CN106709511A - Urban rail transit panoramic monitoring video fault detection method based on depth learning - Google Patents
Urban rail transit panoramic monitoring video fault detection method based on depth learning Download PDFInfo
- Publication number
- CN106709511A CN106709511A CN201611125030.0A CN201611125030A CN106709511A CN 106709511 A CN106709511 A CN 106709511A CN 201611125030 A CN201611125030 A CN 201611125030A CN 106709511 A CN106709511 A CN 106709511A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- video
- layer
- fault
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an urban rail transit panoramic monitoring video fault detection method based on depth learning. The method comprises a data set construction process, a model training generation process and an image classification recognition process. The data set construction process processes a definition abnormity video, a colour cast abnormity video and a normal video in an urban rail transit panoramic monitoring video. A training set and a test set are classified. The model training generation process comprises model training and model test. The model training is to train a fault video image recognition model based on a convolution neural network. The convolutional neural network comprises a plurality of convolution layers and a plurality of full connection layers. The model test is to calculate the test accuracy. If expectation is not fulfilled, the fault video image recognition model is optimized. The image classification recognition process comprises the steps that a single-frame image to be recognized is input into the model, and the fault video image recognition model outputs an image classification result to complete the fault image detection of the urban rail transit panoramic monitoring video.
Description
Technical field
The invention belongs to deep learning field, and in particular to the urban track traffic overall view monitoring video based on deep learning
Fault detection method.
Background technology
In recent years with the growth of private savings automobile quantity, urban traffic blocking problem is asked greatly as what city life faced
Topic, and the generation of traffic accident is also more and more, and these problems increasingly have influence on the normal life of the ordinary municipal people.Cause
The normal operation of this urban track traffic overall view monitoring video is particularly important, and this requires in real time to hand over city rail
The running status of logical overall view monitoring video is checked, corresponding failure is recognized in time and is overhauled in time.Wherein panorama refers to
Be that single monitoring in multiple monitor videos and the spliced image of area three-dimensional Model Fusion or overall view monitoring system is regarded
Frequently, and urban track traffic overall view monitoring video be refer to overall view monitoring system in single monitor video.Urban track traffic is complete
The picture of scape monitoring belongs to digital picture, and the characteristic that the pixel with digital picture is represented can be by Computer Storage and treatment
The characteristics of.And the detection of urban track traffic overall view monitoring video failure can be converted into the detection of video quality with examine
It is disconnected.Detection and diagnosis to video quality are the evaluation to digital picture quality, and the evaluation of digital picture quality that is to say
To distortion and the identification of degraded image and diagnosis.Image quality decrease be usually expressed as image definition abnormal (image blurring),
Situations such as there is striped in colour cast, snowflake (noise), picture.Traditional fault picture method for objectively evaluating main thought is all to rely on
Artificial experience come the feature extracting and analyze image, by contrasting the difference between failure picture and normal picture, particularly picture
The difference of vegetarian refreshments determines whether to be failure picture.These algorithms during application, the selection of picture feature and definition
Quality determines the good and bad degree of testing result, with significant limitation.
The content of the invention
In order to overcome prior art largely manually to extract the wave in the time and efforts caused by features in actual applications
Take, the present invention proposes a kind of urban track traffic overall view monitoring video fault detection method based on deep learning.The method bag
Include data set structure, model training and generation, picture classification and recognize three parts.Existing video is converted into two field picture first
Image data set is built, acquisition convolutional neural networks model, the network for then being obtained using training are trained to view data
Model is classified to image, exports the picture classification of test and the accuracy rate of identification.
Technical solution of the present invention provides a kind of urban track traffic overall view monitoring video fault detect based on deep learning
Method, including data set building process and model training generating process, picture classification identification process,
The data set building process, including to the definition anomalous video in urban track traffic overall view monitoring video,
Colour cast anomalous video and normal video are processed, and video are converted into single-frame images, and dimension of picture is normalized
Treatment, builds definition abnormal image data set, colour cast abnormal image data set and normal image data collection, arbitrarily extracts
Image is divided into training set and test set, the image data set needed for obtaining；
The model training generating process, including model training and model measurement,
The model training, including the fault video image recognition model based on convolutional neural networks is trained, institute
Stating convolutional neural networks includes multiple convolutional layers and multiple full articulamentums；Training method is as follows,
It is multiple using the picture in training set by being input to convolutional neural networks as bottom data after Data Format Transform
Convolutional layer is cascaded, and the view data to being input into carries out convolution algorithm, abstract image feature, the multiple characteristic patterns of generation；
Used as input, incoming full articulamentum turns feature by each full articulamentum completely for the output of last convolutional layer
Chemical conversion one-dimensional vector output；Determine the error between real output value and desired value, backpropagation is carried out to network, adjust network
Parameter；
Then whether training of judgement error restrains, if otherwise returning to training initiating terminal input picture carries out feature learning, if
It is whether judgment models training iterations reaches predetermined iterations, when predetermined iterations is reached, model instruction
White silk terminates, and otherwise returns to training initiating terminal input picture and continues to train；
Training obtains the fault video image recognition model based on convolutional neural networks after terminating；
The model measurement, including the picture input model in test set is trained into gained fault video image recognition model
Network bottom layer, fault video image recognition model exports the classification results of picture, includes that quality is normal and abnormal quality two
Kind of classification, finally compares with the respective labels that have set in test set, the accuracy rate of test is calculated, if be not reaching to pre-
Phase, then optimize fault video image recognition model, until reaching expection, obtain final fault video image recognition model；
The picture classification identification process, including single frames picture input model training generating process to be identified is obtained
The network bottom layer of fault video image recognition model, fault video image recognition model exports the classification results of picture, completes real
Existing urban track traffic overall view monitoring video failure picture detection.
And, the optimization fault video image recognition model, including the extension for carrying out image data set.
And, the optimization fault video image recognition model, including increase the number of convolutional layer, reduce full articulamentum
Number.
And, original fault video image recognition model, the convolutional neural networks are complete including 3 convolutional layers and 3
Articulamentum.
And, original fault video image recognition model include input, convolutional layer C1, down-sampled layer S1, convolutional layer C2,
Down-sampled layer S2, convolutional layer C3, down-sampled layer S3, full articulamentum fc4, full articulamentum fc5, full articulamentum fc6 and output layer.
And, during optimization fault video image recognition model, it is described by convolutional neural networks be changed to include 4 convolutional layers with
2 full articulamentums.
And, during optimization fault video image recognition model, by convolutional neural networks be changed to include input, convolutional layer C1,
Down-sampled layer S1, convolutional layer C2, down-sampled layer S2, convolutional layer C3, down-sampled layer S3, convolutional layer C4, down-sampled layer S4, full connection
Layer fc5, full articulamentum fc6 and output layer.
And, the optimization fault video image recognition model, including the adjustment for carrying out prototype network parameter.
And, whether the training of judgement error restrains, and is realized based on error loss function.
The advantage of the invention is that：Compared with original method, the urban track traffic overall view monitoring based on deep learning is regarded
Frequency fault detection method, when image classification is recognized, is the automatic learning characteristic from big data, rather than using the spy of hand-designed
Levy model.The feature of hand-designed relies primarily on the Heuristics of designer, is difficult with the advantage of big data, the design of feature
In only allow a small amount of parameter occur, and deep learning can include thousands of parameter, for training the data of deep learning
More, the robustness of deep learning algorithm, generalization ability are stronger.On the other hand, deep learning is not explicit feature extraction,
But implicitly from training sample learning characteristic, therefore, except the undesired change that need not pay close attention to every layer of weights and character representation.Power
The shared characteristic of value effectively reduces the order of magnitude of weights, also reduces the difficulty of training.Sample can be directly as deep learning
The input of model, feature selecting and extraction without carrying out complexity to sample image using algorithm in addition.Therefore, the present invention is being protected
High degree of automation while card accuracy rate, with important market value.
Brief description of the drawings
Fig. 1 is the overall process flow figure of the embodiment of the present invention
Fig. 2 is the model training process flow diagram flow chart of the embodiment of the present invention
Fig. 3 is the model measurement process flow diagram flow chart of the embodiment of the present invention
Fig. 4 is the convolutional neural networks network model of the fault video image recognition of the embodiment of the present invention
Fig. 5 is the convolutional neural networks structure chart of the fault video image recognition of the embodiment of the present invention
Fig. 6 is the convolutional neural networks Network Optimization Model of the fault video image recognition of the embodiment of the present invention
Fig. 7 is the convolutional neural networks network optimization structure chart of the fault video image recognition of the embodiment of the present invention
Fig. 8 is the convolutional neural networks model and basic convolutional neural networks Model Identification after the optimization of the embodiment of the present invention
Accuracy rate compares figure
Fig. 9 is the activation primitive figure of the convolutional neural networks of the fault video image recognition of the embodiment of the present invention
Specific implementation method
Specific implementation method of the invention is described in detail below in conjunction with accompanying drawing and case study on implementation.
