CN115690730A - High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation - Google Patents
High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation Download PDFInfo
- Publication number
- CN115690730A CN115690730A CN202211159681.7A CN202211159681A CN115690730A CN 115690730 A CN115690730 A CN 115690730A CN 202211159681 A CN202211159681 A CN 202211159681A CN 115690730 A CN115690730 A CN 115690730A
- Authority
- CN
- China
- Prior art keywords
- picture
- foreign matter
- mask
- abnormal
- speed rail
- 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
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a high-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation. The method comprises the stages of single-classification pre-training, exception generation, single-classification fine tuning and testing. In the single-classification pre-training stage, training is only carried out on a normal data set of the overhead line system, a mask self-encoder is designed, and semantic features under a railway scene are learned from each frame of picture extracted from a monitoring video; in the anomaly generation stage, foreign matters in other scenes are transferred to a high-speed rail scene by Poisson fusion to generate catenary anomaly data; in the single-classification fine adjustment stage, fine adjustment is carried out on a mask self-encoder aiming at abnormal data; and in the testing stage, each contact net picture is subjected to abnormity detection by calculating an abnormity score. According to the method, real abnormal samples are not needed, the learned model has high detection accuracy and generalization aiming at multiple types of unknown foreign matters in multiple scenes, and the safety of the high-speed railway operation is effectively guaranteed.
Description
Technical Field
The invention belongs to the field of intelligent transportation, and relates to a high-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation, which can learn a normal mode from data lacking abnormal samples in a centralized manner and automatically identify various unknown foreign matters invading a contact net.
Background
With the continuous development of social economy, a high-speed railway becomes an important trip mode, and the prevention of emergencies occurring in the operation of high-speed railways becomes an important subject. In an actual operation environment, due to the influence of strong wind along the line, foreign matters (such as mulching films, plastic bags and kites are common) are easily attached to a high-speed rail contact net, and the power supply of an electric locomotive is influenced, so that the normal operation of the train is threatened.
At present, foreign matter net hanging events are mainly checked in China through a manual observation mode, railway line contact net videos shot and transmitted back by a vehicle-mounted camera based on a high-speed rail locomotive are visually recognized whether the contact net is attached with the foreign matters, the detection method is low in efficiency, a large amount of manpower is consumed, and the reliability still has a large promotion space.
In recent years, image processing technology based on a neural network is rapidly developed, target detection in the image processing technology plays a great role in the fields of video monitoring and automatic driving, and especially small target detection makes a major breakthrough. The development of the new technologies provides a powerful tool for detecting foreign matters in a high-speed rail contact network scene, the cost of manual observation can be greatly reduced, and the accuracy and timeliness of detection are improved.
However, the high-speed rail catenary foreign matter detection task has several challenges as follows compared with other detection tasks. (1) Because the net hanging event belongs to an emergency event, the available abnormal video samples are very limited, so that the common target detection task cannot be performed, and enough target samples are provided in the training process; (2) Foreign matters invaded by a high-speed rail contact network are diverse, and the high-speed rail contact network has targets which are obviously different from the background of the contact network, such as mulching films and kites, and also has branches and plastic bags which are not obviously small targets, so that various characteristics of abnormal samples need to be detected; (3) Due to the openness of the real world, foreign matter invasion which has never been seen before, such as Kongming lantern, unmanned aerial vehicle and the like, can occur, and therefore the detection model is required to have better generalization capability.
Disclosure of Invention
In consideration of the characteristics of few samples, multiple types and unknown foreign matter invasion existence in a high-speed rail contact network scene, the invention provides a high-speed rail contact network foreign matter detection method and system based on single classification and abnormal generation.
The purpose of the invention is realized by the following technical scheme:
the invention provides a high-speed rail contact net foreign matter detection method based on single classification and abnormal generation, which comprises the following steps:
s1, single-classification pre-training: extracting high-level semantic features of a normal image of the overhead line system by adopting a mask self-encoder, and pre-training the mask self-encoder based on normal image data of the overhead line system;
s2, exception generation: taking a normal image of the contact network as source domain data, taking a foreign matter image as target domain data, modeling according to a source domain background of the contact network, combining a foreign matter image of the target domain, and generating a foreign matter image data set of the contact network by utilizing Poisson fusion;
s3, single-classification fine adjustment: fine-tuning the mask self-encoder which is pre-trained by using the generated abnormal image data set;
and S4, in the testing stage, calculating the abnormal score of each contact net picture, and judging the contact net picture as abnormal when the score exceeds a threshold value.
