories of contact network insulating sub-targets detected by the network, and A refers to the number of anchor frames generated on each feature point by the feature layer; the position regression subnetwork sequentially comprises 4 convolutions of 256 channels and 1 convolution of 5 multiplied by A, wherein 5 refers to 5 parameter adjustment states of an anchor frame, and A refers to the number of prior frames generated on each feature point by the feature layer; w is the width of the convolution kernel and H is the height of the convolution kernel.
S4, training an insulator target detection model based on the improved RetinaNet according to the contact network insulator data set to obtain an optimal insulator detection model;
s5, detecting an image to be detected according to the optimal insulator detection model to obtain a type prediction result and an anchor frame position parameter adjustment result;
the anchor frame is represented by a rotating rectangular frame, and the expression of the rotating rectangular frame is as follows:
B=(x,y,w,h,θ)
wherein, B is a rotating rectangular frame.
In this embodiment, the image to be detected is scaled to 640 × 640, and then the image is sent to the optimal insulator detection model for detection, where the prediction result of the optimal insulator detection model includes the prediction result of the type of the insulator sub-target and the adjustment result of the anchor frame position parameter.
In this embodiment, in the improved feature extraction network G _ ResNet, a Ghost module (the structure of which is shown in fig. 4) is used to obtain an intrinsic feature map and a similar feature map through 1 × 1 convolution and simple linear transformation (Conv in fig. 4 represents a convolution process, and Identity represents a simple linear transformation process), and the intrinsic feature map and the similar feature map are spliced to obtain an effect similar to that of a common convolution operation; then, after the length and the width of the image are compressed by the feature layers for 3 times, 4 times and 5 times respectively, outputting 3 initial feature maps C3, C4 and C5 with different scales; then adding a CBAM attention mechanism, and suppressing irrelevant targets through attention weights to obtain 3 feature maps C3', C4' and C5' with attention weights; the sizes of C3', C4' and C5' are 80 × 80, 40 × 40 and 20 × 20, respectively. Sending 3 feature graphs C3', C4' and C5' with attention weights into a feature pyramid network FPN, combining shallow layer high-resolution features and deep layer semantic features through upsampling and side edge connection, completing multi-scale feature fusion, and obtaining 5 effective feature layers P3, P4, P5, P6 and P7; the sizes of P3, P4, P5, P6 and P7 are 80 × 80, 40 × 40, 20 × 20, 10 × 10,5 × 5 respectively; p3, P4 and P5 are obtained by C3', C4' and C5 'combined upsampling and side connection, and P6 and P7 are obtained by upsampling on the basis of C5'.
In the embodiment, in order to more accurately fit the contact network insulator sub-targets, angle parameters are introduced into the prediction network, and the preset anchor frame and the prediction frame use rotating rectangular frames in the expression form of (x, y, w, h, theta); the rotating rectangular frame uses a long edge representation method, as shown in fig. 5, each parameter x, y, w, h, and theta respectively represents the horizontal and vertical coordinates of the center of the rectangular frame, and the angle through which the long edge and the short edge of the rectangular frame and the positive direction of the x-axis rotate clockwise to the long edge of the rectangular frame has a value range of [0,180 ].
S6, adjusting the preset anchor frame containing the contact net insulator sub-target according to the type prediction result and the anchor frame position parameter adjustment result, outputting a contact net insulator target detection result image with the prediction frame, and completing detection of the contact net insulator visible light image.
Claims (10)
1. A contact net insulator visible light image detection method is characterized by comprising the following steps:
s1, acquiring a contact net insulator visible light image, and preprocessing the contact net insulator visible light image to obtain a contact net insulator image set;
s2, constructing a contact net insulator data set according to the contact net insulator image set;
s3, constructing an insulator target detection model based on the improved RetinaNet;
s4, training an insulator target detection model based on the improved RetinaNet according to the contact network insulator data set to obtain an optimal insulator detection model;
s5, detecting an image to be detected according to the optimal insulator detection model to obtain a type prediction result and an anchor frame position parameter adjustment result;
s6, adjusting the preset anchor frame containing the contact net insulator sub-target according to the type prediction result and the anchor frame position parameter adjustment result, outputting a contact net insulator target detection result image with the prediction frame, and completing detection of the contact net insulator visible light image.
