CN110766011B - Contact net nut abnormity identification method based on deep multistage optimization - Google Patents
Contact net nut abnormity identification method based on deep multistage optimization Download PDFInfo
- Publication number
- CN110766011B CN110766011B CN201911363436.6A CN201911363436A CN110766011B CN 110766011 B CN110766011 B CN 110766011B CN 201911363436 A CN201911363436 A CN 201911363436A CN 110766011 B CN110766011 B CN 110766011B
- Authority
- CN
- China
- Prior art keywords
- nut
- area
- position information
- network
- image
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a contact net nut abnormity identification method based on deep multistage optimization, which comprises the following steps: acquiring a contact network image, detecting the contact network image by using the trained residual attention convolution neural network weight, and acquiring position information of a railway contact network nut area; intercepting the nut area from an original image according to the position information of the nut area, and inputting the nut area into a depth multi-stage self-coding network to obtain a result graph; and (3) carrying out binarization processing on the original image and the result image by using a maximum inter-class variance method, counting the difference degree between the original image and the result image by using a mean square error method, and judging whether the contact net nut is abnormal or not. The method has the advantages of less labor required, high identification speed, capability of effectively identifying whether the contact net nut is abnormal or not, great improvement of detection efficiency and guarantee of the power supply safety of the contact net.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to a contact net nut abnormity identification method based on depth multistage optimization.
Background
In recent years, the mileage of the high-speed railway in China exceeds 2.2 kilometers, and the high-speed railway is connected with various major cities, so that the high-speed railway can conveniently and quickly provide convenient transportation service for the public. However, with the laying of high-speed rail networks, the distribution of railway contact networks is wider and wider, and various fault detections therewith are also increasingly diversified, such as railway contact network nut abnormality and the like. Through the identification to the nut is unusual, can confirm whether the nut lacks to and the position of disappearance fast accurately to can play an important role in the maintenance of guaranteeing the normal work of railway contact net.
The detection and the discernment of railway contact net nut anomaly are usually traditional, all shoot the contact net through 4C image acquisition device at night earlier, judge whether have the unusual condition of nut through the mode that the analysis room staff looked the picture to record its information, this method efficiency is lower and miss detection or wrong detection phenomenon take place occasionally.
In recent years, some learners try to identify the nut abnormality of the railway contact system by using a method of target detection, but compared with the common target detection, the identification of the nut abnormality of the railway contact system has the following difficulties: 1) shooting angles of images are different, and nut area acquisition is incomplete; 2) the proportion of the nuts of the railway contact network in the whole image is small; 3) the position and angle difference of the nut of the railway contact network in different images is large; 4) the quality of the acquired 4C image is not stable. Therefore, the common target monitoring method cannot be directly applied to nut abnormity identification in the railway contact network 4C image. The characteristics of the nut in the image need to be fully researched according to the characteristics of the 4C image of the railway contact network, and the automatic nut identification technology of the railway contact network is designed.
Through the above description, how to accurately find out the position of the nut of the railway contact network and determine whether the nut is abnormal is an urgent problem to be solved.
Disclosure of Invention
The invention aims to overcome the related problem of the existing railway contact net nut abnormity identification, and the problem is solved in an efficient and accurate mode. According to the contact net nut abnormity identification method based on deep multistage optimization, the characteristic learning deep neural network which is automatically extracted by the small target area robustness is fully utilized to automatically detect the nut area, the image characteristic of whether the nut is abnormal or not is extracted by the automatic encoder, the multistage characteristics are effectively fused according to the weight of various characteristics automatically learned by the self-encoding network, the high-precision railway contact net nut abnormity identification result is obtained, and the method has a good application prospect.
In order to achieve the purpose, the invention adopts the technical scheme that:
a contact net nut abnormity identification method based on depth multistage optimization comprises the following steps:
acquiring a contact network image, detecting the contact network image by using a residual error attention convolution neural network obtained by training, and acquiring position information of a railway contact network nut area;
intercepting the nut area from an original image according to the position information of the nut area, inputting the nut area as an input value into a trained self-coding network, and processing the nut area through the self-coding network to obtain a result image;
and (C) binarizing the result graph after the self-coding network processing and the unprocessed original graph according to the maximum inter-class variance method, and then counting the difference degree between the result graph and the original graph to judge whether the nut is missing.