Further, the step S1 specifically includes:
inputting a preprocessed RGB picture x under a normal scene of the high-speed rail contact network, splitting the RGB picture x into M lines and N columns of M multiplied by N sub-pictures, and utilizing a random mask matrix I to form an element R M×N Covering the subgraph, and aiming at the characteristic that most of floaters are concentrated above the graph in the scene of a high-speed rail contact network, designing the probability distribution of a random mask matrix as follows:
wherein I m,n =1 indicates that the subgraph numbered m, n, respectively, from top to bottom and from left to right is covered; for the covered subgraph, a mask feature learner is adopted to calculate the substitute value of each pixel:
whereinA substitution value, f, representing a pixel point with coordinates (a, b) in the picture p A mask feature mapping network representing a mapping from original pixel values to mask values, shared by the totality of pixel values; after the replacement of the pixel value is completed, the pixel value is processed by the encoder f e And decoder f d Composed reconstructed picture generator for generating reconstructed pictureThe optimization goal of mask autoencoder pre-training is:
wherein l rec A mask reconstruction error of a high-speed rail contact network image pixel level, D is a contact network training set only containing normal pictures, I a,b Whether the pixel points (a and b) are covered is indicated, when the pixel points (a and b) are 1, the pixel points are masked, the pixel points are indicated to pass through the mask, and when the pixel points are 0, the pixel points are indicated to not pass through the mask; e x~D Representing the mathematical expectation that the pictures obey the distribution of the training set; l oc For single classification error of contact net, lambda 1 For contact net single classification error l oc The weight of (c); h is contact net image passing through encoder f e To output of (c).
Further, the image preprocessing in step S1 specifically includes:
for original picture with width W and height H, X belongs to R 3×H×W Firstly, intercepting the upper half area sky background picture, then down-sampling the intercepted picture, compressing the size to 1/2 to form a preprocessed picture
Further, in the step S1, the mask feature learner uses a forward neural network; the encoder uses a ResNet18 encoding network, and the decoder uses a ResNet18 decoding network.
Further, in the step S2, generating a foreign object center coordinate in the catenary source area image according to the size of the foreign object; fusing a foreign object into an original image using Poisson fusion to form a foreign object image dataset D n 。
Further, in the step S2, the foreign matter picture is zoomed, and the picture of the high-speed rail catenary after the upper half area is cutWherein the randomly selected center point (i, j) satisfies:
w′<i<W-w′
h and W are the width and the height of the original image of the high-speed rail contact net, and W 'and H' are the width and the height of the zoomed foreign matter image; taking the central point (i, j) as a fusion center, adding foreign matters into the high-speed rail picture by Poisson fusion to form a contact net foreign matter image data set D n 。
Further, the step S3 specifically includes:
generating a foreign matter detection frame according to the size of the foreign matter and the central point, and generating a substitute mask value by using a mask feature learner obtained by pre-training aiming at the pixel value in the detection frame;
for the foreign matter picture after mask processing, f obtained by pre-training e ,f d Fine adjustment is carried out, and the optimization target is as follows:
wherein l rec′ Mask reconstruction error for high-speed rail catenary image pixel level, D n Is a catenary training set containing only abnormal images, I' a,b Whether the pixel points (a and b) are contained by the foreign matter detection frame or not is indicated, when the pixel points (a and b) are contained by the foreign matter detection frame, the pixel points are represented as being contained when the pixel points (a and b) are 1, and when the pixel points are not contained when the pixel points (a and b) are 0;representing a mathematical expectation when the pictures obey the distribution of the abnormal image training set; l oc′ For single classification error of contact net, lambda 2 For contact net single classification error l oc′ The weight of (c); h is mask contact net image passing coder f e Output of (E), E x~D Representing the mathematical expectation that the picture obeys the normal image training set distribution.