2. The method for detecting the visible light image of the insulator of the overhead line system according to claim 1, wherein the method for constructing the insulator subset of the overhead line system in the step S2 comprises the following steps:
s201, performing data enhancement operation on the contact network insulator image set to obtain an initial data set;
and S202, labeling the contact net insulation sub-targets in the initial data set, and storing the type information and the position information of the contact net insulation sub-targets in a VOC data set format to obtain a contact net insulator data set.
3. The method for detecting the visible light image of the contact network insulator according to claim 2, wherein the data enhancement operation in the step S201 includes rotating, turning, brightness conversion and contrast conversion of the image in the contact network insulator image set.
4. The contact network insulator visible light image detection method according to claim 1, wherein the insulator target detection model based on the improved RetinaNet in the step S3 comprises an improved feature extraction network G _ ResNet, a CBAM attention mechanism, a feature pyramid network FPN and a prediction network which are connected in sequence;
the improved feature extraction network G _ ResNet is used for stacking the residual bottleneck structure by using a Ghost module to obtain a stacking result, obtaining an initial feature map of the image to be detected based on the stacking result, and inputting the initial feature map into a CBAM attention mechanism;
the CBAM attention mechanism is configured to generate an attention weight suppression irrelevant target according to the initial feature map, obtain a feature map with attention weight, and input the feature map with attention weight to a feature pyramid network FPN, where a calculation process expression of the CBAM attention mechanism is as follows:
wherein F is an initial characteristic diagram; m c Operate for channel attention;is element-by-element multiplication; f' is the channel attention module output; m is a group of s For spatial attention operations; f' is a feature map with attention weight;
the feature pyramid network FPN is used for performing multi-scale feature fusion by utilizing upsampling and side edge connection according to the feature graph with the attention weight to obtain a fused feature graph, and inputting the fused feature graph into a prediction network;
and the prediction network is used for obtaining a type prediction result and an anchor frame position parameter adjustment result by utilizing multiple convolutions according to the fused feature map.
5. The method for detecting the visible light image of the contact net insulator according to claim 4, wherein the improved feature extraction network G _ ResNet uses a Ghost module to stack a residual bottleneck structure, and the computing process expression of the Ghost module is as follows:
wherein X is an image to be detected; f' is a convolution kernel; y' is an intrinsic characteristic diagram of the image to be detected; y is i ' is the ith intrinsic mapping of the intrinsic feature map of the image to be measured; phi i,j To generate y i ' a linear transformation of the jth similar feature map; y is ij Is y i ' th and j-th similar feature maps.
6. The contact net insulator visible light image detection method according to claim 4, wherein the CBAM attention mechanism comprises a channel attention module and a space attention module which are connected in sequence;
the channel attention module is used for giving attention weight on channel dimension to the initial feature map to obtain a channel feature map, and inputting the channel feature map into the space attention module;
the spatial attention module is used for giving attention weight on spatial dimension to the channel feature map to obtain a feature map with attention weight.
7. The visible light image detection method for the contact net insulator according to claim 4, wherein the prediction network comprises a classification sub-network and a position regression sub-network;
the classification sub-network is used for predicting the existence probability of the target object at each spatial position on the fused feature map by utilizing multiple convolutions to obtain a category prediction result;
and the position regression subnetwork is used for regressing the offset of each anchor frame on the fused feature map to a nearby real target by utilizing multiple convolutions to obtain an anchor frame position parameter adjustment result.
8. The method for detecting the visible light image of the contact network insulator according to claim 7, wherein the classification sub-network comprises 4 convolutional layers of 256 convolutional kernels and 1 convolutional layer of K x A, which are sequentially connected, wherein K is the number of the types of the contact network insulator sub-targets detected by the network, and A is the number of anchor frames generated on each feature point by the feature layer;
the position regression subnetwork comprises 4 convolution layers of 256 convolution kernels and 1 convolution layer of 5 multiplied by B which are connected in sequence, wherein 5 is 5 parameter adjustment states of an anchor frame, and B is the number of prior frames generated on each feature point by the feature layer.