Preferably, the method comprises the following steps of (A) obtaining an image of the overhead line system, detecting the image by using a residual attention convolutional neural network obtained by training, and obtaining position information of a nut area of the railway overhead line system, wherein the method comprises the following steps:
(A1) selecting an area where the nut is located in the training sample picture through a manual frame, and obtaining position information of a target area where the nut is located in the training sample picture to be used as a label for storage;
(A2) scaling the training sample into a training sample matrix with a set pixel size, taking the training sample matrix and a label of corresponding position information as input, training a residual attention convolution neural network, and obtaining parameters of the corresponding neural network;
(A3) inputting the railway contact network image to a residual error attention convolution neural network, and calculating by using the trained network parameters to obtain the position information and the corresponding confidence coefficient of the nut in the railway contact network image;
(A4) and screening all the position information by using non-maximum suppression, screening out the position information with the confidence coefficient lower than a threshold value, and outputting the position information with the confidence coefficient higher than the threshold value.
Preferably, the step (A2) comprises the steps of:
(1) scaling an input training sample image into a characteristic diagram of 416 × 416, and performing random tone, brightness and angle adjustment on the characteristic diagram to serve as an input value of a network;
(2) using an expansion convolution kernel of 3 × 3, generating 13 × 13 multidimensional vectors by using input values through a 32-layer residual attention convolution neural network, dividing an original image into 169 meshes of 13 × 13, wherein each mesh outputs 5 pieces of position information and a confidence coefficient corresponding to each piece of position information, and each piece of position information comprises the position of the central point of the target and the length and width of the target;
(3) importing the position information of the step (2) into a loss function to obtain a loss value, performing back propagation on the residual attention convolution neural network according to the loss value, and updating the weight parameter;
(4) and when the loss value is lower than 0.06 and the detection accuracy rate of more than 95% is obtained in the test set, the corresponding weight parameter is stored as the optimal weight parameter.
Preferably, the random range of hue is 1 to 1.5 times, the random range of brightness is 1 to 1.5 times, and the random range of angle is 0 to 15 degrees.
The loss function of the preferred residual attention convolutional neural network is:
wherein,
is a loss value;
169 for the number of divided regions;
the number of target frames predicted for each region is 5;
is as followsWhether the j-th prediction frame of the i areas contains the target is 1 if the j-th prediction frame contains the target, and is 0 if the j-th prediction frame does not contain the target;
、
the abscissa and the ordinate of the center point of the jth prediction frame of the ith area are used as the coordinates;
、
the length and width of the jth prediction box of the ith area;
the abscissa and the ordinate of the real target in the ith area are shown;
the length and width of the real target in the ith area;
a prediction class confidence for the jth prediction box for the ith region;
the confidence of the prediction category of the real target of the ith area;
the j-th prediction frame of the i-th area is 1 if not included, and 0 if included.
Preferably, the step (B) of intercepting the nut area from the original image according to the position information of the nut area, inputting the intercepted nut area as an input value into the trained self-coding network, and obtaining a result graph after the input value is processed by the self-coding network includes the following steps:
(B1) intercepting a nut picture from an original image according to the position information of the nut area given in the step (A), scaling the nut picture into a set pixel size, and inputting the nut picture into a coding network as a training sample;
(B2) inputting training samples into an encoding network to be calculated with initial weight parameters, calculating results after convolutional encoding and transposed convolutional decoding processes with a training sample application loss function to obtain a loss value, updating the weight parameters by using a back propagation algorithm, stopping training when the loss value is lower than 0.03, wherein the training sample application loss function is as follows:
wherein x is an input picture, y is a picture after encoding and decoding, N is the number of elements in a group of data,
is a loss value;
(B3) and inputting the nut image to a self-coding network, and performing convolutional coding and transposed convolutional decoding by using the trained weight parameters to obtain a result graph.