Further, in step S4, the process of calculating the abnormality score is as follows:
dividing the original picture into M multiplied by N sub-pictures, performing mask for T times according to the mask strategy in the step S1, performing pixel value replacement based on a mask characteristic learning device, and inputting the generated T pictures into an encoder f e And decoder f d Generating a corresponding reconstructed picture, wherein the final abnormal score A is as follows:
wherein h is i Andand respectively representing the encoder output characteristics and the decoder output reconstructed picture of the picture after the ith mask.
Further, in step S4, the abnormality score threshold τ is set according to the following formula:
whereinAverage value of abnormal scores, max, for training set of abnormal images x~D A x Maximum value of abnormal score of normal image training set, lambda 3 Weights for the training set of abnormal images; and when the abnormal score of the test picture exceeds a threshold value, judging the test picture as abnormal.
The invention provides a high-speed rail contact net foreign matter detection system based on single classification and abnormal generation, which comprises the following components:
single classification pre-training module: extracting high-level semantic features of a normal image of the overhead line system by adopting a mask self-encoder, and pre-training the mask self-encoder based on normal image data of the overhead line system;
an exception generation module: taking a normal image of the contact network as source domain data, taking a foreign matter image as target domain data, modeling according to a source domain background of the contact network, combining a foreign matter image of the target domain, and generating a foreign matter image data set of the contact network by utilizing Poisson fusion;
single-classification fine adjustment module: fine-tuning the mask self-encoder which is pre-trained by using the generated abnormal image data set;
an anomaly scoring module: and calculating the abnormal score of each contact net picture in the test stage, and judging the contact net picture as abnormal when the score exceeds a threshold value.
The invention has the beneficial effects that:
1. according to the invention, a single classification model is adopted for learning, foreign body samples are not required to be introduced in the training process, the condition that original data lack abnormal samples is met, the problem of sample imbalance is solved, and the accuracy of foreign body detection is greatly enhanced.
2. According to the invention, an abnormal data enhancement method is adopted, high-quality high-speed rail scene abnormal invasion overhead line system data is generated, and an effective detection model verification data set is provided.
3. The invention utilizes the mask self-encoder to enhance the semantic feature characterization capability in the support vector description method and greatly improves the generalization of the model to the detection of various foreign matters in the actual scene.
Drawings
Fig. 1 is a schematic diagram of a high-speed rail catenary foreign matter detection method based on single classification and abnormal generation according to an exemplary embodiment.
Fig. 2 is a schematic diagram of a structure of a mask self-encoder according to an exemplary embodiment.
FIG. 3 is a diagram of the effect of an anomaly data test provided by an exemplary embodiment.
Detailed Description
The embodiments of the present invention are illustrated in the drawings according to the principle of the technology, and should not be considered as all of the invention, and should not be considered as the limitation or limitation of the technical solution of the present invention.
As shown in fig. 1, an embodiment of the present invention provides a high-speed rail catenary foreign matter detection method based on single classification and abnormal generation, which is used for detecting catenary foreign matter and floating matter from a catenary monitoring video collected by a train-mounted camera, and includes the following steps:
s1, single-classification pre-training: and extracting high-level semantic features of the normal image of the contact network by adopting a mask self-encoder, and pre-training the mask self-encoder based on the normal image data of the contact network.
The flow of processing pictures by the mask autoencoder is as shown in fig. 2, which is as follows:
for original picture with width W and height H, X belongs to R 3×H×W Preprocessing is carried out, firstly, the sky background picture of the upper half area is intercepted, then the intercepted picture is subjected to down sampling, the size is compressed to 1/2, and a preprocessed picture is formed
Inputting a preprocessed RGB picture x under a normal scene of the high-speed rail contact network, splitting the RGB picture x into M lines and N columns of M multiplied by N sub-pictures, and utilizing a random mask matrix I to form an element R M×N Covering the subgraph, and designing random mask matrix probability distribution according to the characteristic that most of floaters are concentrated above the subgraph in a high-speed rail contact network scene as follows:
wherein I m,n =1 indicates that the subgraphs numbered m, n, respectively, from top to bottom and from left to right are covered.
In this embodiment, a picture x is split into 8 subgraphs of 2 rows and 4 columns, and the probability that each subgraph in the first row is covered is designed to be e -1 The probability of each subgraph in the second row being masked is
After the mask picture is determined, a forward neural network with an input dimension of 2 and an output dimension of 3 is used for generating a substitute value of the masked pixel, and the formula is as follows:
whereinA substitution value, f, representing a pixel point with coordinates (a, b) in the picture p A mask feature mapping network, representing the mapping of mask features from original pixel values to mask values, is shared by the totality of pixel values. The mask picture finally obtained is
Inputting mask picture into ResNet18 coding network f e And the intermediate characteristic h with 256 dimensions of output dimension belongs to R 256 。
Inputting intermediate features into ResNet18 decoding network f d Returning a reconstructed picture of the same size as the input picture
The optimization goals for mask autoencoder pre-training are:
wherein l rec A mask reconstruction error of a high-speed rail contact network image pixel level, D is a contact network training set only containing normal pictures, I a,b Whether the pixel points (a and b) are covered is indicated, when the pixel points (a and b) are 1, the pixel points are masked, the pixel points are indicated to pass through the mask, and when the pixel points are 0, the pixel points are indicated to not pass through the mask; e x~D Representing the mathematical expectation that the pictures obey the distribution of the training set; l oc For single classification error of contact net, lambda 1 For contact net single classification error l oc The weight of (c); h is mask contact net image passing coder f e To output of (c).
S2, exception generation: and taking a normal picture of the contact network as source domain data, taking a foreign matter picture as target domain data, modeling according to a source domain background of the contact network, combining a foreign matter image of the target domain, and generating a foreign matter image data set of the contact network by utilizing Poisson fusion.
Specifically, a foreign matter picture such as a balloon, a Kongming lantern, a plastic bag and the like is obtained, the foreign matter picture is zoomed, and the size w '× h' of the foreign matter picture is controlled to be within the range of 120 × 120 pixels.
High-speed rail contact net picture after cutting out the upper half areaWherein the randomly selected center point (i, j) satisfies:
w′<i<W-w′
taking the central point (i, j) as a fusion center, adding foreign matters into the high-speed rail picture by Poisson fusion to form a contact net foreign matter image data set D n 。
S3, single-classification fine adjustment: and utilizing the generated abnormal image data set to finely adjust the mask self-encoder after the pre-training is finished.
Specifically, a 120 × 120 alien detection frame is generated from the alien center point, and a mask feature learner obtained by pre-training is used to generate an alternative mask value for the pixel values in the detection frame.
For the foreign matter picture after mask processing, f obtained by pre-training e ,f d Fine adjustment is carried out, and the optimization target is as follows:
wherein l rec′ Mask reconstruction error for high-speed rail catenary image pixel level, D n To include only generation of abnormal imagesTraining set of contact net of' a,b Whether the pixel points (a and b) are contained by the foreign matter detection frame or not is indicated, when the pixel points (a and b) are contained by the foreign matter detection frame, the pixel points are represented as being contained when the pixel points (a and b) are 1, and when the pixel points are not contained when the pixel points (a and b) are 0;representing a mathematical expectation when the pictures obey the distribution of the abnormal image training set; l oc′ For single classification error of contact net, lambda 2 For contact net single classification error l oc′ The weight of (c); h is the contact net image passing through the encoder f e Output of (E), E x~D Representing the mathematical expectation that the picture obeys the normal image training set distribution.
And S4, in the testing stage, calculating the abnormal score of each contact net picture, and judging the contact net picture as a foreign matter when the score exceeds a threshold value.
Specifically, the test image is divided into 2 × 4 sub-images, masking is performed 10 times according to the masking policy in step S1, pixel value replacement is performed based on the mask feature learner, and the generated 10 pictures are input to the encoder f e And a decoder f d Generating a corresponding reconstructed picture, wherein the final abnormal score A is as follows:
wherein h is i Andand respectively representing the encoder output characteristic and the decoder output reconstructed picture of the picture after the ith mask.
The abnormality score threshold τ is set according to the following equation:
whereinAverage of anomaly scores for a training set of anomalous images,max x~D A x Maximum value of abnormal score of normal image training set, lambda 3 Weights for the training set of abnormal images; and when the abnormal score of the test picture exceeds a threshold value, judging the test picture as abnormal.
The invention also provides a high-speed rail contact net foreign matter detection system based on single classification and abnormal generation, which comprises:
single classification pre-training module: extracting high-level semantic features of a normal image of the overhead line system by adopting a mask self-encoder, and pre-training the mask self-encoder based on normal image data of the overhead line system;
an exception generation module: taking a normal image of the contact network as source domain data, taking a foreign matter image as target domain data, modeling according to a source domain background of the contact network, combining a foreign matter image of the target domain, and generating a foreign matter image data set of the contact network by utilizing Poisson fusion;
single-classification fine-tuning module: fine-tuning the mask self-encoder which is pre-trained by using the generated abnormal image data set;
an anomaly scoring module: and calculating the abnormal score of each contact net picture in the test stage, and judging the contact net picture as abnormal when the score exceeds a threshold value.
The implementation of each module refers to the steps of the high-speed rail contact net foreign matter detection method embodiment based on single classification and abnormal generation.
Fig. 3 is an abnormal data test effect diagram, and when foreign matters such as unmanned aerial vehicles, kites, plastic bags, balloons and the like appear above a high-speed rail overhead line system, the method can give out higher abnormal scores. The invention can detect various unknown foreign matters, meets the requirements of railway departments and effectively ensures the safety of railway operation.
The above embodiments are only preferred and feasible embodiments of the present invention, and are used for illustrating the technical solutions of the present invention, and not for limiting the protection scope of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, modifications may be made to the embodiments described in the foregoing embodiments, or some of the features may be replaced with equivalents without departing from the spirit and scope of the claims and their equivalents, and thus such modifications and substitutions are intended to be within the scope of the claims.
Claims (10)
1. A high-speed rail contact net foreign matter detection method based on single classification and abnormal generation is characterized by comprising the following steps:
s1, single-classification pre-training: extracting high-level semantic features of a normal image of the overhead line system by adopting a mask self-encoder, and pre-training the mask self-encoder based on normal image data of the overhead line system;
s2, exception generation: taking a normal picture of the contact network as source domain data, taking a foreign matter picture as target domain data, modeling according to a source domain background of the contact network, combining a foreign matter image of the target domain, and generating a foreign matter image data set of the contact network by utilizing Poisson fusion;
s3, single-classification fine adjustment: fine-tuning the mask self-encoder which is pre-trained by using the generated abnormal image data set;
and S4, in the testing stage, calculating the abnormal score of each contact net picture, and judging the contact net picture as abnormal when the score exceeds a threshold value.
2. The method for detecting the foreign matter in the high-speed rail catenary based on the single classification and the abnormal generation as claimed in claim 1, wherein the step S1 is specifically as follows:
inputting a preprocessed RGB picture x under a normal scene of the high-speed rail contact network, splitting the RGB picture x into M rows and N columns of sub-pictures with the number of M multiplied by N, and utilizing a random mask matrix I belonging to R M×N Covering the subgraph, and aiming at the characteristic that most of floaters are concentrated above the graph in the scene of a high-speed rail contact network, designing the probability distribution of a random mask matrix as follows:
wherein I m,n =1 indicates that the subgraphs numbered from top to bottom and from left to right, respectively m, n, are covered; for the covered subgraph, a mask feature learner is adopted to calculate the substitute value of each pixel:
whereinA substitution value, f, representing a pixel point with coordinates (a, b) in the picture p A mask feature mapping network representing a mapping from original pixel values to mask values, shared by the totality of pixel values; after the replacement of the pixel value is completed, the pixel value is processed by the encoder f e And decoder f d Composed reconstructed picture generator for generating reconstructed pictureThe optimization target of mask self-encoder pre-training is as follows:
wherein l rec A mask reconstruction error of a high-speed rail contact network image pixel level, D is a contact network training set only containing normal pictures, I a,b Whether the pixel points (a and b) are covered is indicated, when the pixel points (a and b) are 1, the pixel points are masked, the pixel points are indicated to pass through the mask, and when the pixel points are 0, the pixel points are indicated to not pass through the mask; e x~D Representing the mathematical expectation that the pictures obey the distribution of the training set; l. the oc For single classification error of contact net, lambda 1 For contact net single classification error l oc The weight of (c); h is contact net image passing through encoder f e To output (d).
