CN110633668A - Railway shunting signal lamp detection method and system based on binary convolution neural network - Google Patents

Railway shunting signal lamp detection method and system based on binary convolution neural network Download PDF

Info

Publication number
CN110633668A
CN110633668A CN201910857618.2A CN201910857618A CN110633668A CN 110633668 A CN110633668 A CN 110633668A CN 201910857618 A CN201910857618 A CN 201910857618A CN 110633668 A CN110633668 A CN 110633668A
Authority
CN
China
Prior art keywords
neural network
binarization
network
training
signal lamp
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
Application number
CN201910857618.2A
Other languages
Chinese (zh)
Inventor
蔡永斌
朱玉虎
卫星
李百奇
盛典墨
洪予晨
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei Feiyang Electrical Co ltd
Hefei Locomotive Depot of China Railway Shanghai Group Co Ltd
Original Assignee
Hefei Feiyang Electrical Machinery Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hefei Feiyang Electrical Machinery Co Ltd filed Critical Hefei Feiyang Electrical Machinery Co Ltd
Priority to CN201910857618.2A priority Critical patent/CN110633668A/en
Publication of CN110633668A publication Critical patent/CN110633668A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
    • G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Multimedia (AREA)
  • Probability & Statistics with Applications (AREA)
  • Train Traffic Observation, Control, And Security (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a method and a system for detecting a railway shunting signal lamp based on a binary convolution neural network, which comprises the following steps: s1, preparing a data set, collecting a video in front of a train head acquired when a train runs, cutting the video into pictures with set sizes, and manually screening the pictures containing a target to obtain a target picture; then dividing the target picture into a training set and a testing set according to the set proportion of the blue light and the white light; s2, network construction, namely constructing a binary convolution neural network framework by utilizing a training set; s3, network training, namely training a binarization convolutional neural network by using a binarization method through a test set to obtain a target network; and S4, running the network, and carrying out real-time target detection on the shunting signal lamp of the railway by using the target network. The method carries out binarization aiming at the weight of the neural network to accelerate the operation of the neural network and reduce the memory consumption of the weight. The method has the advantages of high accuracy and reliability, and can accurately detect the signal lamp in front of the train.

Description

Railway shunting signal lamp detection method and system based on binary convolution neural network
Technical Field
The invention relates to the field of image detection, in particular to a railway shunting signal lamp detection method and system based on a binary convolution neural network.
Background
China has rapid railway development, and the freight volume of trains per year is continuously increased. Under the large environment of rapid development of railways, railway shunting accidents frequently occur, disastrous life, economic loss and serious social influence are caused, the number of shunting accidents occurring on the railways accounts for more than half of the total number of all traffic accidents, and the main reason is manual misjudgment of a crew on the state of a signal lamp.
The development of convolutional neural networks, which exhibit their reliable performance in the field of object recognition and detection, has been used in real-world applications, has greatly driven the development of tasks in the field of computer vision. Meanwhile, training the neural network model by using a large amount of image resources is also one of the key factors for the rapid development of the computer vision field. The existing deep neural networks based on different tasks have good effect when being deployed on a server. Full-precision CNN-based recognition systems require significant memory and computational resources, but they are generally not suitable for smaller devices such as cell phones and embedded electronic devices.
Disclosure of Invention
The technical problem to be solved by the invention is how to solve.
The invention solves the technical problems through the following technical means:
the railway shunting signal lamp detection method based on the binary convolution neural network comprises the following steps:
s1, preparing a data set, collecting a video in front of a locomotive head acquired when a train runs, cutting the video into pictures with set sizes, and manually screening the pictures containing a target to obtain a target picture; then dividing the target picture into a training set and a testing set according to the set proportion of the blue light and the white light;
s2, network construction, namely constructing a binary convolution neural network framework by utilizing a training set;
s3, network training, namely training a binarization convolutional neural network by using a binarization method through a test set to obtain a target network;
and S4, running the network, and carrying out real-time target detection on the shunting signal lamp of the railway by using the target network.
Preferably, the step S1 specifically includes:
s11, data collection, namely placing a camera on the locomotive head to obtain a video, cutting the collected video at a set frame rate, manually screening pictures containing targets (shunting signal lamps), and finally obtaining pictures containing signal lamps at a set amount;
s12, manufacturing a training set and a testing set, carrying out artificial target detection and labeling on the photos, and dividing the photos into the training set and the testing set according to the ratio of blue light to white light of 19: 1.
Preferably, the binarizing convolutional neural network framework in step S2 is specifically: firstly, generating a feature map based on a binary convolutional neural network of the convolutional neural network, extracting a regional prediction frame by using 9 candidate frames, and finally screening the most accurate target prediction frame containing a shunting signal lamp and an identification result by using non-maximum suppression, wherein the front end of the model uses a residual error network to extract features, the rear end of the model generates feature maps with different sizes through convolution, and the convolutional layer is divided into two scales according to the size.
Preferably, the step S3 specifically includes: and (3) performing binarization operation on the convolution layer parameters by adopting a determination method:
Figure BDA0002195723640000021
wherein W is a real weight, WbThe weight after binarization has only two values of '+ 1', '1';
in order to limit the convolutional neural network to have the binarization weight, this embodiment uses a binary convolution kernel and a scale parameter to approximately replace the original convolution kernel, and performs binarization operation according to equation (2):
Figure BDA0002195723640000022
wherein A x W represents the operation of a convolution layer, A represents the input, the dimension is C x win x hin, W represents the convolution kernel, the dimension is C x W x h, C is a binary convolution kernel, k is a scale parameter,
Figure BDA0002195723640000023
is calculated for convolution without multiplication.
To satisfy equation (2), a and k should be optimized:
A*＝argmax{WTA}s.tA∈{+1，-1}nformula (3)
Namely: a. the*Sign (w) formula (4)
Figure BDA0002195723640000024
Figure BDA0002195723640000025
Wherein A is*Is the optimal solution of A, k*Is the optimal solution of k, | W | | non-woven phosphorl1Representing the L1 norm.
Preferably, the step S3 specifically includes: in the process of training the binary convolution neural network, the weight is binarized only in forward propagation and backward propagation, binarization is used when the forward propagation is carried out for the first time, binarization is not carried out during the backward propagation and parameter updating, and a full-precision value is used; the inverse gradient formula is as follows:
Figure BDA0002195723640000031
preferably, in the calculation process of the gradient descent process, a random gradient descent method is adopted, and the following operations are performed:
Figure BDA0002195723640000032
and when the fastest descending direction is calculated, randomly selecting one data to calculate.
The invention also provides a railway shunting signal lamp detection system based on the binary convolution neural network, which comprises
The data set preparation module is used for collecting a video in front of a train head acquired when a train runs, cutting the video into pictures with set sizes, and manually screening the pictures containing a target to obtain a target picture; then dividing the target picture into a training set and a testing set according to the set proportion of the blue light and the white light;
the network construction module is used for constructing a binary convolutional neural network framework by utilizing the training set;
the network training module is used for training a binary convolution neural network by using a binary method through a test set to obtain a target network;
and the network operation module is used for carrying out real-time target detection on the railway shunting signal lamp by utilizing a target network.
Preferably, the data set preparation module comprises:
the data collection unit is used for placing a camera on the locomotive head to obtain a video, cutting the collected video at a set frame rate, manually screening pictures containing targets (signal lamps), and finally obtaining pictures containing the signal lamps at a set amount;
and the production units of the training set and the test set are used for carrying out artificial target detection and marking on the photos, and dividing the photos into the training set and the test set according to the ratio of blue light to white light of 19: 1.
Preferably, the specific training method of the network training module is as follows: and (3) performing binarization operation on the convolution layer parameters by adopting a determination method:
Figure BDA0002195723640000033
wherein W is a real weight, WbThe weight after binarization has only two values of '+ 1', '1';
in order to limit the convolutional neural network to have the binarization weight, this embodiment uses a binary convolution kernel and a scale parameter to approximately replace the original convolution kernel, and performs binarization operation according to equation (2):
Figure BDA0002195723640000041
wherein A x W represents the operation of a convolution layer, A represents the input, the dimension is C x win x hin, W represents the convolution kernel, the dimension is C x W x h, C is a binary convolution kernel, k is a scale parameter,is calculated for convolution without multiplication.
To satisfy equation (2), a and k should be optimized:
A*＝argmax{WTA}s.tA∈{+1，-1}nformula (3)
Namely: a. the*Sign (w) formula (4)
Figure BDA0002195723640000043
Figure BDA0002195723640000044
Wherein A is*Is the optimal solution of A, k*Is the optimal solution of k, | W | | non-woven phosphorl1Representing the L1 norm.
Preferably, in the process of training the binary convolution neural network, the weight is binarized only in forward propagation and backward propagation, binarization is used when the forward propagation is carried out for the first time, binarization is not carried out during the backward propagation and parameter updating, and a full-precision value is used; the inverse gradient formula is as follows:
Figure BDA0002195723640000045
in the calculation process of the gradient descent process, a random gradient descent method is adopted, and the operation is as follows:
Figure BDA0002195723640000046
and when the fastest descending direction is calculated, randomly selecting one data to calculate.
The invention has the advantages that: the invention aims to provide a railway shunting signal lamp detection method based on a binarization convolutional neural network, which solves the problems of manual identification of shunting signal lamps in a locomotive crew department and the deployment of the convolutional neural network on an embedded device, and aims at carrying out binarization on the weight of the neural network to accelerate the operation of the neural network and reduce the memory consumption of the weight. The method has the advantages of high accuracy and reliability, and can accurately detect the signal lamp in front of the locomotive.
The invention creatively realizes the compression and acceleration of the network effectively by means of establishing the convolutional neural network as the binary convolutional neural network and quantizing all parameters. The invention creatively combines the binary convolution neural network with the embedded type, can effectively solve the problem of 'two-head one-super' of the train, and has great development space; the invention innovatively compresses the convolutional neural network, thereby greatly promoting the combination of deep learning and embedding.
Drawings
FIG. 1 is a flow chart of a method in an embodiment of the present invention;
FIG. 2 is a diagram of a full-precision convolutional neural network in accordance with an embodiment of the present invention;
FIG. 3 is a diagram of a binarization training structure in the embodiment of the present invention;
fig. 4 is a diagram illustrating a functional effect of practical application of the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention, and it is obvious that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, fig. 2, fig. 3 and fig. 4, a method for detecting a shunting signal lamp of a railway based on a binary convolutional neural network comprises the following steps:
step 1, in a sample preparation stage, a camera is placed at the locomotive head to obtain a video, the collected video is cut at 20fps, a picture is cut into pictures with the resolution of 1280 multiplied by 720, then the size of the pictures is adjusted to 416 multiplied by 3, the pictures containing targets (signal lamps) are manually screened, and finally 40 ten thousand pictures containing the signal lamps are obtained. And (3) manufacturing a training set and a testing set: and (3) carrying out artificial target detection and marking on the pictures, wherein signal lamps are divided into 2 types of blue lamps and 2 types of white lamps, and are respectively marked as '1', '0', and the ratio of 19: a scale of 1 randomly divides the data set into a training set and a test set. The data comprises picture data under different backgrounds, and the robustness of the network is improved.
Step 2, a network construction stage, in which a full-precision convolutional neural network is created, the full-precision convolutional neural network is a forward-based deep neural network, in this embodiment, 14 binarization convolutional layers are constructed, 5 downsampling layers and 2 full-connection layers form a feedforward convolutional neural network, two convolutional cores of (1 × 1) and (3 × 3) are constructed in the convolutional layers to alternately perform feature extraction on input, and filters with fixed size and step length are used to perform maximum extraction in sequence. The invention determines through testing to evaluate using softmax for candidate regions. The characteristic graphs obtained by the convolutional layers can be fully utilized, namely the high convolutional layers have larger receptive field, large objects can be detected conveniently, the low convolutional layers have high resolution, and small objects can be detected conveniently.
The invention adopts anchor boxes (candidate boxes) obtained by clustering K-means from a training set, when K is 9, an anchor box has the best effect, the anchor box has 9 anchor boxes with different sizes, each anchor box has an evaluation score for the detection of the most target object, wherein the number of the anchor boxes is an extremely important hyper-parameter, and finally the size of the anchor box is calculated by using a K-means algorithm, and the sizes of the anchor boxes are respectively as follows: 12 × 18, 170 × 194, 8 × 10, 74 × 86, 37 × 40, 63 × 77, 24 × 30, 113 × 134, 17 × 24.
Finally, the invention uses a maximum suppression algorithm to screen the preliminarily screened prediction frames, and the non-maximum suppression algorithm respectively screens the blue lamps and the white lamps, for example, for the blue lamps:
a is b × c type (1)
Where b is the probability of the target being contained in the prediction box and c is the probability of the target being a blue light. The non-maximum suppression algorithm is to sort the a of each prediction box in the increasing direction, select the largest a to perform IOU calculation with other a, delete the preselection box with the threshold value larger than 60%, then operate the preselection box with the second largest a in the same way, finally the remaining preselection box is the result obtained by the embodiment, and finish the detection of the traffic lights.
In the network training stage, two methods are used for training the binaryzation of the binaryzation convolutional neural network through training set samples: in the determining method and the probability statistical method, considering that Stochastic (probability statistical method) needs hardware to generate a random number, the present embodiment performs binarization operation on convolutional layer parameters by using a Derterministic (determining method):
Figure BDA0002195723640000061
wherein W is a real weight, WbThe weight after binarization has only two values of '+ 1', '1'. Although this is a deterministic operation, averaging many of the input weights of the hidden unit can compensate for the lost information. Wherein the first and last layers of the binary convolutional neural network maintain full-precision weights.
In order to limit the convolutional neural network to have the binarization weight, this embodiment uses a binary convolution kernel and a scale parameter to approximately replace the original convolution kernel, and performs binarization operation according to equation (2):
Figure BDA0002195723640000062
where A x W represents the operation of a convolutional layer, A represents the input, the dimension is C x win x hin, W represents the convolution kernel (or called weight), the dimension is C x W x h, C is a binary convolution kernel, k is a scale parameter,
Figure BDA0002195723640000063
is calculated for convolution without multiplication.
To satisfy the above equation, a and k should have optimal values:
A*＝argmax{WTA}s.tA∈{+1，-1}nformula (4)
Namely: a. the*Sign (w) formula (5)
Figure BDA0002195723640000071
Wherein A is*Is the optimal solution of A, k*Is the optimal solution of k, | W | | non-woven phosphorl1Representing the L1 norm.
In the gradient descent process calculation and the CNN training process, the present embodiment binarizes the weight only in forward propagation and backward propagation. The forward propagation uses binarization when the first time, and the backward propagation and parameter updating periods do not use binarization, and the full precision value is used. The inverse gradient formula is as follows:
Figure BDA0002195723640000072
the embodiment adopts a random gradient descent method, and the adoption of the gradient descent method enables the embodiment to train the network better, and the operation is as follows:
Figure BDA0002195723640000073
and when the fastest descending direction is calculated, randomly selecting one data to calculate.
And 4, in the network operation stage, inputting the data set of the shunting signal lamp of the railway as input into a trained binary convolution neural network, simulating the characteristic response of the CNN by the binary convolution neural network, and obtaining the detection of the target signal lamp through network operation.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims (10)

1. The railway shunting signal lamp detection method based on the binary convolution neural network is characterized by comprising the following steps of: the method comprises the following steps:
s1, preparing a data set, collecting a video in front of a train head acquired when a train runs, cutting the video into pictures with set sizes, and manually screening the pictures containing a target to obtain a target picture; then dividing the target picture into a training set and a testing set according to the set proportion of the blue light and the white light;
s2, network construction, namely constructing a binary convolution neural network framework by utilizing a training set;
s3, network training, namely training a binarization convolutional neural network by using a binarization method through a test set to obtain a target network;
and S4, running the network, and carrying out real-time target detection on the shunting signal lamp of the railway by using the target network.
2. The railway shunting signal lamp detection method based on the binarization convolutional neural network as claimed in claim 1, characterized in that: the step S1 specifically includes:
s11, data collection, namely placing a camera on the locomotive head to obtain a video, cutting the collected video at a set frame rate, manually screening pictures containing targets (signal lamps), and finally obtaining pictures containing the signal lamps at a set amount;
s12, manufacturing a training set and a testing set, carrying out artificial target detection and labeling on the photos, and dividing the photos into the training set and the testing set according to the ratio of blue light to white light of 19: 1.
3. The railway shunting signal lamp detection method based on the binarization convolutional neural network as claimed in claim 1, characterized in that: the binary convolution neural network framework in the step S2 specifically includes: firstly, generating a feature map based on a binary convolutional neural network of the convolutional neural network, extracting a regional prediction frame by using 9 candidate frames, and finally screening the most accurate target prediction frame containing a signal lamp and an identification result by using non-maximum suppression, wherein the front end of the model uses a residual error network to extract features, the rear end of the model generates feature maps with different sizes through convolution, and the convolutional layer is divided into two scales according to the size.
4. The railway shunting signal lamp detection method based on the binarization convolutional neural network as claimed in claim 1, characterized in that: the step S3 specifically includes: and (3) performing binarization operation on the convolution layer parameters by adopting a determination method:
Figure FDA0002195723630000011
wherein W is a real weight, WbThe weight after binarization has only two values of '+ 1', '1';
in order to limit the convolutional neural network to have the binarization weight, this embodiment uses a binary convolution kernel and a scale parameter to approximately replace the original convolution kernel, and performs binarization operation according to equation (2):
Figure FDA0002195723630000021
wherein A x W represents the operation of a convolution layer, A represents the input, the dimension is C x win x hin, W represents the convolution kernel, the dimension is C x W x h, C is a binary convolution kernel, k is a scale parameter,
Figure FDA0002195723630000022
is calculated for convolution without multiplication.
To satisfy equation (2), a and k should be optimized:
A*＝argmax{WTA}s.t A∈{+1,-1}nformula (3)
Namely: a. the*Sign (w) formula (4)
Figure FDA0002195723630000023
Figure FDA0002195723630000024
Wherein A is*Is the optimal solution of A, k*Is the optimal solution of k, | W | | non-woven phosphorl1Representing the L1 norm.
5. The railway shunting signal lamp detection method based on the binarization convolutional neural network as claimed in claim 4, characterized in that: the step S3 specifically includes: in the process of training the binary convolution neural network, the weight is binarized only in forward propagation and backward propagation, binarization is used when the forward propagation is carried out for the first time, binarization is not carried out during the backward propagation and parameter updating, and a full-precision value is used; the inverse gradient formula is as follows:
Figure FDA0002195723630000025
6. the railway shunting signal lamp detection method based on the binarization convolutional neural network as claimed in claim 5, characterized in that: in the calculation process of the gradient descent process, a random gradient descent method is adopted, and the operation is as follows:
Figure FDA0002195723630000026
and when the fastest descending direction is calculated, randomly selecting one data to calculate.
7. Railway shunting signal lamp detecting system based on binarization convolution neural network, its characterized in that: comprises that
The data set preparation module is used for collecting a video in front of a locomotive acquired when a train runs, cutting the video into pictures with set sizes, and manually screening the pictures containing the target to obtain a target picture; then dividing the target picture into a training set and a testing set according to the set proportion of the blue light and the white light;
the network construction module is used for constructing a binary convolutional neural network framework by utilizing the training set;
the network training module is used for training a binary convolution neural network by using a binary method through a test set to obtain a target network;
and the network operation module is used for carrying out real-time target detection on the railway shunting signal lamp by utilizing a target network.
8. The railway shunting signal lamp detection system based on the binarization convolutional neural network as claimed in claim 7, wherein: the dataset preparation module comprises:
the data collection unit is used for placing a camera on the locomotive head to obtain a video, cutting the collected video at a set frame rate, manually screening pictures containing targets (signal lamps), and finally obtaining pictures containing the signal lamps at a set amount;
and the production units of the training set and the test set are used for carrying out artificial target detection and marking on the photos, and dividing the photos into the training set and the test set according to the ratio of blue light to white light of 19: 1.
9. The railway shunting signal lamp detection system based on the binarization convolutional neural network as claimed in claim 7, wherein: the specific training method of the network training module comprises the following steps: and (3) performing binarization operation on the convolution layer parameters by adopting a determination method:
Figure FDA0002195723630000031
wherein W is a real weight, WbThe weight after binarization has only two values of '+ 1', '1';
in order to limit the convolutional neural network to have the binarization weight, this embodiment uses a binary convolution kernel and a scale parameter to approximately replace the original convolution kernel, and performs binarization operation according to equation (2):
wherein A x W represents the operation of a convolution layer, A represents the input, the dimension is C x win x hin, W represents the convolution kernel, the dimension is C x W x h, C is a binary convolution kernel, k is a scale parameter,
Figure FDA0002195723630000033
is calculated for convolution without multiplication.
To satisfy equation (2), a and k should be optimized:
A*＝argmax{WTA}s.t A∈{+1,-1}nformula (3)
Namely: a. the*Sign (w) formula (4)
Figure FDA0002195723630000034
Wherein A is*Is the optimal solution of A, k*Is the optimal solution of k, | W | | non-woven phosphorl1Representing the L1 norm.
10. The railway shunting signal lamp detection system based on the binarization convolutional neural network as claimed in claim 9, wherein: in the process of training the binary convolution neural network, the weight is binarized only in forward propagation and backward propagation, binarization is used when the forward propagation is carried out for the first time, binarization is not carried out during the backward propagation and parameter updating, and a full-precision value is used; the inverse gradient formula is as follows:
in the calculation process of the gradient descent process, a random gradient descent method is adopted, and the operation is as follows:
Figure FDA0002195723630000042
and when the fastest descending direction is calculated, randomly selecting one data to calculate.
CN201910857618.2A 2019-09-09 2019-09-09 Railway shunting signal lamp detection method and system based on binary convolution neural network Pending CN110633668A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910857618.2A CN110633668A (en) 2019-09-09 2019-09-09 Railway shunting signal lamp detection method and system based on binary convolution neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910857618.2A CN110633668A (en) 2019-09-09 2019-09-09 Railway shunting signal lamp detection method and system based on binary convolution neural network

Publications (1)

Publication Number Publication Date
CN110633668A true CN110633668A (en) 2019-12-31

Family

ID=68972543

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910857618.2A Pending CN110633668A (en) 2019-09-09 2019-09-09 Railway shunting signal lamp detection method and system based on binary convolution neural network

Country Status (1)

Country Link
CN (1) CN110633668A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111711585A (en) * 2020-06-11 2020-09-25 西安交通大学 Real-time signal sequence detection method based on deep learning
CN111723690A (en) * 2020-06-03 2020-09-29 北京全路通信信号研究设计院集团有限公司 Circuit equipment state monitoring method and system
CN113098805A (en) * 2021-04-01 2021-07-09 清华大学 Efficient MIMO channel feedback method and device based on binarization neural network
WO2022213307A1 (en) * 2021-04-07 2022-10-13 Nokia Shanghai Bell Co., Ltd. Adaptive convolutional neural network for object detection

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108304786A (en) * 2018-01-17 2018-07-20 东南大学 A kind of pedestrian detection method based on binaryzation convolutional neural networks
CN108304844A (en) * 2018-01-30 2018-07-20 四川大学 Agricultural pest recognition methods based on deep learning binaryzation convolutional neural networks
CN108765506A (en) * 2018-05-21 2018-11-06 上海交通大学 Compression method based on successively network binaryzation
CN110070119A (en) * 2019-04-11 2019-07-30 北京工业大学 A kind of handwritten numeral image recognition classification method based on binaryzation deep neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108304786A (en) * 2018-01-17 2018-07-20 东南大学 A kind of pedestrian detection method based on binaryzation convolutional neural networks
CN108304844A (en) * 2018-01-30 2018-07-20 四川大学 Agricultural pest recognition methods based on deep learning binaryzation convolutional neural networks
CN108765506A (en) * 2018-05-21 2018-11-06 上海交通大学 Compression method based on successively network binaryzation
CN110070119A (en) * 2019-04-11 2019-07-30 北京工业大学 A kind of handwritten numeral image recognition classification method based on binaryzation deep neural network

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111723690A (en) * 2020-06-03 2020-09-29 北京全路通信信号研究设计院集团有限公司 Circuit equipment state monitoring method and system
CN111723690B (en) * 2020-06-03 2023-10-20 北京全路通信信号研究设计院集团有限公司 Method and system for monitoring state of circuit equipment
CN111711585A (en) * 2020-06-11 2020-09-25 西安交通大学 Real-time signal sequence detection method based on deep learning
CN113098805A (en) * 2021-04-01 2021-07-09 清华大学 Efficient MIMO channel feedback method and device based on binarization neural network
WO2022213307A1 (en) * 2021-04-07 2022-10-13 Nokia Shanghai Bell Co., Ltd. Adaptive convolutional neural network for object detection

Similar Documents

Publication Publication Date Title
CN110633668A (en) Railway shunting signal lamp detection method and system based on binary convolution neural network
CN106778595B (en) Method for detecting abnormal behaviors in crowd based on Gaussian mixture model
US20210319561A1 (en) Image segmentation method and system for pavement disease based on deep learning
CN100545867C (en) Aerial shooting traffic video frequency vehicle rapid checking method
CN112132119B (en) Passenger flow statistical method and device, electronic equipment and storage medium
CN108446630A (en) Airfield runway intelligent control method, application server and computer storage media
CN110175613A (en) Street view image semantic segmentation method based on Analysis On Multi-scale Features and codec models
CN109871799B (en) Method for detecting mobile phone playing behavior of driver based on deep learning
CN113065474B (en) Behavior recognition method and device and computer equipment
CN110008854B (en) Unmanned aerial vehicle image highway geological disaster identification method based on pre-training DCNN
CN111898581A (en) Animal detection method, device, electronic equipment and readable storage medium
CN108615226A (en) A kind of image defogging method fighting network based on production
CN105574550A (en) Vehicle identification method and device
CN114283469B (en) Improved YOLOv4-tiny target detection method and system
CN108197544B (en) Face analysis method, face filtering method, face analysis device, face filtering device, embedded equipment, medium and integrated circuit
CN110222604A (en) Target identification method and device based on shared convolutional neural networks
CN114022770A (en) Mountain crack detection method based on improved self-attention mechanism and transfer learning
CN112052744A (en) Environment detection model training method, environment detection method and device
CN112528934A (en) Improved YOLOv3 traffic sign detection method based on multi-scale feature layer
CN102360434B (en) Target classification method of vehicle and pedestrian in intelligent traffic monitoring
CN109934170A (en) A kind of mine resources statistical method based on computer vision
CN108648210A (en) It is a kind of static state complex scene under fast multi-target detection method and device
CN112907138B (en) Power grid scene early warning classification method and system from local to whole perception
CN113159146A (en) Sample generation method, target detection model training method, target detection method and device
CN112488043A (en) Unmanned aerial vehicle target detection method based on edge intelligence

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
TA01 Transfer of patent application right

Effective date of registration: 20211213

Address after: 230012 Hedong Locomotive Section 005, Northeast Bengbu Road, Xinzhan District, Hefei City, Anhui Province

Applicant after: CHINA RAILWAY SHANGHAI GROUP Co.,Ltd. HEFEI CENTRAL OFFICE

Applicant after: HEFEI FEIYANG ELECTRICAL Co.,Ltd.

Address before: 230001 No. 20, Xiangyi Road, Xiaomiao Town Industrial Zone, Shushan District, Hefei City, Anhui Province

Applicant before: HEFEI FEIYANG ELECTRICAL Co.,Ltd.

TA01 Transfer of patent application right
CB03 Change of inventor or designer information

Inventor after: Li Yingzhi

Inventor after: Hong Yuchen

Inventor after: Wei Xing

Inventor after: Fan Xiangye

Inventor after: Zhu Weijun

Inventor after: Fang Xiangming

Inventor after: Cai Yongbin

Inventor after: Zhu Yuhu

Inventor after: Li Baiqi

Inventor after: Sheng Dianmo

Inventor before: Cai Yongbin

Inventor before: Zhu Yuhu

Inventor before: Wei Xing

Inventor before: Li Baiqi

Inventor before: Sheng Dianmo

Inventor before: Hong Yuchen

CB03 Change of inventor or designer information