Preferably, the step (C) of binarizing the result graph after the self-coding network processing and the unprocessed original graph according to the maximum inter-class variance method, and then counting the difference degree between the result graph and the original graph to judge whether the nut is missing comprises the following steps:
(C1) carrying out binarization processing on a result image output from the coding network and an unprocessed original image by using a maximum inter-class variance method to obtain a binarized image;
(C2) and counting the difference between the binarization result graph and the original graph by using a mean square error method, and if the difference is greater than a threshold value, determining that the nut is abnormal.
The invention has the beneficial effects that: the contact net nut abnormity identification method based on deep multistage optimization fully utilizes the characteristic learning residual attention convolution neural network automatically extracted by the small target region robust to automatically detect the nut region, adopts the automatic encoder to extract the image characteristic of whether the nut is abnormal or not, effectively fuses the results corresponding to different characteristics according to the probability characteristic weighting method to obtain the high-precision railway contact net nut abnormity identification result, is suitable for railway contact net nut abnormity identification, and has the following advantages:
(1) under the condition of unstable image quality, the accuracy of detection and identification is ensured in a mode of weighted fusion of multiple feature probabilities;
(2) the position of the nut area of the railway contact network is judged according to the possibility of each area possibly containing the nut without pre-positioning, so that the position of the nut is not omitted, and the area not containing the nut in the image can be effectively filtered;
(3) in the identification process, multi-level feature fusion is added, and the self-coding network automatically learns the weight of various features, so that the multi-level features are effectively fused to adapt to images acquired under various environments, and the identification precision is effectively improved.
Drawings
FIG. 1 is a flow chart of the method for identifying the abnormality of the nut of the contact net based on the deep multistage optimization;
FIG. 2 is a first simulation diagram for identifying the abnormality of the contact net nut of the invention;
FIG. 3 is a second simulation diagram for identifying the abnormality of the nut of the contact net of the invention;
fig. 4 is a third simulation diagram for identifying the abnormality of the contact net nut.
Detailed Description
The invention will be further described with reference to the accompanying drawings.
As shown in fig. 1, the method for identifying the abnormality of the nut of the contact network based on the deep multistage optimization comprises the following steps:
the step (A) of obtaining an image of a contact network and detecting the image by utilizing a residual error attention convolution neural network obtained by training to obtain the position information of a railway contact network nut area comprises the following steps,
(A1) manually framing the area where the nut is located in the training sample picture, wherein the framing range does not exceed the range of 10 pixel points outside the boundary of the nut image, and acquiring the position information of the target area where the nut is located in the training sample picture and storing the position information as a label;
(A2) the method comprises the following steps of scaling a training sample into a training sample matrix with a set pixel size, taking the training sample matrix and a label of corresponding position information as input, training a residual attention convolution neural network, and obtaining parameters of the corresponding neural network, wherein the method comprises the following steps:
(1) scaling an input training sample image into a characteristic diagram of 416 × 416, and performing random hue (1 time to 1.5 times), brightness (1 time to 1.5 times) and angle adjustment (0 degree to 15 degrees) on the characteristic diagram to serve as input values of the network;
(2) using an expansion convolution kernel of 3 × 3, generating 13 × 13 multidimensional vectors by using input values through a 32-layer residual attention convolution neural network, dividing an original image into 169 meshes of 13 × 13, wherein each mesh outputs 5 pieces of position information and a confidence coefficient corresponding to each piece of position information, and each piece of position information comprises the position of the central point of the target and the length and width of the target;
(3) importing the position information in the step (2) into a loss function to obtain a loss value, performing back propagation on the residual attention convolution neural network according to the loss value, and updating the weight parameter, wherein the loss function is as follows:
wherein,
is a loss value;
169 for the number of divided regions;
the number of target frames predicted for each region is 5;
whether the jth prediction frame of the ith area contains the target or not is 1 if the jth prediction frame contains the target, and is 0 if the jth prediction frame does not contain the target;
、
the abscissa and the ordinate of the center point of the jth prediction frame of the ith area are used as the coordinates;
、
the length and width of the jth prediction box of the ith area;
the abscissa and the ordinate of the real target in the ith area are shown;
the length and width of the real target in the ith area;
a prediction class confidence for the jth prediction box for the ith region;
the confidence of the prediction category of the real target of the ith area;
whether the jth prediction frame of the ith area does not contain the target or not is 1 if not, and is 0 if containing;
(4) when the loss value is lower than 0.06 and the detection accuracy rate is over 95 percent in the test set, the corresponding weight parameter is taken as the optimal weight parameter to be stored;
(A3) inputting the railway contact network image to a residual error attention convolution neural network, and calculating by using the trained network parameters to obtain the position information and the corresponding confidence coefficient of the nut in the railway contact network image;
(A4) all the position information is screened by using non-maximum suppression, the position information with the highest confidence coefficient is selected from a plurality of position information with the overlapping degree exceeding 0.5 and stored, the position information with the confidence coefficient lower than a threshold value (preferably, the threshold value is 0.6) is screened out, and the position information with the confidence coefficient higher than the threshold value is output.