3. The method for detecting the foreign matter in the overhead line system of the high-speed rail based on the single classification and the abnormal generation as claimed in claim 2, wherein the image preprocessing in the step S1 is specifically as follows:
4. The method for detecting the foreign matter in the overhead line system of the high-speed rail based on the single classification and the abnormal generation as claimed in claim 2, wherein in the step S1, the mask feature learner adopts a forward neural network; the encoder uses a ResNet18 encoding network, and the decoder uses a ResNet18 decoding network.
5. The high-speed rail contact network foreign matter detection method based on single classification and abnormal generation as claimed in claim 1, wherein in the step S2, foreign matter center coordinates are generated in a contact network source area image according to the size of the foreign matter; fusion of foreign objects into the original image using Poisson fusion, constituting a foreign object image dataset D n 。
6. The high-speed rail catenary foreign matter detection method based on single classification and abnormal generation as claimed in claim 5, wherein in the step S2, the foreign matter picture is zoomed, and the high-speed rail catenary picture after the upper half area is cut outWherein the randomly selected center point (i, j) satisfies:
w′<i<W-w′
h and W are the width and the height of the original image of the high-speed rail contact net, and W 'and H' are the width and the height of the zoomed foreign matter image; taking the central point (i, j) as a fusion center, adding foreign matters into the high-speed rail picture by Poisson fusion to form a contact net foreign matter image data set D n 。
7. The method for detecting the foreign matter in the high-speed rail catenary based on the single classification and the abnormal generation as claimed in claim 2, wherein the step S3 is specifically as follows:
generating a foreign matter detection frame according to the size of the foreign matter and the central point, and generating a substitute mask value by using a mask feature learner obtained by pre-training aiming at the pixel value in the detection frame;
for the foreign matter picture after mask processing, f obtained by pre-training e ,f d Fine adjustment is carried out, and the optimization target is as follows:
wherein l rec′ Mask reconstruction error for high-speed rail catenary image pixel level, D n Is a catenary training set containing only abnormal images, I' a,b Whether the pixel points (a and b) are contained by the foreign matter detection frame or not is indicated, when the pixel points (a and b) are contained by the foreign matter detection frame, the pixel points are represented as being contained when the pixel points (a and b) are 1, and when the pixel points are not contained when the pixel points (a and b) are 0;representing a mathematical expectation when the pictures obey the distribution of the abnormal image training set; l oc′ For single classification error of contact net,λ 2 For contact net single classification error l oc′ The weight of (c); h is mask contact net image passing coder f e Output of (E), E x~D Representing the mathematical expectation that the picture obeys the normal image training set distribution.
8. The method for detecting the foreign matters in the overhead line system of the high-speed rail based on the single classification and the abnormal generation as claimed in claim 2, wherein in the step S4, the process of calculating the abnormal score is as follows:
dividing the original picture into M multiplied by N sub-pictures, performing mask for T times according to the mask strategy in the step S1, performing pixel value replacement based on a mask characteristic learning device, and inputting the generated T pictures into an encoder f e And a decoder f d Generating a corresponding reconstructed picture, wherein the final abnormal score A is as follows:
9. The method for detecting the foreign matter in the overhead line system of the high-speed rail based on the single classification and the abnormal generation as claimed in claim 2, wherein in the step S4, the abnormal score threshold τ is set according to the following formula:
whereinAverage value of abnormal scores, max, for abnormal image training set x~D A x Training set for normal imagesMaximum value of abnormal score of (a), λ 3 Weights for the training set of abnormal images; and when the abnormal score of the test picture exceeds a threshold value, judging the test picture as abnormal.