9. The method for detecting the visible light image of the contact network insulator according to claim 1, wherein the loss function expression of the insulator target detection model based on the improved RetinaNet in the step S3 is as follows:
wherein L is a loss function of the insulator target detection model based on the improved RetinaNet; n is the total number of the preset anchor frames; t is t n ' as foreground/background, t when anchor frame is foreground n Taking 1 as background t n ' take 0; (x, y, w, h, theta) are prediction box coordinates; (x, y) are coordinates of the center point of the prediction frame; (w, h) are the predicted frame length and short sides; theta is an angle which is passed by clockwise rotation of the positive direction of the x axis to the long edge of the prediction frame, and the value range is [0,180 ]; l is a radical of an alcohol reg Is the position regression loss; l is a radical of an alcohol cls To categorical losses; t is t ny ' is the difference between the y coordinate parameter of the n prediction frame and the corresponding parameter of the preset anchor frame; t is t ny Representing the difference between the y coordinate parameter of the nth preset anchor frame and the corresponding parameter of the real frame; p is n Representing the probability distribution of the confidence coefficient of the nth preset anchor frame; t is t n A category label of a contact net insulating sub-target corresponding to the nth preset anchor frame; lambda [ alpha ] 1 And λ 2 Are parameters that balance each loss weight.
10. The contact network insulator visible light image detection method according to claim 1, wherein in the step S5, the anchor frame is represented by a rotating rectangular frame, and an expression of the rotating rectangular frame is as follows:
B=(x,y,w,h,θ)
wherein, B is a rotating rectangular frame.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211406265.2A CN115661117A (en) | 2022-11-10 | 2022-11-10 | Contact net insulator visible light image detection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211406265.2A CN115661117A (en) | 2022-11-10 | 2022-11-10 | Contact net insulator visible light image detection method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115661117A true CN115661117A (en) | 2023-01-31 |
Family
ID=85021863
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211406265.2A Pending CN115661117A (en) | 2022-11-10 | 2022-11-10 | Contact net insulator visible light image detection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115661117A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116524203A (en) * | 2023-05-05 | 2023-08-01 | 吉林化工学院 | Vehicle target detection method based on attention and bidirectional weighting feature fusion |
-
2022
- 2022-11-10 CN CN202211406265.2A patent/CN115661117A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116524203A (en) * | 2023-05-05 | 2023-08-01 | 吉林化工学院 | Vehicle target detection method based on attention and bidirectional weighting feature fusion |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110827251B (en) | Power transmission line locking pin defect detection method based on aerial image | |
Mayr et al. | Weakly supervised segmentation of cracks on solar cells using normalized L p norm | |
CN110717532A (en) | Real-time detection method for robot target grabbing area based on SE-RetinaGrasp model | |
CN113592822B (en) | Insulator defect positioning method for electric power inspection image | |
CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
CN112818969A (en) | Knowledge distillation-based face pose estimation method and system | |
CN110059539A (en) | A kind of natural scene text position detection method based on image segmentation | |
CN112950576B (en) | Power transmission line defect intelligent identification method and system based on deep learning | |
CN114743119A (en) | High-speed rail contact net dropper nut defect detection method based on unmanned aerial vehicle | |
CN113901928A (en) | Target detection method based on dynamic super-resolution, and power transmission line component detection method and system | |
CN115661117A (en) | Contact net insulator visible light image detection method | |
CN112419333A (en) | Remote sensing image self-adaptive feature selection segmentation method and system | |
CN114359619A (en) | Incremental learning-based power grid defect detection method, device, equipment and medium | |
CN113252701A (en) | Cloud edge cooperation-based power transmission line insulator self-explosion defect detection system and method | |
CN116612098A (en) | Insulator RTV spraying quality evaluation method and device based on image processing | |
Sridharan et al. | Deep learning-based ensemble model for classification of photovoltaic module visual faults | |
CN114445615A (en) | Rotary insulator target detection method based on scale invariant feature pyramid structure | |
Hao et al. | PKAMNet: a transmission line insulator parallel-gap fault detection network based on prior knowledge transfer and attention mechanism | |
CN116503398B (en) | Insulator pollution flashover detection method and device, electronic equipment and storage medium | |
CN116485802B (en) | Insulator flashover defect detection method, device, equipment and storage medium | |
CN111091533B (en) | Battery piece EL defect detection method based on improved SSD algorithm | |
CN116994161A (en) | Insulator defect detection method based on improved YOLOv5 | |
CN116563844A (en) | Cherry tomato maturity detection method, device, equipment and storage medium | |
CN116385950A (en) | Electric power line hidden danger target detection method under small sample condition | |
CN116029440A (en) | Ultra-short-term power prediction method and device for photovoltaic power station |
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 |