Step (B), the nut area is intercepted from the original image according to the position information of the nut area, and after being scaled to 300X 300 image matrix, the nut area is input into the trained self-coding network as an input value, and a result graph is obtained after the self-coding network processing, the method comprises the following steps:
(B1) intercepting a nut picture from an original image according to the position information of the nut area given in the step (A), scaling the nut picture to 300 x 300 pixel size, and inputting the nut picture serving as a training sample into a self-coding network;
(B2) inputting training samples into an encoding network to be calculated with initial weight parameters, calculating results after convolutional encoding and transposed convolutional decoding processes with a training sample application loss function to obtain a loss value, updating the weight parameters by using a back propagation algorithm, stopping training when the loss value is lower than 0.03, wherein the training sample application loss function is as follows:
wherein x is an input picture, y is a picture after encoding and decoding, N is the number of elements in a group of data,
is a loss value;
(B3) and inputting the nut image to a self-coding network, and performing convolutional coding and transposed convolutional decoding by using the trained weight parameters to obtain a result graph.
And (C) binarizing a result graph after the self-coding network processing and an original graph which is not processed according to a maximum inter-class variance method, and then counting the difference degree between the result graph and the original graph to judge whether the nut is missing or not, wherein the method comprises the following steps:
(C1) and performing binarization processing on the result image output from the coding network and the unprocessed original image by using a maximum inter-class variance method to obtain a binarized image, wherein the maximum inter-class variance method has the following formula:
wherein g is the inter-class variance,
the proportion of the image pixels smaller than the threshold value T,
the proportion of image pixels larger than the threshold value T,
is the average gray value of the image pixels smaller than the threshold T,
the average gray value of the image pixel points is larger than the threshold value T;
the range of T is 0 to 255, when the value of T maximizes the inter-class variance g, the value of T is taken as a threshold value, pixels smaller than T are set to be 0, and pixels larger than T are set to be 255, so that a binary image is obtained;
(C2) counting the difference between the result graph of the binarization and the original graph by using a mean square error method, and if the difference is greater than a threshold value (preferably, the threshold value is 0.3), determining that the nut is abnormal, wherein the formula is as follows:
wherein MSE is the degree of difference, n is the total number of pixels,
for the ith pixel point of the binarization result graph,
the pixel is the ith pixel of the binary original image.
The contact net nut abnormity identification method based on deep multistage optimization, in a specific embodiment, such as the contact net nut abnormity identification simulation graphs of fig. 2-4, can display the detection and identification results of high-speed rail contact net nut abnormity, wherein the squares in fig. 2-4 are the results of nut abnormity identification, namely nut abnormity areas. According to the detection result, the method can effectively identify the nut abnormity.
In summary, the method for identifying the abnormality of the nut of the contact network based on the deep multi-stage optimization fully utilizes the feature learning deep neural network which is automatically extracted by the robustness of the small target region to automatically detect the nut region, adopts the automatic encoder to extract the image feature of whether the nut is abnormal or not, effectively fuses the results corresponding to different features according to the probability feature weighting method to obtain the high-precision identification result of whether the nut of the contact network of the railway is abnormal or not, is suitable for identifying the abnormality of the nut of the contact network of the railway, and has the following advantages:
(1) under the condition of unstable image quality, the accuracy of detection and identification is ensured in a mode of weighted fusion of multiple feature probabilities;
(2) the position of the nut area of the railway contact network is judged according to the possibility of each area possibly containing the nut without pre-positioning, so that the position of the nut is not omitted, and the area not containing the nut in the image can be effectively filtered;
(3) in the identification process, multi-level feature fusion is added, and the self-coding network automatically learns the weight of various features, so that the multi-level features are effectively fused to adapt to images acquired under various environments, and the identification precision is effectively improved.
