CN109325547A - Non-motor vehicle image multi-tag classification method, system, equipment and storage medium - Google Patents
Non-motor vehicle image multi-tag classification method, system, equipment and storage medium Download PDFInfo
- Publication number
- CN109325547A CN109325547A CN201811240000.3A CN201811240000A CN109325547A CN 109325547 A CN109325547 A CN 109325547A CN 201811240000 A CN201811240000 A CN 201811240000A CN 109325547 A CN109325547 A CN 109325547A
- Authority
- CN
- China
- Prior art keywords
- motor vehicle
- network model
- sorter network
- vehicle image
- training
- 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
- 238000003860 storage Methods 0.000 title claims abstract description 27
- 238000000605 extraction Methods 0.000 claims abstract description 42
- 239000000284 extract Substances 0.000 claims abstract description 13
- 238000004364 calculation method Methods 0.000 claims abstract description 9
- 230000003252 repetitive Effects 0.000 claims description 16
- 230000000875 corresponding Effects 0.000 claims description 13
- 238000000034 method Methods 0.000 description 13
- 239000000203 mixture Substances 0.000 description 9
- 230000004913 activation Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 5
- 210000002569 neurons Anatomy 0.000 description 5
- 230000001537 neural Effects 0.000 description 4
- 238000009826 distribution Methods 0.000 description 3
- 238000003475 lamination Methods 0.000 description 3
- 230000003287 optical Effects 0.000 description 3
- 235000013399 edible fruits Nutrition 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000005291 magnetic Effects 0.000 description 2
- 238000004321 preservation Methods 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 229910002056 binary alloy Inorganic materials 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 210000004027 cells Anatomy 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000002596 correlated Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 239000003365 glass fiber Substances 0.000 description 1
- 210000003702 immature single positive T cell Anatomy 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006011 modification reaction Methods 0.000 description 1
- 230000000051 modifying Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000002093 peripheral Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000000644 propagated Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 210000000352 storage cell Anatomy 0.000 description 1
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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
Abstract
The present invention provides a kind of non-motor vehicle image multi-tag classification method, system, equipment and storage mediums, the label of the non-motor vehicle image includes the classification results of multiple attributes, described method includes following steps: the non-motor vehicle image of test is inputted in trained sorter network model, the sorter network model include feature extraction layer and with the attribute multiple taxons correspondingly；The feature extraction layer of the sorter network model extracts the feature in test image；Multiple taxons of the sorter network model are respectively according to the classification results of each attribute of the feature calculation of extraction；The classification results of each attribute are merged, the label of the non-motor vehicle image as test.Present invention employs a sorter network models, and non-motor vehicle multiple attributive classification can be thus achieved, and training is convenient, and nicety of grading is high.
Description
Technical field
The present invention relates to technical field of image processing more particularly to a kind of non-motor vehicle image multi-tag classification method, it is
System, equipment and storage medium.
Background technique
Representative method of the convolutional neural networks as deep learning, can learn image characteristics extraction, in major part automatically
Computer Vision Task in have good effect.But in the application in certain fields, such as non-motor vehicle image domains, often
Need to obtain the output of class label under multiple attributes of target image.Multiple polytypic networks can be used in Normal practice
Complete the image recognition of multiple class labels, but this method will lead to algorithm time-consuming with the scale of categorical attribute in application and
Linearly increasing, efficiency is very low.
Summary of the invention
For the problems of the prior art, the purpose of the present invention is to provide a kind of non-motor vehicle image multi-tag classification sides
Non-motor vehicle multiple attributive classification can be thus achieved using a sorter network model in method, system, equipment and storage medium, training
Convenient, nicety of grading is high.
The embodiment of the present invention provides a kind of non-motor vehicle image multi-tag classification method, the label of the non-motor vehicle image
Classification results including multiple attributes, described method includes following steps:
The non-motor vehicle image of test is inputted in trained sorter network model, the sorter network model includes spy
Levy extract layer and with the attribute multiple taxons correspondingly；
The feature extraction layer of the sorter network model extracts the feature in the non-motor vehicle image of test；
Multiple taxons of the sorter network model are respectively according to the classification knot of each attribute of the feature calculation of extraction
Fruit；
The classification results of each attribute are merged, the label of the non-motor vehicle image as test.
Optionally, multiple taxons of the sorter network model are respectively according to each attribute of the feature calculation of extraction
Classification results include the following steps:
The non-motor vehicle image that multiple taxons of the sorter network model calculate separately the test belong to pair
The probability of each classification in the attribute answered, classification results of the maximum classification of select probability as corresponding attribute.
Optionally, the feature extraction layer includes an at least convolutional layer and at least a pond layer, the taxon are
Softmax layers.
Optionally, it is additionally provided with the first full articulamentum between the feature extraction layer and taxon and multiple branches connect entirely
Layer is connect, the multiple full articulamentum of branch and the taxon correspond, and the output of the feature extraction layer passes through described
Be connected to the full articulamentum of the multiple branch after first full articulamentum, the output of each full articulamentum of branch be connected to pair
The taxon answered.
Optionally, the output of the described first full articulamentum is by a dropout layers of one second full articulamentum of input, and described the
Two full articulamentums are connected to the full articulamentum of the multiple branch by a dropout layers.
Optionally, the sorter network model is trained using following steps:
Building includes the sorter network model of feature extraction layer and multiple taxons, the taxon and non-motor vehicle
The attribute of image corresponds；
Obtain training set, the training set include training image data and with number of tags corresponding to each training image
According to the label data includes the classification of image path and training image in each attribute；
The training set is inputted into the sorter network model and is iterated training, the penalty values of each taxon are added
Loss of the power summation as sorter network, repetitive exercise to model are restrained；
Save the sorter network model that training is completed.
Optionally, after the building includes feature extraction layer and the sorter network model of multiple taxons, further include
Following steps:
Trained weight file on ImageNet public data collection is obtained to propose the feature of the sorter network model of building
Layer and the first full articulamentum is taken to be initialized；
The full articulamentum of multiple branches and multiple taxons to the sorter network model of building carry out random initializtion.
Optionally, described that the training set input sorter network model is iterated training, include the following steps:
Batch size, initial learning rate and the maximum number of iterations of the setting training sorter network model；
Using sorter network model described in the training set repetitive exercise, after every repetitive exercise i times learning rate multiplied by k value,
Learning rate as successive iterations training, wherein i is the cycle times of default regularized learning algorithm rate, and k is the adjustment of preset learning rate
Coefficient, and k < 1；
After training reaches maximum number of iterations, judge whether the penalty values of the sorter network model are less than preset threshold；
If it is, repetitive exercise is completed；
Otherwise, continue using sorter network model described in the training set repetitive exercise, until the sorter network model
Penalty values are less than preset threshold.
The embodiment of the present invention also provides a kind of non-motor vehicle image multi-tag categorizing system, applied to the non-motor vehicle
Image multi-tag classification method, the system comprises:
Image input module, the non-motor vehicle image for that will test inputs in trained sorter network model, described
Sorter network model include feature extraction layer and with the attribute multiple taxons correspondingly；
Characteristic extracting module extracts the non-motor vehicle figure of test for the feature extraction layer using the sorter network model
Feature as in；
Image classification module, for multiple taxons using the sorter network model respectively according to the feature of extraction
Calculate the classification results of each attribute；
As a result output module, for the classification results of each attribute to be merged, the non-motor vehicle image as the test
Label.
The embodiment of the present invention also provides a kind of non-motor vehicle image multi-tag sorting device, comprising:
Processor；
Memory, wherein being stored with the executable instruction of the processor；
Wherein, the processor is configured to more to execute the non-motor vehicle image via the executable instruction is executed
The step of labeling method.
The embodiment of the present invention also provides a kind of computer readable storage medium, and for storing program, described program is performed
Described in Shi Shixian the step of non-motor vehicle image multi-tag classification method.
Non-motor vehicle image multi-tag classification method, system, equipment and storage medium provided by the present invention have following
Advantage:
The present invention extracts feature by deep learning and multiple taxons combine, using a sorter network model
To realize non-motor vehicle multiple attributive classification, training is convenient, and nicety of grading is high, to solve multiple using multiple in the prior art
The problem of the step of image classification model is cumbersome, inefficiency.
Detailed description of the invention
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon.
Fig. 1 is the flow chart of the non-motor vehicle image multi-tag classification method of one embodiment of the invention；
Fig. 2 is the dimension style exemplary diagram of the non-motor vehicle image of the training set of one embodiment of the invention；
Fig. 3 is the structural schematic diagram of the sorter network model of one embodiment of the invention；
Fig. 4 is the structural schematic diagram of the non-motor vehicle image multi-tag categorizing system of one embodiment of the invention；
Fig. 5 is the structural schematic diagram of the non-motor vehicle image multi-tag sorting device of one embodiment of the invention；
Fig. 6 is the structural schematic diagram of the computer storage medium of one embodiment of the invention.
Specific embodiment
Example embodiment is described more fully with reference to the drawings.However, example embodiment can be with a variety of shapes
Formula is implemented, and is not understood as limited to embodiment set forth herein；On the contrary, thesing embodiments are provided so that the present invention will
Fully and completely, and by the design of example embodiment comprehensively it is communicated to those skilled in the art.It is identical attached in figure
Icon note indicates same or similar structure, thus will omit repetition thereof.
As shown in Figure 1, the embodiment of the present invention provides a kind of non-motor vehicle image multi-tag classification method, the non-motor vehicle
The label of image includes the classification results of multiple attributes, and described method includes following steps:
S100: the non-motor vehicle image of test is inputted in trained sorter network model, the sorter network model
Including feature extraction layer and with the attribute multiple taxons correspondingly；
S200: the feature extraction layer of the sorter network model extracts the feature in the non-motor vehicle image of test；
S300: multiple taxons of the sorter network model are respectively according to point of each attribute of the feature calculation of extraction
Class result；
S400: the classification results of each attribute are merged, the label of the non-motor vehicle image as test.
Each attribute can be the preset attribute that mutex relation is not present to each other, such as multiple attributes may include moving
Power type is applicable in gender, vehicular applications, whether has parasols etc., accordingly, in power type attribute in the following, non-motor vehicle
Electric vehicle, motorcycle, bicycle etc. can be divided into, applicable gender can be divided into male's vehicle, Nv Xingche, and vehicular applications can be divided into
Take out vehicle, express delivery vehicle, inoperative vehicle etc., if having parasols that can be divided into has parasols and without parasols, one non-maneuver
The label of vehicle can be the combination of the classification results of multiple attributes, such as the label of a non-motor vehicle can be electric vehicle, female
Property vehicle, has parasols at inoperative vehicle.
Therefore, the present invention extracts feature by deep learning and multiple taxons combine, using a sorter network mould
Non-motor vehicle multiple attributive classification can be thus achieved in type, and training is convenient, and nicety of grading is high.
In this embodiment, multiple taxons of the sorter network model are each according to the feature calculation of extraction respectively
The classification results of attribute, include the following steps:
The non-motor vehicle image that multiple taxons of the sorter network model calculate separately the test belong to pair
The probability of each classification in the attribute answered, classification results of the maximum classification of select probability as corresponding attribute.Wherein, divide
Class unit can use softmax layer, it is softmax layers trained in the feature vector extracted by feature extraction layer of input, it is defeated
Result out is the vector of a T*1, and the value of T corresponds to the other number of attribute lower class, and each numerical tabular in the vector of T*1
What is shown is the probability that feature belongs to the next classification of the attribute, classification knot of the maximum classification of select probability as the taxon
Fruit, the classification of the selection is the classification that the non-motor vehicle image tested most likely belongs under the attribute, so as to accurate
The highest classification results of non-motor vehicle image accuracy tested.
In this embodiment, the feature extraction layer includes an at least convolutional layer and an at least pond layer, the grouping sheet
Member is softmax layers.It is additionally provided with the first full articulamentum between the feature extraction layer and taxon and multiple branches connect entirely
Layer is connect, the multiple full articulamentum of branch and the taxon correspond, and the output of the feature extraction layer passes through described
Be connected to the full articulamentum of the multiple branch after first full articulamentum, the output of each full articulamentum of branch be connected to pair
The taxon answered.Further, the output of the described first full articulamentum can be connected entirely by a dropout layers of input one second
Layer is connect, the second full articulamentum is connected to the full articulamentum of the multiple branch by a dropout layers.In convolutional neural networks
Every layer of convolutional layer is made of several convolution units, and the parameter of each convolution unit is to optimize to obtain by back-propagation algorithm
's.The purpose of convolution algorithm is to extract the different characteristic of input, and first layer convolutional layer may can only extract some rudimentary features
Such as edge, lines and angle level, the network of more layers can from low-level features the more complicated feature of iterative extraction.Pond layer
Sampling layer is made equally to be made of multiple characteristic faces after convolutional layer, corresponding one layer thereon of its each characteristic face
A characteristic face, the number of characteristic face will not be changed.Pond layer is intended to the resolution ratio by reducing characteristic face to be had
The feature of space-invariance.Pond layer plays the role of second extraction feature, its each neuron carries out local acceptance region
Pondization operation.Common pond method has maximum pondization to take the maximum point of local acceptance region intermediate value, mean value pondization i.e. to part
All values in acceptance region average, random pool etc., this example is mainly using maximum pond method.
Therefore, the reasonable cooperation of convolutional layer and pond layer can preferably extract the feature of the non-motor vehicle image of test,
To further increase the accuracy that taxon is classified.It will be further in the specific example shown in figure 2 and figure 3 below
Introduce each layer of concrete application.
In this embodiment, the sorter network model is trained using following steps:
Building includes the sorter network model of feature extraction layer and multiple taxons, the taxon and non-motor vehicle
The attribute of image corresponds；
Obtain training set, the training set include training image data and with number of tags corresponding to each training image
According to the label data includes the classification of image path and training image in each attribute；Training image data can be reality
What is be now collected has been carried out the non-motor vehicle image for classifying and knowing classification results, is each according to known classification results
The non-motor vehicle image of a training increases label, and trained non-motor vehicle image and label are added in training set together, with
Sorter network model is trained；
The training set is inputted into the sorter network model and is iterated training, the penalty values of each taxon are added
Loss of the power summation as sorter network, repetitive exercise to model are restrained；The damage of each taxon is allowed for due to losing
Mistake value, in an iterative process, the continuous reduction of total losses value, also just optimization improves the corresponding grouping sheet of each attribute simultaneously
The recognition accuracy of member；
The sorter network model that training is completed is saved, the sorter network model after convergence can be used for test of the invention
Non-motor vehicle image Classification and Identification, and can choose the accurate non-motor vehicle image of identification in use and be newly added to instruction
Practice and concentrates, training set of enriching constantly, and periodically re -training sorter network model, continue to optimize the identification of sorter network model
Effect, the accuracy rate of model identification to be continuously improved in use.
In this embodiment, after the sorter network model of the building including feature extraction layer and multiple taxons,
Further include following steps:
Trained weight file on ImageNet public data collection is obtained to propose the feature of the sorter network model of building
Layer and the first full articulamentum is taken to be initialized；Since ImageNet has had the convolutional layer of comparative maturity, pond layer and complete
It directly can be used to initialize sorter network model of the invention by the weight file of articulamentum, the embodiment, can
To greatly save the time of sorter network model training；
The full articulamentum of multiple branches and multiple taxons to the sorter network model of building carry out random initializtion, with
Machine initialization can initialize weight using normal distribution, however, the present invention is not limited thereto.
In this embodiment, described that the training set input sorter network model is iterated training, including such as
Lower step:
Batch size, initial learning rate and the maximum number of iterations of the setting training sorter network model；
Using sorter network model described in the training set repetitive exercise, after every repetitive exercise i times learning rate multiplied by k value,
Learning rate as successive iterations training, wherein i is the cycle times of default regularized learning algorithm rate, and k is the adjustment of preset learning rate
Coefficient, and k < 1；
During training pattern, in order to which the training speed of balance model has selected relatively suitable study with loss
Rate, but the loss of training set may decline and just no longer decline to a certain extent, at this time can be by suitably reducing learning rate
One step reduces loss, but the decline of learning rate can extend the training required time.Therefore, which uses learning rate gradually
The method of decaying finds new balance between training time and reduction loss, and appropriate by setting maximum number of iterations
The controlled training time；
After training reaches maximum number of iterations, judge whether the penalty values of the sorter network model are less than preset threshold；
If it is, illustrating that sorter network model has reached convergence, repetitive exercise is completed；
Otherwise, continue using sorter network model described in the training set repetitive exercise, until the sorter network model
Penalty values are less than preset threshold, that is, train until the convergence of sorter network model.
Below with reference to Fig. 2 and Fig. 3, non-motor vehicle image multi-tag of the invention is further described with a specific example
Classification method.In the specific example, the non-motor vehicle image multi-tag classification includes the following steps:
Step 1: non-motor vehicle multi-tag data set arranges, and step 1 is to correspond to the instruction of above-mentioned sorter network model
The step of training set is obtained during practicing；The process of specific steps one is as follows:
Step 1.1: obtaining largely trained non-motor vehicle image from reality scene, be picture number；
Step 1.2: the markup information of image being designed, is divided by multiple attributes, respectively with one-bit digital
Position is marked to indicate, every group of attribute includes a variety of attributes, is indicated respectively with the number of 1~N.Annotation formatting are as follows: [image path
Diameter] [the attribute class code of feature 1] [the attribute class code of feature 2] ..., such as include non-motor vehicle image shown in Fig. 2
Five attribute markup informations mask method；
Step 1.3: all non-motor vehicle images being labeled with the above mask method, to label information and image
Information is organized into corresponding data set respectively.Data set is LMDB format.
Step 2: non-motor vehicle multi-tag sorter network model construction, the sorter network model of building is as shown in figure 3, step
Had in rapid two for the building mode of five convolutional layers, three pond layers and the full articulamentum that are used in above-described embodiment
The building process of body introduction, concrete model is as follows:
Step 2.1: network is by input layer, five convolutional layers, three pond layers, 2+N articulamentums, N number of full when training
Softmax layers of composition, wherein the value of N is identified non-motor vehicle attribute number.Step 2.2:data input layer is followed by the first volume
Lamination and Relu activation primitive, and it is followed by first standardization processing layer in Relu activation primitive, connect first pond below herein
Change layer maximum value pond；
Step 2.3: pond is followed by the second convolutional layer, adds Relu activation primitive after the second convolution, and be followed by activation primitive
Second batch standardization processing layer, standardization processing are followed by the second pond layer, and the second pond layer uses maximum value pond；
In deep neural network, usually using one kind cry correct linear unit (Rectified linear unit,
Relu) as the activation primitive of neuron.By ReLU realize it is sparse after model can preferably excavate correlated characteristic, be fitted
Training data.
Step 2.4: continuously connecing three convolutional layers after the second pond layer, put in order and activate letter for third convolutional layer, Relu
Number, Volume Four lamination, Relu activation primitive, the 5th convolutional layer, Relu activation primitive；
Step 2.5: being followed by third pond layer in previous step, pond is followed by the first full articulamentum, and in the first full connection
Layer is followed by Relu activation primitive and dropout layers, then connects the second full articulamentum, the first full articulamentum and the second full articulamentum are set
It sets identical；Each node of full articulamentum is connected with upper one layer of all nodes, comprehensive for feature that front is extracted
Altogether.
Step 2.6: being followed by the full articulamentum of N number of branch in the second full articulamentum, the respectively full articulamentum 1 of branch, branch is complete
The full articulamentum n of articulamentum 2 ... branch, connect after each full articulamentum of branch again one it is softmax layers corresponding, as corresponding
Taxon, attribute classification output of the output of taxon as final every class characteristics of image, respectively taxon 1,
Taxon 2 ... taxon n.
In the present invention, after multiple convolutional layers and pond layer, 2 full articulamentums is connected to and N number of branch connects entirely
Layer.Each neuron in full articulamentum is connect entirely with all neurons of its preceding layer.Full articulamentum can integrate volume
With the local message of class discrimination in lamination or pond layer.In order to promote the network performance of convolutional neural networks, Quan Lian
The excitation function of each neuron of layer is connect using ReLU function.The output valve of each full articulamentum of branch is delivered to one
Softmax layers are classified.Softmax layers of algorithm can be understood as normalizing, and such as current picture classification has x kind, that process
Softmax layers of output is exactly the vector of x dimension.First value in vector is exactly the probability that current image belongs to the first kind
It is worth, second value in vector is exactly that the sum of vector that current image belongs to that this x of the probability value ... of the second class is tieed up is 1.
Step 3: the training of non-motor vehicle multi-tag sorter network, step 3 correspond to the instruction of above-mentioned sorter network model
The training set is inputted into the sorter network model in white silk and is iterated trained step, by each grouping sheet in training process
Loss of the penalty values weighted sum of member as sorter network, repetitive exercise to model are restrained, and the process of specific steps three is as follows:
Step 3.1: according to the LMDB data set arranged in step 1, first calculating the mean value file of training dataset, save
For the format of .binaryproto file, and refer to the position for determining binary system mean value file in training network；
Step 3.2: using the training method of finetune, using the trained weight on ImageNet public data collection
File initializes current network carry out portion hierarchical weight, and other layers are carried out with the mode of random initializtion, institute as above
It states, in this example, using weight file trained on ImageNet public data collection to convolutional layer, pond layer, first
Full articulamentum and the second full articulamentum initialization, and random initializtion is carried out to the full articulamentum of branch and taxon, it is random first
Beginningization can initialize weight using normal distribution, however, the present invention is not limited thereto；
Step 3.3: in batch size, initial learning rate and the maximum number of iterations of the setting training sorter network model
When, setting crowd size batch_size is 32, and initial learning rate is 0.001, and maximum number of iterations is 200000 times, using step
Mode carry out learning rate modification in the training process, learning rate is multiplied by 0.9 after every iteration 1000 times, using stochastic gradient descent
Algorithm training data sets 10000 preservation primary network models of every iteration；Stochastic gradient descent algorithm can be accelerated to update each
A layer of weight file and biased data, referring in each iterative process at random here, sample will be upset at random, upset
Parameter caused by can effectively reducing between sample updates cancellation problem.In most basic stochastic gradient descent algorithm, parameter
Each step is updated by subtracting its gradient, and large-scale machine learning task, stochastic gradient descent algorithm are showed
Performance is very considerable.
When training, training sample and label are inputted into the network of initialization, calculate the penalty values of each softmax layers of input
Weighted sum is as final loss, and each softmax layers of weight is 1/N, but invention is not limited thereto, specific weight value
Distribution, which can according to need, to be configured.By two steps of continuous propagated forward and backpropagation, repetition training makes
Losing in training process constantly reduces, until reaching maximum the number of iterations；
After training is completed every time, using the network model of preservation as the pre-training model of next iteration training, avoid
Occur the loss of model data in the training process, continue to train, after training reaches maximum the number of iterations, judges that loss is
It is no to reach preset threshold hereinafter, if it is, end training, otherwise continues training until loss reaches preset threshold hereinafter, instruction
Practicing convergence then terminates to train.
Step 4: non-motor vehicle multi-tag image classification, step 1 and step 3 are all non-motor vehicle multi-tag images
The preparation process of sorter network model, and step 4 corresponds to above-mentioned steps S100~step S400, i.e., using trained
The step of sorter network model classifies to image, specifically, the process of step 4 are as follows:
Step 4.1: the test data pre-processed being sent into trained sorter network model, extracts non-motor vehicle image
Feature corresponds to step S100 and S200；
Step 4.2: the non-motor vehicle characteristics of image extracted being sent into softmax layers, N number of feature is exported and belongs to specified genus
Property in each classification probability, take the classification of maximum probability in each attribute as the classification results of the attribute, by N group identification
Classification results merge into final output list of categories, that is, correspond to step S300 and S400.
Due to the structure for the sorter network model that the present invention uses, non-motor vehicle image can be disposably obtained in multiple categories
Property under classification results, and operated without individually constructing disaggregated model for each attribute, therefore by above-mentioned step, i.e.,
The quick multi-tag classification of non-motor vehicle image can be achieved.
As shown in figure 4, the embodiment of the present invention also provides a kind of non-motor vehicle image multi-tag categorizing system, it is applied to described
Non-motor vehicle image multi-tag classification method, the system comprises:
Image input module M100, the non-motor vehicle image for that will test input in trained sorter network model,
The sorter network model include feature extraction layer and with the attribute multiple taxons correspondingly；
Characteristic extracting module M200, for being extracted in test image using the feature extraction layer of the sorter network model
Feature；
Image classification module M300, for multiple taxons using the sorter network model respectively according to extraction
The classification results of each attribute of feature calculation；
As a result output module M400, for the classification results of each attribute to be merged, the non-motor vehicle as the test
The label of image.
Therefore, feature extraction layer and image classification module that the present invention passes through deep learning in characteristic extracting module M200
Multiple taxons combine in M300, and non-motor vehicle multiple attributive classification can be thus achieved using a sorter network model, training
Convenient, nicety of grading is high.
The embodiment of the present invention also provides a kind of non-motor vehicle image multi-tag sorting device, including processor；Memory,
In be stored with the executable instruction of the processor；Wherein, the processor is configured to next via the executable instruction is executed
The step of executing the non-motor vehicle image multi-tag classification method.
Person of ordinary skill in the field it is understood that various aspects of the invention can be implemented as system, method or
Program product.Therefore, various aspects of the invention can be embodied in the following forms, it may be assumed that complete hardware embodiment, complete
The embodiment combined in terms of full Software Implementation (including firmware, microcode etc.) or hardware and software, can unite here
Referred to as circuit, " module " or " system ".
The electronic equipment 600 of this embodiment according to the present invention is described referring to Fig. 5.The electronics that Fig. 5 is shown
Equipment 600 is only an example, should not function to the embodiment of the present invention and use scope bring any restrictions.
As shown in figure 5, electronic equipment 600 is showed in the form of universal computing device.The component of electronic equipment 600 can wrap
It includes but is not limited to: at least one processing unit 610, at least one storage unit 620, (including the storage of the different system components of connection
Unit 620 and processing unit 610) bus 630, display unit 640 etc..
Wherein, the storage unit is stored with program code, and said program code can be held by the processing unit 610
Row, so that the processing unit 610 executes described in this specification above-mentioned electronic prescription circulation processing method part according to this
The step of inventing various illustrative embodiments.For example, the processing unit 610 can execute step as shown in fig. 1.
The storage unit 620 may include the readable medium of volatile memory cell form, such as random access memory
Unit (RAM) 6201 and/or cache memory unit 6202 can further include read-only memory unit (ROM) 6203.
The storage unit 620 can also include program/practical work with one group of (at least one) program module 6205
Tool 6204, such program module 6205 includes but is not limited to: operating system, one or more application program, other programs
It may include the realization of network environment in module and program data, each of these examples or certain combination.
Bus 630 can be to indicate one of a few class bus structures or a variety of, including storage unit bus or storage
Cell controller, peripheral bus, graphics acceleration port, processing unit use any bus structures in a variety of bus structures
Local bus.
Electronic equipment 600 can also be with one or more external equipments 700 (such as keyboard, sensing equipment, bluetooth equipment
Deng) communication, can also be enabled a user to one or more equipment interact with the electronic equipment 600 communicate, and/or with make
Any equipment (such as the router, modulation /demodulation that the electronic equipment 600 can be communicated with one or more of the other calculating equipment
Device etc.) communication.This communication can be carried out by input/output (I/O) interface 650.Also, electronic equipment 600 can be with
By network adapter 660 and one or more network (such as local area network (LAN), wide area network (WAN) and/or public network,
Such as internet) communication.Network adapter 660 can be communicated by bus 630 with other modules of electronic equipment 600.It should
Understand, although not shown in the drawings, other hardware and/or software module can be used in conjunction with electronic equipment 600, including but unlimited
In: microcode, device driver, redundant processing unit, external disk drive array, RAID system, tape drive and number
According to backup storage system etc..
The embodiment of the present invention also provides a kind of computer readable storage medium, and for storing program, described program is performed
Described in Shi Shixian the step of non-motor vehicle image multi-tag classification method.In some possible embodiments, of the invention
Various aspects are also implemented as a kind of form of program product comprising program code, when described program product is set in terminal
When standby upper operation, said program code is for making the terminal device execute the above-mentioned electronic prescription circulation processing method of this specification
Described in part according to the present invention various illustrative embodiments the step of.
Refering to what is shown in Fig. 6, describing the program product for realizing the above method of embodiment according to the present invention
800, can using portable compact disc read only memory (CD-ROM) and including program code, and can in terminal device,
Such as it is run on PC.However, program product of the invention is without being limited thereto, in this document, readable storage medium storing program for executing can be with
To be any include or the tangible medium of storage program, the program can be commanded execution system, device or device use or
It is in connection.
Described program product can be using any combination of one or more readable mediums.Readable medium can be readable letter
Number medium or readable storage medium storing program for executing.Readable storage medium storing program for executing for example can be but be not limited to electricity, magnetic, optical, electromagnetic, infrared ray or
System, device or the device of semiconductor, or any above combination.The more specific example of readable storage medium storing program for executing is (non exhaustive
List) include: electrical connection with one or more conducting wires, portable disc, hard disk, random access memory (RAM), read-only
Memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable compact disc read only memory
(CD-ROM), light storage device, magnetic memory device or above-mentioned any appropriate combination.
The computer readable storage medium may include in a base band or the data as the propagation of carrier wave a part are believed
Number, wherein carrying readable program code.The data-signal of this propagation can take various forms, including but not limited to electromagnetism
Signal, optical signal or above-mentioned any appropriate combination.Readable storage medium storing program for executing can also be any other than readable storage medium storing program for executing
Readable medium, the readable medium can send, propagate or transmit for by instruction execution system, device or device use or
Person's program in connection.The program code for including on readable storage medium storing program for executing can transmit with any suitable medium, packet
Include but be not limited to wireless, wired, optical cable, RF etc. or above-mentioned any appropriate combination.
The program for executing operation of the present invention can be write with any combination of one or more programming languages
Code, described program design language include object oriented program language-Java, C++ etc., further include conventional
Procedural programming language-such as " C " language or similar programming language.Program code can be fully in user
It calculates and executes in equipment, partly executes on a user device, being executed as an independent software package, partially in user's calculating
Upper side point is executed on a remote computing or is executed in remote computing device or server completely.It is being related to far
Journey calculates in the situation of equipment, and remote computing device can pass through the network of any kind, including local area network (LAN) or wide area network
(WAN), it is connected to user calculating equipment, or, it may be connected to external computing device (such as utilize ISP
To be connected by internet).
In conclusion compared with prior art, non-motor vehicle image multi-tag classification method provided by the present invention is
System, equipment and storage medium have the advantage that
The present invention extracts feature by deep learning and multiple taxons combine, using a sorter network model
To realize non-motor vehicle multiple attributive classification, training is convenient, and nicety of grading is high, to solve multiple using multiple in the prior art
The problem of the step of image classification model is cumbersome, inefficiency.
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that
Specific implementation of the invention is only limited to these instructions.For those of ordinary skill in the art to which the present invention belongs, exist
Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to of the invention
Protection scope.
Claims (11)
1. a kind of non-motor vehicle image multi-tag classification method, which is characterized in that the label of the non-motor vehicle image includes more
The classification results of a attribute, described method includes following steps:
The non-motor vehicle image of test is inputted in trained sorter network model, the sorter network model includes that feature mentions
Take layer and with the attribute multiple taxons correspondingly；
The feature extraction layer of the sorter network model extracts the feature in the non-motor vehicle image of test；
Multiple taxons of the sorter network model are respectively according to the classification results of each attribute of the feature calculation of extraction；
The classification results of each attribute are merged, the label of the non-motor vehicle image as test.
2. non-motor vehicle image multi-tag classification method according to claim 1, which is characterized in that the sorter network mould
Multiple taxons of type include the following steps: respectively according to the classification results of each attribute of the feature calculation of extraction
The non-motor vehicle image that multiple taxons of the sorter network model calculate separately the test belongs to corresponding
The probability of each classification in attribute, classification results of the maximum classification of select probability as corresponding attribute.
3. non-motor vehicle image multi-tag classification method according to claim 1, which is characterized in that the feature extraction layer
Including an at least convolutional layer and an at least pond layer, the taxon is softmax layers.
4. non-motor vehicle image multi-tag classification method according to claim 1, which is characterized in that the feature extraction layer
It is additionally provided with the full articulamentum of the first full articulamentum and multiple branches between taxon, the full articulamentum of the multiple branch and institute
Taxon one-to-one correspondence is stated, the output of the feature extraction layer is by being connected to the multiple point after the described first full articulamentum
The full articulamentum of branch, the output of each full articulamentum of branch are connected to corresponding taxon.
5. non-motor vehicle image multi-tag classification method according to claim 4, which is characterized in that the described first full connection
The output of layer passes through a dropout layers of connection by a dropout layers of one second full articulamentum of input, the second full articulamentum
To the full articulamentum of the multiple branch.
6. non-motor vehicle image multi-tag classification method according to claim 4, which is characterized in that the sorter network mould
Type is trained using following steps:
Building includes the sorter network model of feature extraction layer and multiple taxons, the taxon and non-motor vehicle image
Attribute correspond；
Obtain training set, the training set include training image data and with label data corresponding to each training image, institute
Stating label data includes the classification of image path and training image in each attribute；
The training set is inputted into the sorter network model and is iterated training, the penalty values weighting of each taxon is asked
With the loss as sorter network, repetitive exercise to model is restrained；
Save the sorter network model that training is completed.
7. non-motor vehicle image multi-tag classification method according to claim 6, which is characterized in that the building includes spy
Further include following steps after levying extract layer and the sorter network model of multiple taxons:
Feature extraction layer of the trained weight file to the sorter network model of building on acquisition ImageNet public data collection
It is initialized with the first full articulamentum；
The full articulamentum of multiple branches and multiple taxons to the sorter network model of building carry out random initializtion.
8. non-motor vehicle image multi-tag classification method according to claim 7, which is characterized in that described by the training
Collection inputs the sorter network model and is iterated training, includes the following steps:
Batch size, initial learning rate and the maximum number of iterations of the setting training sorter network model；
Using sorter network model described in the training set repetitive exercise, after every repetitive exercise i times learning rate multiplied by k value, as
The learning rate of successive iterations training, wherein i is the cycle times of default regularized learning algorithm rate, and k is preset learning rate adjustment system
Number, and k < 1；
After training reaches maximum number of iterations, judge whether the penalty values of the sorter network model are less than preset threshold；
If it is, repetitive exercise is completed；
Otherwise, continue using sorter network model described in the training set repetitive exercise, until the loss of the sorter network model
Value is less than preset threshold.
9. a kind of non-motor vehicle image multi-tag categorizing system, which is characterized in that be applied to any one of claims 1 to 8 institute
The non-motor vehicle image multi-tag classification method stated, the system comprises:
Image input module, the non-motor vehicle image for that will test input in trained sorter network model, the classification
Network model include feature extraction layer and with the attribute multiple taxons correspondingly；
Characteristic extracting module, for being extracted in the non-motor vehicle image of test using the feature extraction layer of the sorter network model
Feature；
Image classification module, for multiple taxons using the sorter network model respectively according to the feature calculation of extraction
The classification results of each attribute；
As a result output module, for the classification results of each attribute to be merged, the mark of the non-motor vehicle image as the test
Label.
10. a kind of non-motor vehicle image multi-tag sorting device characterized by comprising
Processor；
Memory, wherein being stored with the executable instruction of the processor；
Wherein, the processor is configured to come described in any one of perform claim requirement 1 to 8 via the execution executable instruction
Non-motor vehicle image multi-tag classification method the step of.
11. a kind of computer readable storage medium, for storing program, which is characterized in that described program is performed realization power
Benefit require any one of 1 to 8 described in non-motor vehicle image multi-tag classification method the step of.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811240000.3A CN109325547A (en) | 2018-10-23 | 2018-10-23 | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium |
| PCT/CN2019/111320 WO2020083073A1 (en) | 2018-10-23 | 2019-10-15 | Non-motorized vehicle image multi-label classification method, system, device and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811240000.3A CN109325547A (en) | 2018-10-23 | 2018-10-23 | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109325547A true CN109325547A (en) | 2019-02-12 |
Family
ID=65262678
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811240000.3A Pending CN109325547A (en) | 2018-10-23 | 2018-10-23 | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109325547A (en) |
| WO (1) | WO2020083073A1 (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020083073A1 (en) * | 2018-10-23 | 2020-04-30 | 苏州科达科技股份有限公司 | Non-motorized vehicle image multi-label classification method, system, device and storage medium |
| CN111368931A (en) * | 2020-03-09 | 2020-07-03 | 第四范式（北京）技术有限公司 | Method and device for training image classification model, computer device and storage medium |
| CN111737521A (en) * | 2020-08-04 | 2020-10-02 | 北京微播易科技股份有限公司 | Video classification method and device |
| CN112115880A (en) * | 2020-09-21 | 2020-12-22 | 成都数之联科技有限公司 | Ship pollution monitoring method, system, device and medium based on multi-label learning |
| CN112446439A (en) * | 2021-01-29 | 2021-03-05 | 魔视智能科技(上海)有限公司 | Inference method and system for deep learning model dynamic branch selection |
| CN113313079A (en) * | 2021-07-16 | 2021-08-27 | 深圳市安软科技股份有限公司 | Training method and system of vehicle attribute recognition model and related equipment |
| CN114429638A (en) * | 2022-04-06 | 2022-05-03 | 四川省大数据中心 | Construction drawing examination management system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105654066A (en) * | 2016-02-02 | 2016-06-08 | 北京格灵深瞳信息技术有限公司 | Vehicle identification method and device |
| WO2017158058A1 (en) * | 2016-03-15 | 2017-09-21 | Imra Europe Sas | Method for classification of unique/rare cases by reinforcement learning in neural networks |
| CN107330396A (en) * | 2017-06-28 | 2017-11-07 | 华中科技大学 | A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study |
| CN107886073A (en) * | 2017-11-10 | 2018-04-06 | 重庆邮电大学 | A kind of more attribute recognition approaches of fine granularity vehicle based on convolutional neural networks |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106203330A (en) * | 2016-07-08 | 2016-12-07 | 西安理工大学 | A kind of vehicle classification method based on convolutional neural networks |
| CN108256498A (en) * | 2018-02-01 | 2018-07-06 | 上海海事大学 | A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN |
| CN109325547A (en) * | 2018-10-23 | 2019-02-12 | 苏州科达科技股份有限公司 | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium |
-
2018
- 2018-10-23 CN CN201811240000.3A patent/CN109325547A/en active Pending
-
2019
- 2019-10-15 WO PCT/CN2019/111320 patent/WO2020083073A1/en active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105654066A (en) * | 2016-02-02 | 2016-06-08 | 北京格灵深瞳信息技术有限公司 | Vehicle identification method and device |
| WO2017158058A1 (en) * | 2016-03-15 | 2017-09-21 | Imra Europe Sas | Method for classification of unique/rare cases by reinforcement learning in neural networks |
| CN107330396A (en) * | 2017-06-28 | 2017-11-07 | 华中科技大学 | A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study |
| CN107886073A (en) * | 2017-11-10 | 2018-04-06 | 重庆邮电大学 | A kind of more attribute recognition approaches of fine granularity vehicle based on convolutional neural networks |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020083073A1 (en) * | 2018-10-23 | 2020-04-30 | 苏州科达科技股份有限公司 | Non-motorized vehicle image multi-label classification method, system, device and storage medium |
| CN111368931A (en) * | 2020-03-09 | 2020-07-03 | 第四范式（北京）技术有限公司 | Method and device for training image classification model, computer device and storage medium |
| CN111737521A (en) * | 2020-08-04 | 2020-10-02 | 北京微播易科技股份有限公司 | Video classification method and device |
| CN112115880A (en) * | 2020-09-21 | 2020-12-22 | 成都数之联科技有限公司 | Ship pollution monitoring method, system, device and medium based on multi-label learning |
| CN112446439A (en) * | 2021-01-29 | 2021-03-05 | 魔视智能科技(上海)有限公司 | Inference method and system for deep learning model dynamic branch selection |
| CN112446439B (en) * | 2021-01-29 | 2021-04-23 | 魔视智能科技(上海)有限公司 | Inference method and system for deep learning model dynamic branch selection |
| CN113313079A (en) * | 2021-07-16 | 2021-08-27 | 深圳市安软科技股份有限公司 | Training method and system of vehicle attribute recognition model and related equipment |
| CN114429638A (en) * | 2022-04-06 | 2022-05-03 | 四川省大数据中心 | Construction drawing examination management system |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020083073A1 (en) | 2020-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109325547A (en) | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium | |
| CN110334705B (en) | Language identification method of scene text image combining global and local information | |
| CN112434721A (en) | Image classification method, system, storage medium and terminal based on small sample learning | |
| CN111507378A (en) | Method and apparatus for training image processing model | |
| CN108648020A (en) | User behavior quantization method, system, equipment and storage medium | |
| CN110175248B (en) | Face image retrieval method and device based on deep learning and Hash coding | |
| CN109063719B (en) | Image classification method combining structure similarity and class information | |
| CN108073851B (en) | Grabbing gesture recognition method and device and electronic equipment | |
| CN109817276A (en) | A kind of secondary protein structure prediction method based on deep neural network | |
| CN110084221A (en) | A kind of serializing face critical point detection method of the tape relay supervision based on deep learning | |
| CN112668579A (en) | Weak supervision semantic segmentation method based on self-adaptive affinity and class distribution | |
| CN113326764A (en) | Method and device for training image recognition model and image recognition | |
| CN111666919A (en) | Object identification method and device, computer equipment and storage medium | |
| CN113298817A (en) | High-accuracy semantic segmentation method for remote sensing image | |
| CN112199536A (en) | Cross-modality-based rapid multi-label image classification method and system | |
| CN113822264A (en) | Text recognition method and device, computer equipment and storage medium | |
| CN112364974A (en) | Improved YOLOv3 algorithm based on activation function | |
| CN113807399A (en) | Neural network training method, neural network detection method and neural network detection device | |
| CN110111365B (en) | Training method and device based on deep learning and target tracking method and device | |
| CN110728186A (en) | Fire detection method based on multi-network fusion | |
| CN110110724A (en) | The text authentication code recognition methods of function drive capsule neural network is squeezed based on exponential type | |
| CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| CN113642431A (en) | Training method and device of target detection model, electronic equipment and storage medium | |
| CN111160219B (en) | Object integrity evaluation method and device, electronic equipment and storage medium | |
| CN112560948A (en) | Eye fundus map classification method and imaging method under data deviation |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190212 |
|
| RJ01 | Rejection of invention patent application after publication |