10. The utility model provides a high-speed railway contact net foreign matter detecting system based on single classification and unusual formation which characterized in that includes:
single classification pre-training module: extracting high-level semantic features of a normal image of the overhead line system by adopting a mask self-encoder, and pre-training the mask self-encoder based on normal image data of the overhead line system;
an exception generation module: taking a normal image of the contact network as source domain data, taking a foreign matter image as target domain data, modeling according to a source domain background of the contact network, combining a foreign matter image of the target domain, and generating a foreign matter image data set of the contact network by utilizing Poisson fusion;
single-classification fine-tuning module: fine-tuning the mask self-encoder which is pre-trained by using the generated abnormal image data set;
an anomaly scoring module: and calculating the abnormal score of each contact net picture in the test stage, and judging the contact net picture as abnormal when the score exceeds a threshold value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211159681.7A CN115690730A (en) | 2022-09-22 | 2022-09-22 | High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211159681.7A CN115690730A (en) | 2022-09-22 | 2022-09-22 | High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115690730A true CN115690730A (en) | 2023-02-03 |
Family
ID=85063258
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211159681.7A Pending CN115690730A (en) | 2022-09-22 | 2022-09-22 | High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115690730A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116385807A (en) * | 2023-05-30 | 2023-07-04 | 南京信息工程大学 | Abnormal image sample generation method and device |
CN116933023A (en) * | 2023-09-14 | 2023-10-24 | 德电北斗电动汽车有限公司 | Monitoring method of opposed-piston magnetic force linear generator |
-
2022
- 2022-09-22 CN CN202211159681.7A patent/CN115690730A/en active Pending
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116385807A (en) * | 2023-05-30 | 2023-07-04 | 南京信息工程大学 | Abnormal image sample generation method and device |
CN116385807B (en) * | 2023-05-30 | 2023-09-12 | 南京信息工程大学 | Abnormal image sample generation method and device |
CN116933023A (en) * | 2023-09-14 | 2023-10-24 | 德电北斗电动汽车有限公司 | Monitoring method of opposed-piston magnetic force linear generator |
CN116933023B (en) * | 2023-09-14 | 2023-12-01 | 德电北斗电动汽车有限公司 | Monitoring method of opposed-piston magnetic force linear generator |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110781838B (en) | Multi-mode track prediction method for pedestrians in complex scene | |
CN110348376B (en) | Pedestrian real-time detection method based on neural network | |
CN115690730A (en) | High-speed rail contact net foreign matter detection method and system based on single classification and abnormal generation | |
CN114743119B (en) | High-speed rail contact net hanger nut defect detection method based on unmanned aerial vehicle | |
CN116229295A (en) | Remote sensing image target detection method based on fusion convolution attention mechanism | |
CN112766195B (en) | Electrified railway bow net arcing visual detection method | |
CN112070729A (en) | Anchor-free remote sensing image target detection method and system based on scene enhancement | |
CN109712127A (en) | A kind of electric transmission line fault detection method for patrolling video flowing for machine | |
CN111414807A (en) | Tidal water identification and crisis early warning method based on YO L O technology | |
CN115761537B (en) | Power transmission line foreign matter intrusion identification method oriented to dynamic feature supplementing mechanism | |
CN115546742A (en) | Rail foreign matter identification method and system based on monocular thermal infrared camera | |
CN117011883A (en) | Pedestrian re-recognition method based on pyramid convolution and transducer double branches | |
CN113298817A (en) | High-accuracy semantic segmentation method for remote sensing image | |
CN116052082A (en) | Power distribution station room anomaly detection method and device based on deep learning algorithm | |
CN112395953A (en) | Road surface foreign matter detection system | |
CN115116137A (en) | Pedestrian detection method based on lightweight YOLO v5 network model and space-time memory mechanism | |
Chen et al. | Foreign Object Detection in Railway Images Based on an Efficient Two‐Stage Convolutional Neural Network | |
CN118015562A (en) | Method and system for extracting key frames of traffic accident monitoring video in severe weather | |
CN114022727A (en) | Deep convolution neural network self-distillation method based on image knowledge review | |
CN116994161A (en) | Insulator defect detection method based on improved YOLOv5 | |
CN112907138A (en) | Power grid scene early warning classification method and system from local perception to overall perception | |
CN112837281A (en) | Pin defect identification method, device and equipment based on cascade convolutional neural network | |
CN112052829A (en) | Pilot behavior monitoring method based on deep learning | |
CN116977917A (en) | Infrared image pedestrian detection method | |
CN116453040A (en) | Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault |
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 |