The above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by using the contents of the present specification and the accompanying drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.
Claims (7)
1. The method for identifying the abnormity of the contact net nut based on the deep multistage optimization is characterized by comprising the following steps of:
the method comprises the following steps of (A) obtaining an image of the overhead line system, detecting the image by using a residual error attention convolution neural network obtained by training, and obtaining position information of a railway overhead line system nut area, wherein the method comprises the following steps:
(A1) selecting an area where the nut is located in the training sample picture through a manual frame, and obtaining position information of a target area where the nut is located in the training sample picture to be used as a label for storage;
(A2) the method comprises the following steps of scaling a training sample into a training sample matrix with a set pixel size, taking the training sample matrix and a label of corresponding position information as input, training a residual attention convolution neural network, and obtaining parameters of the corresponding neural network, wherein the method comprises the following steps:
(1) scaling an input training sample image into a characteristic diagram of 416 × 416, and performing random tone, brightness and angle adjustment on the characteristic diagram to serve as an input value of a network;
(2) using an expansion convolution kernel of 3 × 3, generating 13 × 13 multidimensional vectors by using input values through a 32-layer residual attention convolution neural network, dividing an original image into 169 meshes of 13 × 13, wherein each mesh outputs 5 pieces of position information and a confidence coefficient corresponding to each piece of position information, and each piece of position information comprises the position of the central point of the target and the length and width of the target;
(3) introducing the 5 position information obtained in the step (2) into a loss function to obtain a loss value, performing back propagation on the residual attention convolution neural network according to the loss value, and updating the weight parameter;
(4) when the loss value is lower than 0.06 and the detection accuracy rate is over 95 percent in the test set, the corresponding weight parameter is taken as the optimal weight parameter to be stored;
(A3) inputting the railway contact network image to a residual error attention convolution neural network, and calculating by using the trained network parameters to obtain the position information and the corresponding confidence coefficient of the nut in the railway contact network image;
(A4) screening all position information by using non-maximum suppression, screening out the position information with the confidence coefficient lower than a threshold value, and outputting the position information with the confidence coefficient higher than the threshold value;
intercepting the nut area from an original image according to the position information of the nut area, inputting the nut area as an input value into a trained self-coding network, and processing the nut area through the self-coding network to obtain a result image;
and (C) binarizing the result graph after the self-coding network processing and the unprocessed original graph according to the maximum inter-class variance method, and then counting the difference degree between the result graph and the original graph to judge whether the nut is abnormal.
2. The method for identifying the abnormity of the contact net nut based on the deep multistage optimization, according to claim 1, is characterized in that: the random range of hue is 1 to 1.5 times, the random range of brightness is 1 to 1.5 times, and the random range of angle is 0 to 15 degrees.
3. The method for identifying the abnormality of the nut of the overhead line system based on the deep multistage optimization, according to claim 1, is characterized in that a loss function of the residual attention convolution neural network is as follows:
wherein loss is a loss value; p is the number of divided regions, 169; b is the predicted number of target frames of each area, and is 5;
whether the jth prediction frame of the ith area contains the target or not is 1 if the jth prediction frame contains the target, and is 0 if the jth prediction frame does not contain the target; x is the number ofij、yijThe abscissa and the ordinate of the center point of the jth prediction frame of the ith area are used as the coordinates; w is aij、hijThe length and width of the jth prediction box of the ith area; xi、YiThe abscissa and the ordinate of the real target in the ith area are shown; wi、HiThe length and width of the real target in the ith area; c. CijA prediction class confidence for the jth prediction box for the ith region; ciThe confidence of the prediction category of the real target of the ith area;
the j-th prediction frame of the i-th area is 1 if not included, and 0 if included.