Method provided by the present invention is to be detected by the quality of the single-frame images to constituting video and realized.Because
The quality testing of digital picture and diagnosis are built upon to being carried out on the basis of feature extraction, and feature extraction to image with
Classification can sum up in the point that the extraction and analysis of the pixel to pie graph picture again.Therefore knowing for image classification identification correlation is studied in depth
It is sensible when important.Deep learning (deep learning, i.e. DL) algorithm is to be analyzed study for setting up and simulating human brain
Neuromechanism, builds the network structure model of many hidden layers, is trained by mass data, obtains representative feature
Information, realizes the explanation and prediction to data such as image, sound, texts, improves the accuracy of classification and prediction.And know in image
Other field, based on deep learning model particularly convolutional neural networks (convolution neural networks, abbreviation
CNN) model, using than wide.Therefore the present invention is pioneering proposes to answer on the convolutional neural networks model in deep learning model
For urban track traffic overall view monitoring video fault picture identification field detection, the accurate of identification is not only greatly improved
Property, while the waste in it also avoid traditional approach in a large amount of artificial time and efforts extracted caused by feature, reduces fortune
Calculate complexity.
The operating system that embodiment is used is Ubuntu14.04, and CPU is Intel Core i3-4130,3.40GHZ,
Internal memory 4G, GPU are NVIDIA Quadro K600.Embodiment basic procedure is as shown in Figure 1：
1. data set builds
The training and generation of deep learning neural network model are built upon the basis learnt to a large amount of picture features
Upper realization.Present invention research is the urban track traffic monitor video fault detection method based on deep learning, the method
The basis that the model for using is set up is a large amount of fault video images and normal picture.Fault picture be definition abnormal image and
The class of colour cast abnormal image two, picture rich in detail that normal picture is received by human eye, recognizable.To definition anomalous video, color
Inclined anomalous video and normal video are processed, and video are converted into single-frame images, and place is normalized to dimension of picture
Reason, builds definition abnormal image data set, colour cast abnormal image data set and normal image data collection.Arbitrarily extract figure
As being divided into training set, the class of test set two, during model training, training set is used to carry out study and the network ginseng of feature
Several adjustment.Picture in test set is different from training set, is retouched by testing the data not occurred in training set
State the classification performance of network model.The network structure and parameter of model can be adjusted according to test result, the circulation of training is changed
Number of times, the optimization of implementation model.During specific implementation, picture number can be combined and be actually needed selection in each set.
2. model training and generation
The training of model is main with generation to be included training and two processes of test.
(1) model training
After image data collection builds, it is input into as the bottom data of network by data conversion, including training picture number
It is leveldb forms according to integrating by Data Format Transform, realizes leveldb database sharings.Convolutional neural networks enter to data
After the operation such as row convolution, pond, realize that network model is exported.The good and bad precision value (accuracy) of observation descriptive model and
The results such as error loss function end value (loss), adjust network model parameter, until accuracy and the loss convergence for exporting
And tend to a state for stabilization, model training terminates, and the characteristics of image for extracting is stored in the model.The model is used for
Ensuing urban track traffic overall view monitoring video fault detect.Model training flow such as Fig. 2.
First, it is that leveldb forms are input into as bottom data by Data Format Transform based on training image data collection
To network, (convolution 1, convolution 2 ... convolution n) in such as Fig. 2, the view data to being input into carry out convolution fortune to the cascade of multiple convolutional layers
Calculate, abstract image feature, the multiple characteristic patterns of generation.Using multilayer convolution because the feature that one layer of convolution learns is often office
Portion, the convolution number of plies is higher, feature more globalization acquired.When training network model, the output of previous convolutional layer
As the input of this convolutional layer, the output of this convolutional layer again using as the input of next convolutional layer, until finally
One convolutional layer.The output of last convolutional layer can be used as input, incoming full articulamentum (full articulamentum 1, Quan Lian in such as Fig. 2
The full articulamentum n) that connects layer 2 ....Feature is fully converted into one-dimensional vector output by full articulamentum.Determine real output value and desired value
Between error, carry out backpropagation to network, adjust network parameter.Then judge whether training error restrains, in the present invention
Network model in, error loss function (softmax_loss) characterising parameter that selection is commonly used in convolutional neural networks now is adjusted
Whole good and bad degree.When the value of error loss function is smaller and reaches convergence (value of such as error loss function no longer changes,
When being below 1, those skilled in the art can voluntarily preset Rule of judgment to value), you can judge that network parameter now reaches
To optimal, now can determine whether whether model training iterations reaches predetermined iterations, when reaching predetermined iterations
When, model training terminates, and otherwise returns to training initiating terminal input picture and continues to train.Final training terminates to obtain convolutional Neural net
Network model, what is stored in model is the species and characteristic information of the image for extracting.If the value of error loss function is larger,
Needs proceed training, and returning to training initiating terminal input picture carries out feature learning.
(2) model measurement
, it is necessary to the application performance of test model after model training generation.The test process of model is as shown in Figure 3.First, from
Test pictures are concentrated and extract picture, call the disaggregated model for having trained, by the network bottom layer of the image data input model,
Then feature extraction and output (in such as Fig. 3 convolution 1, convolution 2 ... convolution of the view data in the network model by convolutional layer
N), incoming full articulamentum.Feature changed into one-dimensional vector output by n full articulamentum (full connection output 1 in such as Fig. 3, complete
Connection output 2 ... full connection output n), contrast test is carried out with existing feature in disaggregated model, exports the classification knot of picture
Really, namely image classification, have quality normal and two kinds of classifications of abnormal quality, complete to be input into the classification of picture.Finally this is marked
Sign and compare with the label set in test set, calculate the accuracy rate A of test, i.e., have the test of a in m test data
Data classification is correct, then accuracy rate A=a (the correct test data quantity of classification)/m (test data sum).If precision is relatively low,
Illustrate that the network model that training is obtained is not reaching to be expected, i.e. the parameter of network is not optimal solution, is verified using model
During, the classification results to picture are also undesirable.Therefore need continuation to optimize network model and improve its test accuracy rate.This
The optimization that invention passes through three below method implementation model.
Method 1, the extension of image data set
Image data set is the basis that convolutional neural networks model carries out feature learning, and training of its size to model has weight
Want meaning.Image data set is too small, and the feature extracted is not abundant enough, and model will be subject to for the abstracting power of feature learning
Limitation, and then the test accuracy rate of model can be influenceed.Therefore, in order to improve the accuracy rate of model, it is necessary in raw data set
On the basis of, image data set is extended.
There are significant structural and continuity Characteristics due to image and video sequence sheet on time and Spatial Dimension,
And comprising substantial amounts of redundancy, the image processing operations such as translation, rotation, distortion, gaussian sum salt-pepper noise to image, all
Effective training data can be produced, as the measure of extended model training set, the robustness without reducing model.
Method 2, the variation of prototype network structure
The network hierarchical structure of model can influence the test accuracy rate of model, so original convolution neural network model it
On, take certain measure to optimize its network structure.It is appropriate to increase the individual of convolutional layer on the premise of image data set is expanded
Number, reduces the number of full articulamentum, improves the test accuracy rate of model, reduces the value of error loss function, reaches convergence
State.
Method 3, the adjustment of prototype network parameter
The test accuracy rate of model is all pind down due to convolution nuclear volume and size, the number of iterations.For network
Model parameter, present invention further propose that on existing basic model, adjust the number of convolution kernel and the size of convolution kernel with
And iterations, the test accuracy rate of model is improved, reduce the size of error loss function.
The optimization of implementation model can be combined using a method above or multiple methods.
3. picture classification identification
In image classification process, the network model obtained in training process is tested using authentication image.For treating
The picture of identification, the model for calling training to generate carries out Classification and Identification to picture.Picture classification identification specific steps as and survey
Examination process is similar to, and based on identification model and single frames picture to be identified, realizes urban track traffic overall view monitoring video failure picture
Detection.
Embodiment one：
1. the implementation of image data set
Final purpose of the invention is to realize the Fault Identification to urban track traffic overall view monitoring video, is substantially to realize
The identification of fault video image, image definition and colour cast are two particularly important factor of judgment of video quality detection.This
, used as training and the sample set of test, image is for invention selection definition exception picture, colour cast exception picture and normal picture
RGB color pattern.In the process, definition anomalous video, colour cast anomalous video and normal video are processed, will
Monitor video is converted into single-frame images (i.e. frame picture), and dimension of picture is normalized, and builds definition Abnormal Map
As data set, colour cast abnormal image data set and normal image data collection.According to allocation proportion, therefrom abstract image is divided into instruction
Practice collection, test set.Wherein the abnormal picture of definition has 500, and the abnormal picture of colour cast has 500, and normal picture has 500
.100 composition test pictures are therefrom respectively randomly selected, remaining 1200 are then used in model training.Picture size be 256 ×
256, form is jpeg image.After image data collection builds, image data collection is transformed into leveldb database files,
It is input into as the bottom data of network.
2. model training and the implementation of generation
The fault video image recognition model based on convolutional neural networks of the design of the embodiment of the present invention one has multitiered network
Structure, including 3 convolutional layers, 3 down-sampled layers and 3 full articulamentums.It is input to the view data of network bottom layer
Size be 60 × 60, network model as shown in figure 4, wherein have successively data (training data), C1 (CONVOLUTION LAYER,
Convolutional layer), conv1 (convolutional layer output), ReLU activation primitives, S1 (POOLING LAYER, pond layer), pool1 (pond layers
Output), C2 (CONVOLUTION LAYER, convolutional layer), conv2 (convolutional layer output), ReLU activation primitives, S2 (POOLING
LAYER, pond layer), pool2 (output of pond layer), C3 (CONVOLUTION LAYER, convolutional layer), (convolutional layer is defeated for conv3
Go out), ReLU activation primitives, S3 (POOLING LAYER, pond layer), pool3 (output of pond layer), fc4 (INNER_
PRODUCT, full articulamentum), f4 (full articulamentum output), ReLU activation primitives, fc5 (INNER_PRODUCT, full articulamentum),
F5 (full articulamentum output), ReLU activation primitives, fc6 (INNER_PRODUCT, full articulamentum), f6 (full articulamentum output),
Accuracy (training precision) and Loss (training loss).
Connect a pond layer, i.e. view data (data) behind each convolutional layer to be input into by network bottom layer, first
Process of convolution is carried out by convolutional layer (C1), the data after treatment select ReLU functions as activation primitive, then by after treatment
Data carry out the mode that down-sampled, down-sampled mode is maximum pond as incoming pond layer (S1) is input into, and obtain feature
The maximum in region.By after continuous three convolutional layers and the treatment of down-sampled layer, data change into one-dimensional vector and are input to and connect entirely
Layer (f4) is connect, the data after full articulamentum treatment equally select ReLU functions as activation primitive, by continuous two full connections
After layer treatment, the test accuracy rate and error loss function value of final output model.The fault video figure for designing on this basis
As the convolutional neural networks network frame of identification is as shown in Figure 5.
■ is input into：Network mould will be input in data set building process by the image data set of size normalized
In type, 60 × 60 image is chosen in the picture centre of 256 × 256 sizes.Set at the picture batch of network an iteration treatment
Reason number is 60.
■ convolutional layers C1：Convolution is carried out to view data from 16 convolution kernels of 4 × 4 sizes, 16 characteristic patterns are obtained
(Feature map).The degree that the size and number meeting effect characteristicses of convolution kernel are extracted.4 × 4 sizes are used in convolutional layer C1
Convolution kernel, acts on the image of 60 × 60 sizes, by traveling through the region of each 4 × 4 size, obtains 57 × 57 sizes
Characteristic pattern.For same characteristic pattern, the size of convolution kernel is 4 × 4.In the network model, for image block x, when
During from activation primitive ReLU, output y meets：
Y=ReLU (wx+b) formula 1
Wherein w is convolution kernel, and b is bias term, and (0, z), its image is as shown in figure 9, its effect for ReLU (z)=max
It is that independent variable z (z=wx+b) is linearly corrected.When the value of independent variable z is less than or equal to 0, at ReLU functions
After reason, the value of output is 0；When independent variable is more than 0, by after the treatment of ReLU functions, the value of output is z.Now, convolution
Data amount check on layer C1 is 16 × 57 × 57 × 60=3119040.
The down-sampled layer S1 of ■：Down-sampled treatment is carried out to the data on convolutional layer C1, by the way of maximum pond, with big
Small is that 3 × 3 convolution kernel carries out stride and is 2 aggregate statistics to the adjacent area of characteristic image, and takes the maximum in zonule
Value, to reduce the number of data, eliminates the influence of over-fitting.Now, the data amount check on S1 layers is 16 × 28 × 28 × 60=
752640。
■ convolutional layers C2：The layer choosing carries out convolution with 24 convolution kernels of 5 × 5 sizes to the view data being input into, and obtains
24 characteristic patterns (Feature map).It is 24 × 28 × 28 × 60=1128960 by the data amount check of process of convolution.
■ convolutional layers C3 and down-sampled layer S2, S3：The handling process and convolutional layer above and drop of C3, S2, S3 to data
Sample level is consistent, and simply with the increase of the number of plies, size and number for the convolution kernel of feature extraction change, to feature
The abstracting power of extraction also there occurs change.
The full articulamentums of ■：Fc4 is full articulamentum, and the characteristic on convolutional layer C3 is converted into one-dimensional vector input, is had
32 neurons.Now the number of data is 32 × 1 × 1 × 60=1920.Remaining full articulamentum fc5, fc6 and full articulamentum
The processing procedure of fc4 is similar to.
■ output layers：Output layer is connected with full articulamentum fc6, and the output x to full articulamentum fc6 uses softmax_loss
Processed, calculated sample losses value, be next exactly to carry out backpropagation according to penalty values to be progressively updated weight, instead
It is the gradient by calculating each parameter for needing change to the process propagated, is then updated by some optimized algorithms.
The weight learning rate base_lr for setting network is 0.001, and iterations is 4800, and any 120 pictures of choosing are made
It is checking pictures, wherein definition exception picture, colour cast picture and normal picture respectively have chosen 40.By calling training institute
Model is obtained to picture recognition, its accuracy rate Test-A is obtained.Errorless number/the checking of classification results of the accuracy rate of Definition Model
Sum is represented.Training result such as following table：
| Test (nicety of grading of the model to picture) | 75.4% |
| Loss (error loss function value) | 0.76 |
| Test-A (model training output accuracy) | 65.7% |
From the training data shown in subscript, this model is 75.4% to the classification results accuracy rate of picture, and by mistake
Than larger when the value of difference loss function reaches convergence, the parameter of the network model that this explanation is searched out not is optimal solution, is utilized
During model is verified, the classification results to picture are also undesirable.Therefore need to take certain measure to be allowed to optimize, from
And improve its test accuracy rate.Failure based on convolutional neural networks can be realized by three in following examples two method
The optimization of video image identification model, to improve the accuracy rate of Model Identification.
Embodiment two：
The optimization of the fault video image recognition model based on convolutional neural networks is comprised the following steps that：
Step 1, expanded image data collection
If image data set is too small, the feature extracted is not abundant enough, model for feature learning abstracting power just
Can be restricted, and then the test accuracy rate of model can be influenceed；The training set of positive and negative sample imbalance also results in training result
Accuracy rate reduction.Therefore, the model higher in order to obtain accuracy rate, the present invention is clear by 500 of script image data set
Degree is abnormal, 500 colour cast exceptions, 500 normal pictures, extends to definition 3700, picture of exception, colour cast exception picture
4150, (training of two disaggregated models need to ensure positive and negative total sample number more than 10000, wherein quality to normal picture 4150
Normal picture is positive sample, and abnormal quality picture is negative sample)；Script view data is concentrated into the clear comprising 400 of selection
The abnormal picture of degree, 400 colour cast exception pictures and 400 normal picture training datasets that totally 1200 pictures are constituted, extension
To 3000 definition exception pictures, 3500 colour cast exception pictures and 3500 normal picture totally one ten thousand pictures composition training
Data set, the ratio for controlling positive and negative sample training collection is 1/3~1；Remaining is used as test data set.For whole data set
Divide, this method uses " reserving method ", will whole data set be divided into two set of mutual exclusion, it is one of as training number
According to collection, used as test data set, the ratio of controlled training data set and test data set is 1/4~1/2 for another.Training data
Collection keeps the uniformity of data distribution with the division of test data set as far as possible, it is to avoid the diagram factor introduces extra according to partition process
Deviation and influence is produced on final result, i.e. sample class being in similar proportion, such as training data concentrate definition exception picture,
The ratio of colour cast exception picture and normal picture is 31:83:83；And test data concentrates definition exception picture, colour cast exception
The ratio of picture and normal picture also wants approximate 31:83:83.Picture size is 256 × 256, and form is jpeg image.
Step 2, adjusts prototype network structure
The convolutional neural networks model of the fault video image recognition after adjustment comprising 4 convolutional layers, 3 down-sampled layers and
2 full articulamentums.Compared to basic convolutional neural networks model, 1 convolutional layer is increased, reduce 1 full articulamentum.And
View data size to being input to network bottom layer is adjusted, and 100 × 100 are adjusted to by original 60 × 60.In model
Activation primitive and pond mode with it is consistent in basic model, do not change.Network model after adjustment is as shown in fig. 6, wherein
There are data (training data), C1 (CONVOLUTION LAYER, convolutional layer), conv1 (convolutional layer output), ReLU to activate successively
Function, S1 (POOLING LAYER, pond layer), pool1 (output of pond layer), C2 (CONVOLUTION LAYER, convolutional layer),
Conv2 (convolutional layer output), ReLU activation primitives, S2 (POOLING LAYER, pond layer), pool2 (output of pond layer), C3
(CONVOLUTION LAYER, convolutional layer), conv3 (convolutional layer output), ReLU activation primitives, C4 (CONVOLUTION
LAYER, convolutional layer), conv4 (convolutional layer output), ReLU activation primitives, S4 (POOLING LAYER, pond layer), pool4
(output of pond layer), fc5 (INNER_PRODUCT, full articulamentum), f5 (full articulamentum output), ReLU activation primitives, fc6
(INNER_PRODUCT, full articulamentum), f6 (full articulamentum output), Accuracy (training precision) and Loss (training loss).
View data (data) is input into by network bottom layer, and first passing around convolutional layer (C1) carries out process of convolution, after treatment
Data select ReLU (Rectified Linear Units) function as activation primitive, then using the data after treatment as input
Incoming pond layer (S1) carries out the mode that down-sampled, down-sampled mode is maximum pond, obtains the maximum of characteristic sub-areas.
By (wherein convolutional layer 3 does not connect S3 (POOLING below after continuous four convolutional layers and three down-sampled layer treatment
LAYER, pond layer), pool3 (output of pond layer)), data change into one-dimensional vector and are input to full articulamentum (fc5), full connection
Data after layer treatment equally select ReLU functions as activation primitive, final defeated by after continuous two full articulamentum treatment
Go out the test accuracy rate and error loss function value of model.The convolutional neural networks network optimization structure of fault video image recognition
As shown in Figure 7.Convolutional neural networks Optimization Framework on fault video image recognition model, its concrete construction is as follows：
■ is input into：To be input in network model by the image data set of size normalized, 256 × 256
The picture centre of size chooses 100 × 100 image.The number of pictures processed during setting network an iteration is 50.
■ convolutional layers C1：Compared to 16 convolution kernels of 4 × 4 sizes of basic model, from 36 volumes of 5 × 5 sizes
Product collecting image data carry out the convolution that stride is 2, obtain 36 characteristic patterns of 49 × 49 sizes, are designated as 36@49 × 49.Will volume
Product the data obtained and the computing of an activation primitive elder generation, and for it adds a bias term, allow its characteristic value for serving as neuron to deposit
Store up on feature extraction layer C1 layers.Now, the data amount check on C1 layers is 36 × 49 × 49 × 50=4321800.
The down-sampled layer S1 of ■：Down-sampled treatment is carried out to the data on convolutional layer C1, by the way of maximum pond, with big
Small is that 3 × 3 convolution kernel carries out the aggregate statistics that stride is 2 to the adjacent area of characteristic image, then in acquisition subregion
Max values.Now, the data amount check on down-sampled layer S1 is 36 × 24 × 24 × 50=1036800.
■ convolutional layers C2：The feature of image is mainly extracted, process is similar with convolutional layer C1.It is now 5 × 5 big from 96
Small convolution kernel carries out convolution to view data, obtains 96 characteristic patterns, and the number of convolution kernel changes, the characteristic pattern of acquisition
Number also accordingly increase.It is 96 × 12 × 12 × 50=829440 by the data amount check of process of convolution.
■ convolutional layers C3, C4 and down-sampled layer S2, S4：The operation principle of these layers and convolutional layer above and down-sampled
Layer is consistent, and wherein convolutional layer C3, for 3 × 3 region is gone to carry out view data convolution, obtains 128 with 128 convolution kernel sizes
Individual size is 12 × 12 characteristic image.Convolutional layer C4 is with 96 convolution kernel sizes for 3 × 3 region is gone to carry out view data
Convolution, obtains the characteristic image that 96 sizes are for 12 × 12.Down-sampled layer S2 layers is connected to after convolutional layer C2, to the spy for extracting
Levying image carries out maximum pondization treatment, and the mode in pond is identical with down-sampled layer S1, and the data for now exporting are 96 × 12 × 12
× 50=691200.Without the down-sampled layer of connection behind convolutional layer C3, directly enter data on convolutional layer C4.Down-sampled layer
After S4 is connected to convolutional layer C4, pondization operation is also carried out to convolutional layer C4, the data of output are 96 × 6 × 6 × 50=127800.
The full articulamentums of ■：Fc5 is full articulamentum, and the characteristic on down-sampled layer S4 is converted into one-dimensional vector input,
There are 528 neurons.Now the number of data is 528 × 1 × 1 × 50=26400.There are 3 neurons on full articulamentum fc6,
Equivalent to 3 species of picture.
■ output layers：Output layer is connected with full articulamentum fc6, and the output x to full articulamentum fc6 uses softmax_loss
(loss function) is processed, and calculates sample losses value, is next exactly to carry out backpropagation progressively to weight according to penalty values
It is updated, the process of backpropagation is the gradient by calculating each parameter for needing change, is then optimized by some and calculated
Method is updated..
Step 3, adjusts prototype network parameter
Weight learning rate base_lr is used to set basic learning rate, is often used in combination with adjustable strategies lr_policy,
If basic learning rate sets too big, easily across extreme point, if setting too small, local optimum, weight are easily trapped into again
Practise speed base_lr and be traditionally arranged to be 0.001~0.01；Because neural computing is not ensured that under various parameters configuration
Iteration result is restrained, when iteration result is not restrained, it is allowed to which maximum iterations max_iter, maximum iteration is set too
It is small, can cause not restrain, accuracy is very low.Set too big, concussion can be caused, lose time, preferred maximum iteration
Scope is 10000~20000.For the network structure after optimization, it is 0.001 to set its weight learning rate base_lr, maximum
Iterations max_iter is 10000.Randomly select 600 pictures to be verified, be divided into 5 groups, every group 120, and these
Picture is different from the picture for training and testing.Wherein every group definition abnormal image, colour cast image and normal picture have
40.Call training gained model to their Classification and Identifications, the accuracy rate Test-A of the model is obtained by calculating.After optimization
Convolutional neural networks model is compared as follows shown in table with basic convolutional neural networks model test results.
CNN (convolutional neural networks) models and basis CNN (convolutional neural networks) Model Identification accuracy rate ratios after optimization
It is more as shown in Figure 8.
Observation data above can be known, by image data set extension, network hierarchical structure adjustment and model parameter
The measure such as change after, compared with basic model, the test accuracy rate of model is significantly increased.And the error of model is damaged
Function convergence to a minimum value is lost, the network parameter of model obtains optimal solution, illustrates that the model can be realized to Abnormal Map
The identification of picture and classification.Fig. 8 results prove that the model can reach 90% or so to the accuracy rate that image classification is recognized, remote high
In the experimental result of basic model, model realization optimization.
The convolutional neural networks model of the fault video image recognition, detects and depth by by video image quality
Habit is blended, and convolutional neural networks are applied in the middle of urban track traffic monitoring fault video detection.Test result indicate that, should
Method Detection accuracy is high, easy to operate.Joined by building network structure, Optimum learning rate, number of training and iterations etc.
Number, can be effectively ensured the Detection results of model, and the method has very big researching value and practice significance.During specific implementation,
Technical solution of the present invention can be realized using software mode.
Claims (9)
1. a kind of urban track traffic overall view monitoring video fault detection method based on deep learning, it is characterised in that：Including
Data set building process and model training generating process, picture classification identification process,
The data set building process, including to definition anomalous video, the colour cast in urban track traffic overall view monitoring video
Anomalous video and normal video are processed, and video is converted into single-frame images, and dimension of picture is normalized,
Definition abnormal image data set, colour cast abnormal image data set and normal image data collection are built, arbitrarily abstract image
Training set and test set are divided into, the image data set needed for obtaining；
The model training generating process, including model training and model measurement,
The model training, including the fault video image recognition model based on convolutional neural networks is trained, the volume
Product neutral net includes multiple convolutional layers and multiple full articulamentums；Training method is as follows,
Using the picture in training set by being input to convolutional neural networks, multiple convolution as bottom data after Data Format Transform
Level joins, and the view data to being input into carries out convolution algorithm, abstract image feature, the multiple characteristic patterns of generation；
Used as input, be fully converted into for feature by each full articulamentum by incoming full articulamentum for the output of last convolutional layer
One-dimensional vector is exported；Determine the error between real output value and desired value, backpropagation, adjustment network ginseng are carried out to network
Number；
Then whether training of judgement error restrains, if otherwise returning to training initiating terminal input picture carries out feature learning, if then
Whether judgment models training iterations reaches predetermined iterations, when predetermined iterations is reached, model training knot
Beam, otherwise returns to training initiating terminal input picture and continues to train；
Training obtains the fault video image recognition model based on convolutional neural networks after terminating；
The model measurement, including the picture input model in test set is trained the net of gained fault video image recognition model
Network bottom, fault video image recognition model exports the classification results of picture, includes quality normally and the species of abnormal quality two
Not, finally compare with the respective labels that have set in test set, calculate the accuracy rate of test, if being not reaching to be expected,
Then optimize fault video image recognition model, until reaching expection, obtain final fault video image recognition model；
The picture classification identification process, including the failure that single frames picture input model training generating process to be identified is obtained
The network bottom layer of video image identification model, fault video image recognition model exports the classification results of picture, and city is realized in completion
City's track traffic overall view monitoring video failure picture is detected.
2. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 1, its
It is characterised by：The optimization fault video image recognition model, including the extension for carrying out image data set.
3. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 1, its
It is characterised by：The optimization fault video image recognition model, including increase the number of convolutional layer, reduce the individual of full articulamentum
Number.
4. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 3, its
It is characterised by：Original fault video image recognition model, the convolutional neural networks include 3 convolutional layers and 3 full connections
Layer.
5. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 4, its
It is characterised by：Original fault video image recognition model includes that input, convolutional layer C1, down-sampled layer S1, convolutional layer C2, drop are adopted
Sample layer S2, convolutional layer C3, down-sampled layer S3, full articulamentum fc4, full articulamentum fc5, full articulamentum fc6 and output layer.
6. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 4, its
It is characterised by：It is described to be changed to convolutional neural networks to include 4 convolutional layers and 2 during optimization fault video image recognition model
Full articulamentum.
7. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 6, its
It is characterised by：During optimization fault video image recognition model, convolutional neural networks are changed to include that input, convolutional layer C1, drop are adopted
Sample layer S1, convolutional layer C2, down-sampled layer S2, convolutional layer C3, down-sampled layer S3, convolutional layer C4, down-sampled layer S4, full articulamentum
Fc5, full articulamentum fc6 and output layer.
8. the urban track traffic overall view monitoring video fault detection method of deep learning is based on according to claim 1, its
It is characterised by：The optimization fault video image recognition model, including the adjustment for carrying out prototype network parameter.
9. the urban track traffic panorama based on deep learning is supervised according to claim 1 or 2 or 3 or 4 or 5 or 6 or 7 or 8
Control video fault detection method, it is characterised in that：Whether the training of judgement error restrains, and is realized based on error loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611125030.0A CN106709511A (en) | 2016-12-08 | 2016-12-08 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611125030.0A CN106709511A (en) | 2016-12-08 | 2016-12-08 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106709511A true CN106709511A (en) | 2017-05-24 |
Family
ID=58936000
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611125030.0A Pending CN106709511A (en) | 2016-12-08 | 2016-12-08 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106709511A (en) |
Cited By (82)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107316004A (en) * | 2017-06-06 | 2017-11-03 | 西北工业大学 | Space Target Recognition based on deep learning |
| CN107316295A (en) * | 2017-07-02 | 2017-11-03 | 苏州大学 | A kind of fabric defects detection method based on deep neural network |
| CN107396094A (en) * | 2017-08-17 | 2017-11-24 | 上海大学 | The automatic testing method of single camera damage towards in multi-cam monitoring system |
| CN107506695A (en) * | 2017-07-28 | 2017-12-22 | 武汉理工大学 | Video monitoring equipment failure automatic detection method |
| CN107729363A (en) * | 2017-09-06 | 2018-02-23 | 上海交通大学 | Based on GoogLeNet network model birds population identifying and analyzing methods |
| CN108337505A (en) * | 2018-01-30 | 2018-07-27 | 百度在线网络技术（北京）有限公司 | Information acquisition method and device |
| CN108376268A (en) * | 2018-05-21 | 2018-08-07 | 北京京天威科技发展有限公司 | Subway electric motor coach car team examines underbody, vehicle side fault detection system and method |
| CN108460399A (en) * | 2017-12-29 | 2018-08-28 | 华南师范大学 | A kind of child building block builds householder method and system |
| CN108667747A (en) * | 2018-04-28 | 2018-10-16 | 深圳信息职业技术学院 | The method, apparatus and computer readable storage medium of network flow application type identification |
| CN108665093A (en) * | 2018-04-19 | 2018-10-16 | 东南大学 | Traffic accidents Severity forecasting method based on deep learning |
| CN108674453A (en) * | 2018-05-23 | 2018-10-19 | 深圳市阡丘越科技有限公司 | A kind of rail traffic monitoring system and control method based on AI technologies |
| CN108764270A (en) * | 2018-04-03 | 2018-11-06 | 上海大学 | A kind of Information Hiding & Detecting method integrated using convolutional neural networks |
| CN108846332A (en) * | 2018-05-30 | 2018-11-20 | 西南交通大学 | A kind of railway drivers Activity recognition method based on CLSTA |
| CN108985328A (en) * | 2018-06-08 | 2018-12-11 | 佛山市顺德区中山大学研究院 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
| CN108985456A (en) * | 2018-07-25 | 2018-12-11 | 大国创新智能科技(东莞)有限公司 | The number of plies increases and decreases deep learning neural network training method, system, medium and equipment |
| CN109034264A (en) * | 2018-08-15 | 2018-12-18 | 云南大学 | Traffic accident seriousness predicts CSP-CNN model and its modeling method |
| CN109145134A (en) * | 2018-07-27 | 2019-01-04 | 郑州云海信息技术有限公司 | A kind of method and device of detection device failure |
| CN109189767A (en) * | 2018-08-01 | 2019-01-11 | 北京三快在线科技有限公司 | Data processing method, device, electronic equipment and storage medium |
| CN109214308A (en) * | 2018-08-15 | 2019-01-15 | 武汉唯理科技有限公司 | A kind of traffic abnormity image identification method based on focal loss function |
| CN109327479A (en) * | 2018-12-14 | 2019-02-12 | 锐捷网络股份有限公司 | Encrypt recognition methods and the device of stream |
| CN109657600A (en) * | 2018-12-14 | 2019-04-19 | 广东工业大学 | A kind of video area removes altering detecting method and device |
| CN109711377A (en) * | 2018-12-30 | 2019-05-03 | 陕西师范大学 | Standardize examinee's positioning and method of counting in the single-frame images of examination hall monitoring |
| CN109902640A (en) * | 2019-03-05 | 2019-06-18 | 江南大学 | Video quality abnormality detection system and its detection method based on edge calculations and machine learning |
| CN109934106A (en) * | 2019-01-30 | 2019-06-25 | 长视科技股份有限公司 | A kind of user behavior analysis method based on video image deep learning |
| CN109993048A (en) * | 2018-11-26 | 2019-07-09 | 普达迪泰（天津）智能装备科技有限公司 | A kind of bituminous pavement striped filter method and detection system |
| CN110032662A (en) * | 2019-05-16 | 2019-07-19 | 福州大学 | Video image distortion effect model building method based on improved dice loss function |
| CN110097081A (en) * | 2019-03-29 | 2019-08-06 | 广州思德医疗科技有限公司 | A kind of construction method and device of training set |
| CN110097080A (en) * | 2019-03-29 | 2019-08-06 | 广州思德医疗科技有限公司 | A kind of construction method and device of tag along sort |
| CN110131109A (en) * | 2019-04-25 | 2019-08-16 | 浙江大学 | A kind of pneumatic equipment bladess unbalance detection based on convolutional neural networks |
| CN110186471A (en) * | 2019-05-06 | 2019-08-30 | 平安科技（深圳）有限公司 | Air navigation aid, device, computer equipment and storage medium based on history video |
| CN110213788A (en) * | 2019-06-15 | 2019-09-06 | 福州大学 | WSN abnormality detection and kind identification method based on data flow space-time characteristic |
| CN110348337A (en) * | 2019-06-26 | 2019-10-18 | 西安理工大学 | The detection of dancing video human skeleton and extracting method based on convolutional neural networks |
| CN110782692A (en) * | 2019-10-31 | 2020-02-11 | 青岛海信网络科技股份有限公司 | Signal lamp fault detection method and system |
| CN110794254A (en) * | 2018-08-01 | 2020-02-14 | 北京映翰通网络技术股份有限公司 | Power distribution network fault prediction method and system based on reinforcement learning |
| CN110826491A (en) * | 2019-11-07 | 2020-02-21 | 北京工业大学 | Video key frame detection method based on cascading manual features and depth features |
| CN110826636A (en) * | 2019-11-12 | 2020-02-21 | 吉林省民航机场集团公司 | Aircraft anomaly detection system and anomaly detection method thereof |
| CN110942448A (en) * | 2019-10-30 | 2020-03-31 | 江苏大学 | Quantitative phase image identification method based on convolutional neural network |
| CN110988597A (en) * | 2019-12-15 | 2020-04-10 | 云南电网有限责任公司文山供电局 | Resonance type detection method based on neural network |
| CN111144464A (en) * | 2019-12-17 | 2020-05-12 | 金陵科技学院 | Fruit automatic identification method based on CNN-Kmeans algorithm |
| WO2020093211A1 (en) * | 2018-11-05 | 2020-05-14 | 中国科学院计算技术研究所 | Kronecker convolution-based scene segmentation method and system |
| CN111353394A (en) * | 2020-02-20 | 2020-06-30 | 中山大学 | Video behavior identification method based on three-dimensional alternate update network |
| CN111353947A (en) * | 2018-12-24 | 2020-06-30 | 深圳先进技术研究院 | Magnetic resonance parallel imaging method and related equipment |
| CN111369555A (en) * | 2020-03-19 | 2020-07-03 | 昆明理工大学 | Video quality diagnosis method based on deep learning |
| CN111429424A (en) * | 2020-03-20 | 2020-07-17 | 中冶赛迪重庆信息技术有限公司 | Heating furnace inlet abnormity identification method based on deep learning |
| CN111505424A (en) * | 2020-05-06 | 2020-08-07 | 哈尔滨工业大学 | Large experimental device power equipment fault diagnosis method based on deep convolutional neural network |
| CN111563886A (en) * | 2020-04-29 | 2020-08-21 | 安徽国钜工程机械科技有限公司 | Tunnel steel rail surface disease detection method and device based on unsupervised feature learning |
| CN111723739A (en) * | 2020-06-19 | 2020-09-29 | 艾索信息股份有限公司 | Congestion state monitoring method in bus based on convolutional neural network |
| CN111738044A (en) * | 2020-01-06 | 2020-10-02 | 西北大学 | Campus violence assessment method based on deep learning behavior recognition |
| CN111914634A (en) * | 2020-06-23 | 2020-11-10 | 华南理工大学 | Complex-scene-interference-resistant automatic manhole cover type detection method and system |
| CN112101383A (en) * | 2020-09-15 | 2020-12-18 | 遵义师范学院 | Color cast image identification method |
| CN112097365A (en) * | 2020-07-10 | 2020-12-18 | 珠海派诺科技股份有限公司 | Air conditioner fault detection and identification method and device based on prediction and classification model |
| CN112115824A (en) * | 2020-09-07 | 2020-12-22 | 北京豆牛网络科技有限公司 | Fruit and vegetable detection method and device, electronic equipment and computer readable medium |
| CN112183193A (en) * | 2020-08-19 | 2021-01-05 | 中冶南方工程技术有限公司 | Method for monitoring working state of blast furnace tuyere |
| CN112241470A (en) * | 2020-09-24 | 2021-01-19 | 北京影谱科技股份有限公司 | Video classification method and system |
| CN112434730A (en) * | 2020-11-10 | 2021-03-02 | 广东电力信息科技有限公司 | GoogleNet-based video image quality abnormity classification method |
| WO2021068178A1 (en) * | 2019-10-11 | 2021-04-15 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for image quality detection |
| CN112733676A (en) * | 2020-12-31 | 2021-04-30 | 青岛海纳云科技控股有限公司 | Method for detecting and identifying garbage in elevator based on deep learning |
| CN112784805A (en) * | 2021-02-02 | 2021-05-11 | 郑州地铁集团有限公司 | Pantograph video analysis system and method based on image recognition |
| CN112794174A (en) * | 2019-11-13 | 2021-05-14 | 深圳市赛特尔信息技术有限公司 | Real-time video judgment elevator door opening and closing abnormity scheme based on big data |
| CN112836719A (en) * | 2020-12-11 | 2021-05-25 | 南京富岛信息工程有限公司 | Indicator diagram similarity detection method fusing two classifications and three groups |
| CN112949770A (en) * | 2021-04-08 | 2021-06-11 | 深圳市医诺智能科技发展有限公司 | Medical image identification and classification method and terminal |
| CN113011315A (en) * | 2021-03-16 | 2021-06-22 | 华南理工大学 | Subway track identification method based on ultra-fast structure perception depth network |
| CN113139932A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Deep learning defect image identification method and system based on ensemble learning |
| CN113177614A (en) * | 2021-05-27 | 2021-07-27 | 广州白云电器设备股份有限公司 | Image recognition system and method for power supply switch cabinet of urban rail transit |
| CN113191952A (en) * | 2021-06-02 | 2021-07-30 | 重庆渝微电子技术研究院有限公司 | Spliced image evaluation method for 360-degree panoramic looking-around equipment |
| CN113269111A (en) * | 2021-06-03 | 2021-08-17 | 昆山杜克大学 | Elevator abnormal behavior detection method and system based on video monitoring |
| CN113313428A (en) * | 2021-06-29 | 2021-08-27 | 上海煤科信息科技有限公司 | AI fault monitoring method and system of smart mine based on big data |
| CN113379608A (en) * | 2020-03-10 | 2021-09-10 | Tcl科技集团股份有限公司 | Image processing method, storage medium and terminal equipment |
| CN113392853A (en) * | 2021-05-28 | 2021-09-14 | 中汽研（天津）汽车工程研究院有限公司 | Door closing sound quality evaluation and identification method based on image identification |
| CN113505821A (en) * | 2021-06-29 | 2021-10-15 | 重庆邮电大学 | Deep neural network image identification method and system based on sample reliability |
| CN113520810A (en) * | 2020-04-14 | 2021-10-22 | 中国移动通信集团浙江有限公司 | Method and device for assisting visually impaired people to plan advancing action and computing equipment |
| CN113537306A (en) * | 2021-06-29 | 2021-10-22 | 复旦大学 | Image classification method based on progressive growth element learning |
| CN113706524A (en) * | 2021-09-17 | 2021-11-26 | 上海交通大学 | Convolutional neural network reproduction image detection system improved based on continuous learning method |
| CN113762084A (en) * | 2021-08-09 | 2021-12-07 | 江苏大学 | Building night scene light abnormity detection method based on RetinaXNet |
| CN114581861A (en) * | 2022-03-02 | 2022-06-03 | 北京交通大学 | Track area identification method based on deep learning convolutional neural network |
| CN115331155A (en) * | 2022-10-14 | 2022-11-11 | 智慧齐鲁(山东)大数据科技有限公司 | Mass video monitoring point location graph state detection method and system |
| CN115527134A (en) * | 2022-10-27 | 2022-12-27 | 浙江九烁光电工程技术有限公司 | Urban garden landscape lighting monitoring system and method based on big data |
| CN115731513A (en) * | 2022-11-15 | 2023-03-03 | 杭州电子科技大学信息工程学院 | Intelligent park management system based on digital twin |
| CN116542956A (en) * | 2023-05-25 | 2023-08-04 | 广州机智云物联网科技有限公司 | Automatic detection method and system for fabric components and readable storage medium |
| CN116778395A (en) * | 2023-08-21 | 2023-09-19 | 成都理工大学 | Mountain torrent flood video identification monitoring method based on deep learning |
| CN116972914A (en) * | 2023-09-22 | 2023-10-31 | 华夏天信智能物联股份有限公司 | Intelligent testing method and system for frequency conversion integrated machine |
| CN117095411A (en) * | 2023-10-16 | 2023-11-21 | 青岛文达通科技股份有限公司 | Detection method and system based on image fault recognition |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102226907A (en) * | 2011-05-24 | 2011-10-26 | 武汉嘉业恒科技有限公司 | License plate positioning method and apparatus based on multiple characteristics |
| CN105139028A (en) * | 2015-08-13 | 2015-12-09 | 西安电子科技大学 | SAR image classification method based on hierarchical sparse filtering convolutional neural network |
-
2016
- 2016-12-08 CN CN201611125030.0A patent/CN106709511A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102226907A (en) * | 2011-05-24 | 2011-10-26 | 武汉嘉业恒科技有限公司 | License plate positioning method and apparatus based on multiple characteristics |
| CN105139028A (en) * | 2015-08-13 | 2015-12-09 | 西安电子科技大学 | SAR image classification method based on hierarchical sparse filtering convolutional neural network |
Non-Patent Citations (1)
| Title |
|---|
| 邬美银等: "基于卷积神经网络的视频图像失真检测及分类", 《计算机应用研究》 * |
Cited By (110)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107316004A (en) * | 2017-06-06 | 2017-11-03 | 西北工业大学 | Space Target Recognition based on deep learning |
| CN107316295A (en) * | 2017-07-02 | 2017-11-03 | 苏州大学 | A kind of fabric defects detection method based on deep neural network |
| CN107506695A (en) * | 2017-07-28 | 2017-12-22 | 武汉理工大学 | Video monitoring equipment failure automatic detection method |
| CN107396094A (en) * | 2017-08-17 | 2017-11-24 | 上海大学 | The automatic testing method of single camera damage towards in multi-cam monitoring system |
| CN107396094B (en) * | 2017-08-17 | 2019-02-22 | 上海大学 | Automatic testing method towards camera single in multi-cam monitoring system damage |
| CN107729363A (en) * | 2017-09-06 | 2018-02-23 | 上海交通大学 | Based on GoogLeNet network model birds population identifying and analyzing methods |
| CN108460399A (en) * | 2017-12-29 | 2018-08-28 | 华南师范大学 | A kind of child building block builds householder method and system |
| CN108337505A (en) * | 2018-01-30 | 2018-07-27 | 百度在线网络技术（北京）有限公司 | Information acquisition method and device |
| CN108764270B (en) * | 2018-04-03 | 2022-06-14 | 上海大学 | Information hiding detection method integrated by using convolutional neural network |
| CN108764270A (en) * | 2018-04-03 | 2018-11-06 | 上海大学 | A kind of Information Hiding & Detecting method integrated using convolutional neural networks |
| CN108665093A (en) * | 2018-04-19 | 2018-10-16 | 东南大学 | Traffic accidents Severity forecasting method based on deep learning |
| CN108665093B (en) * | 2018-04-19 | 2021-07-27 | 东南大学 | Deep learning-based expressway traffic accident severity prediction method |
| CN108667747A (en) * | 2018-04-28 | 2018-10-16 | 深圳信息职业技术学院 | The method, apparatus and computer readable storage medium of network flow application type identification |
| CN108376268A (en) * | 2018-05-21 | 2018-08-07 | 北京京天威科技发展有限公司 | Subway electric motor coach car team examines underbody, vehicle side fault detection system and method |
| CN108674453A (en) * | 2018-05-23 | 2018-10-19 | 深圳市阡丘越科技有限公司 | A kind of rail traffic monitoring system and control method based on AI technologies |
| CN108846332A (en) * | 2018-05-30 | 2018-11-20 | 西南交通大学 | A kind of railway drivers Activity recognition method based on CLSTA |
| CN108985328A (en) * | 2018-06-08 | 2018-12-11 | 佛山市顺德区中山大学研究院 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
| CN108985456A (en) * | 2018-07-25 | 2018-12-11 | 大国创新智能科技(东莞)有限公司 | The number of plies increases and decreases deep learning neural network training method, system, medium and equipment |
| CN109145134A (en) * | 2018-07-27 | 2019-01-04 | 郑州云海信息技术有限公司 | A kind of method and device of detection device failure |
| CN109189767B (en) * | 2018-08-01 | 2021-07-23 | 北京三快在线科技有限公司 | Data processing method and device, electronic equipment and storage medium |
| CN109189767A (en) * | 2018-08-01 | 2019-01-11 | 北京三快在线科技有限公司 | Data processing method, device, electronic equipment and storage medium |
| CN110794254A (en) * | 2018-08-01 | 2020-02-14 | 北京映翰通网络技术股份有限公司 | Power distribution network fault prediction method and system based on reinforcement learning |
| CN109034264B (en) * | 2018-08-15 | 2021-11-19 | 云南大学 | CSP-CNN model for predicting severity of traffic accident and modeling method thereof |
| CN109214308A (en) * | 2018-08-15 | 2019-01-15 | 武汉唯理科技有限公司 | A kind of traffic abnormity image identification method based on focal loss function |
| CN109034264A (en) * | 2018-08-15 | 2018-12-18 | 云南大学 | Traffic accident seriousness predicts CSP-CNN model and its modeling method |
| WO2020093211A1 (en) * | 2018-11-05 | 2020-05-14 | 中国科学院计算技术研究所 | Kronecker convolution-based scene segmentation method and system |
| CN109993048A (en) * | 2018-11-26 | 2019-07-09 | 普达迪泰（天津）智能装备科技有限公司 | A kind of bituminous pavement striped filter method and detection system |
| CN109657600B (en) * | 2018-12-14 | 2024-02-27 | 广东工业大学 | Video area removal tampering detection method and device |
| CN109327479A (en) * | 2018-12-14 | 2019-02-12 | 锐捷网络股份有限公司 | Encrypt recognition methods and the device of stream |
| CN109657600A (en) * | 2018-12-14 | 2019-04-19 | 广东工业大学 | A kind of video area removes altering detecting method and device |
| CN111353947A (en) * | 2018-12-24 | 2020-06-30 | 深圳先进技术研究院 | Magnetic resonance parallel imaging method and related equipment |
| CN109711377A (en) * | 2018-12-30 | 2019-05-03 | 陕西师范大学 | Standardize examinee's positioning and method of counting in the single-frame images of examination hall monitoring |
| CN109711377B (en) * | 2018-12-30 | 2023-04-18 | 陕西师范大学 | Method for positioning and counting examinees in single-frame image monitored by standardized examination room |
| CN109934106A (en) * | 2019-01-30 | 2019-06-25 | 长视科技股份有限公司 | A kind of user behavior analysis method based on video image deep learning |
| CN109902640A (en) * | 2019-03-05 | 2019-06-18 | 江南大学 | Video quality abnormality detection system and its detection method based on edge calculations and machine learning |
| CN109902640B (en) * | 2019-03-05 | 2023-06-30 | 江南大学 | Video quality anomaly detection system and detection method based on edge calculation and machine learning |
| CN110097081A (en) * | 2019-03-29 | 2019-08-06 | 广州思德医疗科技有限公司 | A kind of construction method and device of training set |
| CN110097080A (en) * | 2019-03-29 | 2019-08-06 | 广州思德医疗科技有限公司 | A kind of construction method and device of tag along sort |
| CN110097080B (en) * | 2019-03-29 | 2021-04-13 | 广州思德医疗科技有限公司 | Construction method and device of classification label |
| CN110131109A (en) * | 2019-04-25 | 2019-08-16 | 浙江大学 | A kind of pneumatic equipment bladess unbalance detection based on convolutional neural networks |
| CN110186471A (en) * | 2019-05-06 | 2019-08-30 | 平安科技（深圳）有限公司 | Air navigation aid, device, computer equipment and storage medium based on history video |
| CN110032662A (en) * | 2019-05-16 | 2019-07-19 | 福州大学 | Video image distortion effect model building method based on improved dice loss function |
| CN110213788A (en) * | 2019-06-15 | 2019-09-06 | 福州大学 | WSN abnormality detection and kind identification method based on data flow space-time characteristic |
| CN110348337A (en) * | 2019-06-26 | 2019-10-18 | 西安理工大学 | The detection of dancing video human skeleton and extracting method based on convolutional neural networks |
| WO2021068178A1 (en) * | 2019-10-11 | 2021-04-15 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for image quality detection |
| CN110942448A (en) * | 2019-10-30 | 2020-03-31 | 江苏大学 | Quantitative phase image identification method based on convolutional neural network |
| CN110782692A (en) * | 2019-10-31 | 2020-02-11 | 青岛海信网络科技股份有限公司 | Signal lamp fault detection method and system |
| CN110826491A (en) * | 2019-11-07 | 2020-02-21 | 北京工业大学 | Video key frame detection method based on cascading manual features and depth features |
| CN110826636A (en) * | 2019-11-12 | 2020-02-21 | 吉林省民航机场集团公司 | Aircraft anomaly detection system and anomaly detection method thereof |
| CN112794174A (en) * | 2019-11-13 | 2021-05-14 | 深圳市赛特尔信息技术有限公司 | Real-time video judgment elevator door opening and closing abnormity scheme based on big data |
| CN110988597A (en) * | 2019-12-15 | 2020-04-10 | 云南电网有限责任公司文山供电局 | Resonance type detection method based on neural network |
| CN111144464A (en) * | 2019-12-17 | 2020-05-12 | 金陵科技学院 | Fruit automatic identification method based on CNN-Kmeans algorithm |
| CN111738044A (en) * | 2020-01-06 | 2020-10-02 | 西北大学 | Campus violence assessment method based on deep learning behavior recognition |
| CN111738044B (en) * | 2020-01-06 | 2024-01-19 | 西北大学 | Campus violence assessment method based on deep learning behavior recognition |
| CN111353394A (en) * | 2020-02-20 | 2020-06-30 | 中山大学 | Video behavior identification method based on three-dimensional alternate update network |
| CN111353394B (en) * | 2020-02-20 | 2023-05-23 | 中山大学 | Video behavior recognition method based on three-dimensional alternate update network |
| CN113379608A (en) * | 2020-03-10 | 2021-09-10 | Tcl科技集团股份有限公司 | Image processing method, storage medium and terminal equipment |
| CN111369555A (en) * | 2020-03-19 | 2020-07-03 | 昆明理工大学 | Video quality diagnosis method based on deep learning |
| CN111429424A (en) * | 2020-03-20 | 2020-07-17 | 中冶赛迪重庆信息技术有限公司 | Heating furnace inlet abnormity identification method based on deep learning |
| CN111429424B (en) * | 2020-03-20 | 2023-08-11 | 中冶赛迪信息技术(重庆)有限公司 | Heating furnace inlet anomaly identification method based on deep learning |
| CN113520810A (en) * | 2020-04-14 | 2021-10-22 | 中国移动通信集团浙江有限公司 | Method and device for assisting visually impaired people to plan advancing action and computing equipment |
| CN111563886B (en) * | 2020-04-29 | 2023-04-07 | 安徽国钜工程机械科技有限公司 | Unsupervised feature learning-based tunnel steel rail surface disease detection method and device |
| CN111563886A (en) * | 2020-04-29 | 2020-08-21 | 安徽国钜工程机械科技有限公司 | Tunnel steel rail surface disease detection method and device based on unsupervised feature learning |
| CN111505424A (en) * | 2020-05-06 | 2020-08-07 | 哈尔滨工业大学 | Large experimental device power equipment fault diagnosis method based on deep convolutional neural network |
| CN111723739A (en) * | 2020-06-19 | 2020-09-29 | 艾索信息股份有限公司 | Congestion state monitoring method in bus based on convolutional neural network |
| CN111914634B (en) * | 2020-06-23 | 2023-06-20 | 华南理工大学 | Automatic detection method and system for well lid class resisting complex scene interference |
| CN111914634A (en) * | 2020-06-23 | 2020-11-10 | 华南理工大学 | Complex-scene-interference-resistant automatic manhole cover type detection method and system |
| CN112097365A (en) * | 2020-07-10 | 2020-12-18 | 珠海派诺科技股份有限公司 | Air conditioner fault detection and identification method and device based on prediction and classification model |
| CN112183193A (en) * | 2020-08-19 | 2021-01-05 | 中冶南方工程技术有限公司 | Method for monitoring working state of blast furnace tuyere |
| CN112115824A (en) * | 2020-09-07 | 2020-12-22 | 北京豆牛网络科技有限公司 | Fruit and vegetable detection method and device, electronic equipment and computer readable medium |
| CN112101383B (en) * | 2020-09-15 | 2022-10-21 | 遵义师范学院 | Color cast image identification method |
| CN112101383A (en) * | 2020-09-15 | 2020-12-18 | 遵义师范学院 | Color cast image identification method |
| CN112241470A (en) * | 2020-09-24 | 2021-01-19 | 北京影谱科技股份有限公司 | Video classification method and system |
| CN112241470B (en) * | 2020-09-24 | 2024-02-02 | 北京影谱科技股份有限公司 | Video classification method and system |
| CN112434730A (en) * | 2020-11-10 | 2021-03-02 | 广东电力信息科技有限公司 | GoogleNet-based video image quality abnormity classification method |
| CN112836719B (en) * | 2020-12-11 | 2024-01-05 | 南京富岛信息工程有限公司 | Indicator diagram similarity detection method integrating two classifications and triplets |
| CN112836719A (en) * | 2020-12-11 | 2021-05-25 | 南京富岛信息工程有限公司 | Indicator diagram similarity detection method fusing two classifications and three groups |
| CN112733676A (en) * | 2020-12-31 | 2021-04-30 | 青岛海纳云科技控股有限公司 | Method for detecting and identifying garbage in elevator based on deep learning |
| CN112784805A (en) * | 2021-02-02 | 2021-05-11 | 郑州地铁集团有限公司 | Pantograph video analysis system and method based on image recognition |
| CN113011315B (en) * | 2021-03-16 | 2022-12-16 | 华南理工大学 | Subway track identification method based on ultra-fast structure perception depth network |
| CN113011315A (en) * | 2021-03-16 | 2021-06-22 | 华南理工大学 | Subway track identification method based on ultra-fast structure perception depth network |
| CN113139932A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Deep learning defect image identification method and system based on ensemble learning |
| CN112949770B (en) * | 2021-04-08 | 2023-12-26 | 深圳市医诺智能科技发展有限公司 | Medical image identification and classification method and terminal |
| CN112949770A (en) * | 2021-04-08 | 2021-06-11 | 深圳市医诺智能科技发展有限公司 | Medical image identification and classification method and terminal |
| CN113177614A (en) * | 2021-05-27 | 2021-07-27 | 广州白云电器设备股份有限公司 | Image recognition system and method for power supply switch cabinet of urban rail transit |
| CN113392853A (en) * | 2021-05-28 | 2021-09-14 | 中汽研（天津）汽车工程研究院有限公司 | Door closing sound quality evaluation and identification method based on image identification |
| CN113191952A (en) * | 2021-06-02 | 2021-07-30 | 重庆渝微电子技术研究院有限公司 | Spliced image evaluation method for 360-degree panoramic looking-around equipment |
| CN113269111A (en) * | 2021-06-03 | 2021-08-17 | 昆山杜克大学 | Elevator abnormal behavior detection method and system based on video monitoring |
| CN113269111B (en) * | 2021-06-03 | 2024-04-05 | 昆山杜克大学 | Video monitoring-based elevator abnormal behavior detection method and system |
| CN113537306A (en) * | 2021-06-29 | 2021-10-22 | 复旦大学 | Image classification method based on progressive growth element learning |
| CN113505821A (en) * | 2021-06-29 | 2021-10-15 | 重庆邮电大学 | Deep neural network image identification method and system based on sample reliability |
| CN113505821B (en) * | 2021-06-29 | 2022-09-27 | 重庆邮电大学 | Deep neural network image identification method and system based on sample reliability |
| CN113313428A (en) * | 2021-06-29 | 2021-08-27 | 上海煤科信息科技有限公司 | AI fault monitoring method and system of smart mine based on big data |
| CN113762084A (en) * | 2021-08-09 | 2021-12-07 | 江苏大学 | Building night scene light abnormity detection method based on RetinaXNet |
| CN113706524A (en) * | 2021-09-17 | 2021-11-26 | 上海交通大学 | Convolutional neural network reproduction image detection system improved based on continuous learning method |
| CN113706524B (en) * | 2021-09-17 | 2023-11-14 | 上海交通大学 | Convolutional neural network image-flipping detection system based on continuous learning method improvement |
| CN114581861A (en) * | 2022-03-02 | 2022-06-03 | 北京交通大学 | Track area identification method based on deep learning convolutional neural network |
| CN114581861B (en) * | 2022-03-02 | 2023-05-23 | 北京交通大学 | Rail region identification method based on deep learning convolutional neural network |
| CN115331155A (en) * | 2022-10-14 | 2022-11-11 | 智慧齐鲁(山东)大数据科技有限公司 | Mass video monitoring point location graph state detection method and system |
| CN115331155B (en) * | 2022-10-14 | 2023-02-03 | 智慧齐鲁(山东)大数据科技有限公司 | Mass video monitoring point location graph state detection method and system |
| CN115527134A (en) * | 2022-10-27 | 2022-12-27 | 浙江九烁光电工程技术有限公司 | Urban garden landscape lighting monitoring system and method based on big data |
| CN115731513A (en) * | 2022-11-15 | 2023-03-03 | 杭州电子科技大学信息工程学院 | Intelligent park management system based on digital twin |
| CN116542956A (en) * | 2023-05-25 | 2023-08-04 | 广州机智云物联网科技有限公司 | Automatic detection method and system for fabric components and readable storage medium |
| CN116542956B (en) * | 2023-05-25 | 2023-11-17 | 广州机智云物联网科技有限公司 | Automatic detection method and system for fabric components and readable storage medium |
| CN116778395B (en) * | 2023-08-21 | 2023-10-24 | 成都理工大学 | Mountain torrent flood video identification monitoring method based on deep learning |
| CN116778395A (en) * | 2023-08-21 | 2023-09-19 | 成都理工大学 | Mountain torrent flood video identification monitoring method based on deep learning |
| CN116972914B (en) * | 2023-09-22 | 2023-12-26 | 华夏天信智能物联股份有限公司 | Intelligent testing method and system for frequency conversion integrated machine |
| CN116972914A (en) * | 2023-09-22 | 2023-10-31 | 华夏天信智能物联股份有限公司 | Intelligent testing method and system for frequency conversion integrated machine |
| CN117095411B (en) * | 2023-10-16 | 2024-01-23 | 青岛文达通科技股份有限公司 | Detection method and system based on image fault recognition |
| CN117095411A (en) * | 2023-10-16 | 2023-11-21 | 青岛文达通科技股份有限公司 | Detection method and system based on image fault recognition |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106709511A (en) | Urban rail transit panoramic monitoring video fault detection method based on depth learning | |
| CN105825235B (en) | A kind of image-recognizing method based on multi-characteristic deep learning | |
| CN112232476B (en) | Method and device for updating test sample set | |
| CN112614077B (en) | Unsupervised low-illumination image enhancement method based on generation countermeasure network | |
| CN110349136A (en) | A kind of tampered image detection method based on deep learning | |
| CN107016406A (en) | The pest and disease damage image generating method of network is resisted based on production | |
| CN107529650A (en) | The structure and closed loop detection method of network model, related device and computer equipment | |
| CN108416327A (en) | A kind of object detection method, device, computer equipment and readable storage medium storing program for executing | |
| CN108009594B (en) | A kind of image-recognizing method based on change grouping convolution | |
| CN109346159A (en) | Case image classification method, device, computer equipment and storage medium | |
| WO2021051987A1 (en) | Method and apparatus for training neural network model | |
| CN109472193A (en) | Method for detecting human face and device | |
| CN109801232A (en) | A kind of single image to the fog method based on deep learning | |
| CN108961308A (en) | A kind of residual error depth characteristic method for tracking target of drift detection | |
| Lu et al. | Blind image quality assessment based on the multiscale and dual‐domains features fusion | |
| CN109241814A (en) | Pedestrian detection method based on YOLO neural network | |
| CN114419413A (en) | Method for constructing sensing field self-adaptive transformer substation insulator defect detection neural network | |
| CN109871845A (en) | Certificate image extracting method and terminal device | |
| CN110288574A (en) | A kind of adjuvant Ultrasonographic Diagnosis hepatoncus system and method | |
| CN115965864A (en) | Lightweight attention mechanism network for crop disease identification | |
| CN116597224A (en) | Potato defect detection method based on improved YOLO V8 network model | |
| CN113689382B (en) | Tumor postoperative survival prediction method and system based on medical images and pathological images | |
| WO2024032010A1 (en) | Transfer learning strategy-based real-time few-shot object detection method | |
| CN111612739A (en) | Deep learning-based cerebral infarction classification method | |
| Fan et al. | Cycle GAN-MF: A Cycle-consistent Generative Adversarial Network Based on Multifeature Fusion for Pedestrian Re-recognition: Cycle GAN-MF |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170524 |