4. The method for identifying the abnormality of the nut of the overhead line system based on the deep multistage optimization, according to claim 1, wherein (a 4): the confidence is set to 0.6.
5. The method for identifying the abnormal nut of the contact network based on the deep multistage optimization as claimed in claim 1, wherein in the step (B), the nut area is intercepted from an original image according to the position information of the nut area, the nut area is input into a trained self-coding network as an input value, and a result graph is obtained after the self-coding network processing, and the method comprises the following steps:
(B1) intercepting a nut picture from an original image according to the position information of the nut area given in the step (A), scaling the nut picture into a set pixel size, and inputting the nut picture into a coding network as a training sample;
(B2) inputting training samples into an encoding network to be calculated with initial weight parameters, calculating results after convolutional encoding and transposed convolutional decoding processes with a training sample application loss function to obtain a loss value, updating the weight parameters by using a back propagation algorithm, stopping training when the loss value is lower than 0.03, wherein the training sample application loss function is as follows:
wherein x is an input picture, y is a picture after encoding and decoding, N is the number of elements in a group of data, loss*Is a loss value;
(B3) and inputting the nut image to a self-coding network, and performing convolutional coding and transposed convolutional decoding by using the trained weight parameters to obtain a result graph.
6. The method for identifying the abnormity of the nut of the contact net based on the deep multistage optimization, which is characterized in that the step (C) is used for binarizing a result graph after the self-coding network processing and an unprocessed original graph according to the maximum inter-class variance method, and then counting the difference degree between the result graph and the original graph to judge whether the nut is missing or not, and comprises the following steps:
(C1) carrying out binarization processing on a result image output from the coding network and an unprocessed original image by using a maximum inter-class variance method to obtain a binarized image;
(C2) and counting the difference between the binarization result graph and the original graph by using a mean square error method, and if the difference is greater than a threshold value, determining that the nut is abnormal.
7. The method for identifying the abnormity of the nut of the contact net based on the deep multistage optimization, which is characterized in that: (C2) and the threshold value is 0.3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911363436.6A CN110766011B (en) | 2019-12-26 | 2019-12-26 | Contact net nut abnormity identification method based on deep multistage optimization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911363436.6A CN110766011B (en) | 2019-12-26 | 2019-12-26 | Contact net nut abnormity identification method based on deep multistage optimization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110766011A CN110766011A (en) | 2020-02-07 |
| CN110766011B true CN110766011B (en) | 2020-04-28 |
Family
ID=69341600
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911363436.6A Active CN110766011B (en) | 2019-12-26 | 2019-12-26 | Contact net nut abnormity identification method based on deep multistage optimization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110766011B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111415326A (en) * | 2020-02-18 | 2020-07-14 | 中国铁道科学研究院集团有限公司 | Method and system for detecting abnormal state of railway contact net bolt |
| CN111553500B (en) * | 2020-05-11 | 2021-05-14 | 北京航空航天大学 | Railway traffic contact net inspection method based on attention mechanism full convolution network |
| CN111709936B (en) * | 2020-06-17 | 2023-10-24 | 广州麦仑智能科技有限公司 | Ream defect detection method based on multistage feature comparison |
| CN111563896B (en) * | 2020-07-20 | 2023-06-02 | 成都中轨轨道设备有限公司 | Image processing method for detecting abnormality of overhead line system |
| CN111951264B (en) * | 2020-08-28 | 2022-03-25 | 浙江中烟工业有限责任公司 | Cigarette package abnormity detection and positioning method based on deep learning |
| CN112528827B (en) * | 2020-12-03 | 2023-04-07 | 和远智能科技股份有限公司 | Automatic detection method for crack loss of high-speed rail contact network power supply equipment |
| CN112861948B (en) * | 2021-01-29 | 2024-04-09 | 南京中兴力维软件有限公司 | Method, device and equipment for identifying working status of siphon type mud discharge equipment |
| CN113012113B (en) * | 2021-03-01 | 2023-04-07 | 和远智能科技股份有限公司 | Automatic detection method for bolt looseness of high-speed rail contact network power supply equipment |
| CN113450357B (en) * | 2021-09-01 | 2021-12-17 | 南昌市建筑科学研究所（南昌市建筑工程质量检测中心） | Segment image online analysis subsystem and subway shield detection system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679250A (en) * | 2017-11-01 | 2018-02-09 | 浙江工业大学 | A kind of multitask layered image search method based on depth own coding convolutional neural networks |
| CN108509950A (en) * | 2018-03-27 | 2018-09-07 | 吴泽彬 | Railway contact line pillar number plate based on probability characteristics Weighted Fusion detects method of identification |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP4495849B2 (en) * | 2000-10-18 | 2010-07-07 | ハードロック工業株式会社 | Locking nut |
| CN106127749A (en) * | 2016-06-16 | 2016-11-16 | 华南理工大学 | The target part recognition methods of view-based access control model attention mechanism |
| CN110059601B (en) * | 2019-04-10 | 2021-04-13 | 西安交通大学 | An Intelligent Fault Diagnosis Method Based on Multi-feature Extraction and Fusion |
| CN110246132A (en) * | 2019-06-23 | 2019-09-17 | 中车青岛四方车辆研究所有限公司 | Rail vehicle bolt looseness detection method and system |
-
2019
- 2019-12-26 CN CN201911363436.6A patent/CN110766011B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679250A (en) * | 2017-11-01 | 2018-02-09 | 浙江工业大学 | A kind of multitask layered image search method based on depth own coding convolutional neural networks |
| CN108509950A (en) * | 2018-03-27 | 2018-09-07 | 吴泽彬 | Railway contact line pillar number plate based on probability characteristics Weighted Fusion detects method of identification |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110766011A (en) | 2020-02-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110766011B (en) | Contact net nut abnormity identification method based on deep multistage optimization | |
| CN108389197A (en) | Transmission line of electricity defect inspection method based on deep learning | |
| CN109740639B (en) | A method, system and electronic device for detecting cloud in remote sensing image of Fengyun satellite | |
| CN113947563B (en) | Cable process quality dynamic defect detection method based on deep learning | |
| CN109544501A (en) | A kind of transmission facility defect inspection method based on unmanned plane multi-source image characteristic matching | |
| CN113989257A (en) | Electric power comprehensive pipe gallery settlement crack identification method based on artificial intelligence technology | |
| CN111507235B (en) | Railway perimeter foreign matter intrusion detection method based on video | |
| CN115908407B (en) | Power equipment defect detection method and device based on infrared image temperature value | |
| CN111008641B (en) | Power transmission line tower external force damage detection method based on convolutional neural network | |
| CN116629465B (en) | Smart power grids video monitoring and risk prediction response system | |
| CN111767874A (en) | Pavement disease detection method based on deep learning | |
| CN111160432A (en) | Automatic classification method and system for panel production defects | |
| CN112785578A (en) | Road crack detection method and system based on U-shaped codec neural network | |
| CN113470316B (en) | A Debris Flow Monitoring and Early Warning Method Based on Self-encoding Single Classification Model | |
| Lam et al. | Real-time traffic status detection from on-line images using generic object detection system with deep learning | |
| CN111191610A (en) | A method for detecting and processing people flow in video surveillance | |
| CN116777839A (en) | Method and system for rapidly detecting power equipment based on infrared image and DETR | |
| CN116385368A (en) | A Dataset Augmentation Method for Photovoltaic Cell Defect Detection Based on Generative Adversarial Network | |
| CN116033119A (en) | High-low voltage line strong discharge monitoring system and method based on video image recognition | |
| CN115018777A (en) | Power grid equipment state evaluation method and device, computer equipment and storage medium | |
| CN112560574A (en) | River black water discharge detection method and recognition system applying same | |
| CN110782431A (en) | High-voltage wire icing area detection method based on deep learning | |
| CN117152665A (en) | Street lamp damage detection method and device and related components | |
| CN113191170B (en) | Public facility two-dimensional code damage inspection effectiveness prediction method and system | |
| CN117274154A (en) | Porcelain insulator string defect identification method and system based on discharge UV video |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |