CN106096670B - Concatenated convolutional neural metwork training and image detecting method, apparatus and system - Google Patents
Concatenated convolutional neural metwork training and image detecting method, apparatus and system Download PDFInfo
- Publication number
- CN106096670B CN106096670B CN201610439342.2A CN201610439342A CN106096670B CN 106096670 B CN106096670 B CN 106096670B CN 201610439342 A CN201610439342 A CN 201610439342A CN 106096670 B CN106096670 B CN 106096670B
- Authority
- CN
- China
- Prior art keywords
- convolutional neural
- image data
- neural networks
- neural network
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Physiology (AREA)
- Databases & Information Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of concatenated convolutional neural metwork training and image detecting method, apparatus and system, wherein, the training method includes: that the image data of image to be learned at least regional area is processed into the image data of the different size of input area of N kind respectively, and N is the integer more than or equal to 2；Respectively using the image data of N kind input area as the input of convolutional neural networks at different levels in N grades of cascade convolutional neural networks, convolutional neural networks at different levels are trained；At least training result that convolutional neural networks at different levels are exported respectively is associated, and the training result after association is back to convolutional neural networks at different levels to adjust the parameter of neural networks at different levels.When training result is propagated to convolutional neural networks at different levels, the parameter of neural networks at different levels can be adjusted, concatenated convolutional neural network is enabled to reach the global optimization of neural network parameter in training.
Description
Technical field
The present invention relates to image real time transfer fields, and in particular to a kind of concatenated convolutional neural metwork training and image detection
Method, apparatus and system.
Background technique
Object detection is the picture for input, the position of certain all type objects is accurately detected out, in computer vision
It is occupied an important position in area of pattern recognition.
Traditional object detecting method based on convolutional neural networks, selects a series of positions, size on picture first
Then the region is directly inputted a convolutional neural networks, obtains classification results by different area to be tested.By suitably setting
The structure for counting convolutional neural networks, can allow computer directly to learn the hiding feature in picture out, avoid engineer spy
Sign, can be applied even more extensively in the detection of various classification objects.But since the calculating time of convolutional neural networks often compares
The calculating time of artificial design features is much longer, therefore the lance that meet visiting degree of testing the speed and detection effect can not get both when detecting
Shield.
For the object detecting method based on convolutional neural networks, commonly single-stage multi-layer convolutional neural networks at present.
Since the preferable single-stage convolutional neural networks structure of classifying quality is more complex, the time for detecting each region is longer, if selecting
Using comprehensive sliding window back-and-forth method for covering picture when selecting area to be tested, then it is slower to detect speed；If using for figure
The algorithms selection most probable of the certain characteristic Designs of piece includes a small amount of region of object, then detects speed and increase, but selecting
Regional stage may miss certain regions comprising object.For object detection problem, due to data mark very labor intensive,
Time, often there are many negative sample in data set, and positive sample is seldom.In training, since positive and negative sample distribution is unbalanced, training effect
Fruit is often poor.
Concatenated convolutional neural network can be made by using multistage multilayer neural network small-sized, that complexity is incremental
Speed is detected with improving while sliding window back-and-forth method, and the positive negative sample for providing different proportion to network not at the same level is instructed
Practice, alleviates two above-mentioned problems to a certain extent.But in traditional cascade network, neural network not at the same level is usual
Separately training, can only make every level-one neural network reach local optimum, the overall performance of multistage multilayer neural network is undesirable.
Summary of the invention
The embodiment of the present invention provides a kind of concatenated convolutional neural network and its training and detection scheme.
According in a first aspect, the embodiment of the invention provides a kind of concatenated convolutional neural network training methods, comprising:
The image data of image to be learned at least regional area is processed into the figure of the different size of input area of N kind respectively
As data, N is the integer more than or equal to 2；Respectively using the image data of N kind input area as N grades of cascade convolutional Neurals
The input of convolutional neural networks at different levels, is trained convolutional neural networks at different levels in network, wherein N grades of cascade convolution minds
It is corresponding with a kind of input area of N kind input area respectively through every level-one convolutional neural networks in network；By convolution minds at different levels
It is associated through at least training result that network exports respectively, and the training result after association is back to convolutional Neurals at different levels
Network is to adjust the parameters of neural networks at different levels.
Optionally, convolutional neural networks at different levels are trained, comprising: according to level 1 volume product neural network the last layer
Output, obtain level 1 volume product neural network feature vector；According to the output of n-th grade of convolutional neural networks the last layer and
The feature vector of (n-1)th grade of convolutional neural networks obtains the feature vector of n-th grade of convolutional neural networks, wherein and n is positive integer,
And 1 < n≤N.
Optionally, the last layer of level 1 volume product neural network is convolutional layer；It is at different levels in 2nd to N grade of convolutional neural networks
The last layer of convolutional neural networks is full articulamentum.
Optionally, convolutional neural networks at different levels are trained, comprising: at least seek convolutional neural networks at different levels respectively
The same level loss；At least training result that convolutional neural networks at different levels are exported respectively is associated, comprising: at least to volumes at different levels
The same level loss that product neural network exports respectively is weighted summation, obtains the global loss of N grades of cascade convolutional neural networks.
Optionally, the same level loss of convolutional neural networks at different levels is at least sought respectively, comprising: seeks convolution minds at different levels respectively
The same level loss function and recurrence bounding box true value loss through network；The same level that at least convolutional neural networks at different levels are exported respectively
Loss is weighted summation, obtains the global loss of N grades of cascade convolutional neural networks, comprising: to convolutional neural networks at different levels
The same level loss function and/or recurrence bounding box true value loss exported respectively is weighted summation, obtains N grades of cascade convolution minds
Global loss through network.
Optionally, the size of i-stage convolutional neural networks input area is less than j-th stage convolutional neural networks input area
Size, wherein i and j is positive integer, and 1≤i < j≤N.
The embodiment of the invention also provides a kind of concatenated convolutional neural metwork training devices, comprising:
Learning data acquiring unit, for the image data of image to be learned at least regional area to be processed into N kind respectively
The image data of different size of input area, the N are the integer more than or equal to 2；Training unit, being used for respectively will be described
Input of the image data of N kind input area as convolutional neural networks at different levels in N grades of cascade convolutional neural networks, at different levels
Convolutional neural networks are trained, wherein every level-one convolutional neural networks difference in the N grades of cascade convolutional neural networks
It is corresponding with a kind of input area of the N kind input area；Back propagation unit, for export convolutional neural networks at different levels respectively
At least a training result is associated, and the training result after association is back to convolutional neural networks at different levels to adjust minds at different levels
Parameter through network.
Optionally, training unit includes: primary vector unit, for according to level 1 volume product neural network the last layer
Output obtains the feature vector of the level 1 volume product neural network；Secondary vector unit, for according to n-th grade of convolutional Neural net
The output of network the last layer and the feature vector of (n-1)th grade of convolutional neural networks, obtain the spy of n-th grade of convolutional neural networks
Levy vector, wherein the n is positive integer, and 1 < n≤N.
Optionally, the training unit includes: that unit is sought in the same level loss, at least seeking convolutional Neurals at different levels respectively
The same level of network is lost；Unit is sought in overall situation loss, at least losing to the same level that convolutional neural networks at different levels export respectively
It is weighted summation, obtains the global loss of the N grades of cascade convolutional neural networks.
The embodiment of the invention also provides a kind of computer storage medium, store for realizing offer of the embodiment of the present invention
Any concatenated convolutional neural network training method computer-readable instruction.For example, described instruction includes: will be wait learn
The image data of habit image at least regional area is processed into the instruction of the image data of the different size of input area of N kind, N respectively
For the integer more than or equal to 2；Respectively using the image data of N kind input area as each in N grades of cascade convolutional neural networks
The input of grade convolutional neural networks, the instruction that convolutional neural networks at different levels are trained, wherein N grades of cascade convolutional Neurals
Every level-one convolutional neural networks in network are corresponding with a kind of input area of N kind input area respectively；By convolutional Neurals at different levels
At least training result that network exports respectively is associated, and the training result after association is back to convolutional Neural nets at different levels
Network is to adjust the instruction of the parameter of neural networks at different levels.
The embodiment of the invention also provides a kind of concatenated convolutional neural metwork training systems, comprising:
Image acquiring device, for obtaining the image data of image to be learned；Memory, for storing program；Processor,
The image data for receiving image to be learned realizes the operation in above-mentioned training method for executing program.
According to second aspect, the embodiment of the invention provides a kind of image detection sides based on concatenated convolutional neural network
Method, comprising:
Obtain the image data of image to be detected；It is established the image data of image to be detected as above-mentioned training method
The input of neural network model image to be detected is detected, obtain the testing result of image to be detected.
Optionally, after the image data for obtaining image to be detected, further includes: partition image data into multiple regions
Obtain the image data of each region；Using the image data of image to be detected as the input of neural network model to figure to be detected
As being detected, the testing result of image to be detected is obtained, comprising: successively regard the image data of each region as nerve respectively
The input of network model detects the image data of each region, obtains the testing result of each region.
Optionally, carrying out detection to image to be detected includes: the N grade concatenated convolutional neural network in neural network model
Classification score is carried out to image data at least one level neural network to calculate；By the classification score of at least one level convolutional neural networks
Calculated value be compared at least one predetermined score value, and determine whether image data includes target according to an at least comparison result
Object.
Optionally, after the image data for obtaining image to be detected, further includes: by the size compression of image data at symbol
Close the size of n-th grade of convolutional neural networks input area, wherein n is positive integer, and 1≤n≤N-1；By the figure of image to be detected
Input as data as neural network model detects image to be detected, obtains the testing result of image to be detected, packet
It includes: compressed image data is inputted into n-th grade of convolutional neural networks；By n-th grade of convolutional neural networks to compressed figure
It is calculated as data carry out the first classification score；If the first classification score is judged as image data less than the first predetermined score value
Not comprising target object.
Optionally, if the first classification score is greater than or equal to the first predetermined score value, further include: by the ruler of image data
It is very little to be compressed into the size for meeting (n+1)th grade of convolutional neural networks input area；Compressed image data is inputted into (n+1)th grade of volume
Product neural network；The second classification score is carried out to compressed image data by (n+1)th grade of convolutional neural networks to calculate；If
Second classification score is then judged as that image data does not include target object less than the second predetermined score value.
Optionally, if the second classification score is greater than or equal to the second predetermined score value, and when n=N-1, then it is judged as figure
As data include target object.
Optionally, after judging that image data includes target object, further includes: output includes the picture number of target object
According to characteristic information.
The embodiment of the invention also provides a kind of image detection devices based on concatenated convolutional neural network, comprising:
Image data acquisition unit, for obtaining the image data of image to be detected；Detection unit, being used for will be described to be checked
The input for the neural network model that the image data of altimetric image is established as above-mentioned training device to described image to be detected into
Row detection, obtains the testing result of described image to be detected.
Optionally, further includes: area division unit obtains each area for described image data to be divided into multiple regions
The image data in domain；The detection unit is for successively regarding the image data of described each region as the neural network respectively
The input of model detects the image data of described each region, obtains the testing result of described each region.
Optionally, the detection unit includes: sub-unit of classifying to obtain, for the N grade cascade in the neural network model
Classification score is carried out to described image data in at least one level neural network of convolutional neural networks to calculate；Comparing unit is used for
The calculated value of the classification score of at least one level convolutional neural networks is compared at least one predetermined score value, and according at least one
Comparison result determines whether described image data include target object.
Optionally, further includes: compression unit, for by the size compression of described image data at meeting n-th grade of convolution mind
Size through network inputs region, wherein n is positive integer, and 1≤n≤N-1；The detection unit is used for compressed institute
It states image data and inputs n-th grade of convolutional neural networks；First classifies to obtain sub-unit, for passing through n-th grade of convolution mind
The first classification score is carried out to compressed described image data through network to calculate；Judging unit, if being used for described first point
Class score is then judged as that described image data do not include target object less than the first predetermined score value.
The embodiment of the invention also provides a kind of computer storage medium, store for realizing offer of the embodiment of the present invention
Any image detecting method based on concatenated convolutional neural network computer-readable instruction.For example, described instruction
It include: the instruction for obtaining the image data of image to be detected；For using the image data of image to be detected as above-mentioned
The input for the neural network model that training method is established detects image to be detected, obtains the testing result of image to be detected
Instruction.
The embodiment of the invention also provides a kind of image detecting systems based on concatenated convolutional neural network, comprising:
Image acquiring device, for obtaining the image data of image to be detected；Memory, for storing program；Processor,
The image data for receiving image to be detected, the operation in above-mentioned detection method is realized for executing program.
According to the third aspect, the embodiment of the invention provides a kind of concatenated convolutional neural networks, comprising:
Cascade N grades of convolutional neural networks, N grades of convolutional neural networks for receiving the different size of input area of N kind
Image data, be trained/detect for the image data respectively to the different size of input area of N kind, N be greater than or wait
In 2 integer；At least training result that convolutional neural networks at different levels export respectively is associated, and the training result after association is returned
Reach convolutional neural networks at different levels.
Technical solution provided in an embodiment of the present invention handles the image data of image to be learned at least regional area respectively
At the image data of N kind input area, and N grades of cascade convolutional neural networks are separately input into, then, to N grades of cascade volumes
Product neural network is trained, therefore training result, is being propagated to convolutional neural networks at different levels by the association of an at least training result
When, the parameter of neural networks at different levels can be adjusted, concatenated convolutional neural network is enabled to reach neural network ginseng in training
Several global optimizations.
Detailed description of the invention
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art be briefly described, it should be apparent that, it is described below
Attached drawing is some embodiments of the present invention, for those of ordinary skill in the art, before not making the creative labor
It puts, is also possible to obtain other drawings based on these drawings.
Fig. 1 is a kind of concatenated convolutional neural network training method flow chart in the embodiment of the present invention 1；
Fig. 2 is a kind of concatenated convolutional neural metwork training device principle block diagram in the embodiment of the present invention 1；
Fig. 3 is a kind of image detecting method flow chart based on concatenated convolutional neural network in the embodiment of the present invention 2；
Fig. 4 is image detecting method flow chart of the another kind based on concatenated convolutional neural network in the embodiment of the present invention 2；
Fig. 5 is a kind of image detection device functional block diagram based on concatenated convolutional neural network in the embodiment of the present invention 2；
Fig. 6 schematically shows the department of computer science of the terminal device or server that are suitable for being used to realize the embodiment of the present application
The structural schematic diagram of system.
Specific embodiment
Technical solution of the present invention is clearly and completely described below in conjunction with attached drawing, it is clear that described implementation
Example is a part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill
Personnel's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
As long as in addition, the non-structure each other of technical characteristic involved in invention described below different embodiments
It can be combined with each other at conflict.
Embodiment 1
In order to realize the training of Multilevel ANN global optimum, present embodiment discloses a kind of concatenated convolutional neural network instructions
Practice method, referring to FIG. 1, being the concatenated convolutional neural network training method flow chart, this method comprises the following steps:
Step S110 obtains the image data of image to be learned at least regional area.In a particular embodiment, cunning can be used
Dynamic window selection method selects learning region of at least regional area of image to be learned as the image.In optional embodiment
In, each learning region can be irised out with for example rectangular bounding box, and according to objects in images zone boundary frame true value
Registration, to mark whether the learning region includes band detection object, in order to neural network learning training.In the present embodiment
In, each learning region can be tuned into as preset standard dimensions, such as 48*48, it should be noted that the present embodiment
And in following embodiments, in the case where not making especially to declare, specific numerical value does not constitute limit to the technical solution of the present embodiment
System, it should be understood that convenient for those skilled in the art understand that technical solution and the example lifted.
The image data of image to be learned at least regional area is processed into the figure of N kind input area by step S120 respectively
As data.In the present embodiment, N kind input area has different sizes between any two, and N is the integer more than or equal to 2.Optional
Embodiment in, the image data of image to be learned at least regional area can be compressed respectively, to obtain inputting in N
The image data in region.
The image data of N kind input area is input to N grades of cascade convolutional neural networks respectively by step S130.At this
In embodiment, each grade in N grades of cascade convolutional neural networks is a pair of with a kind of input area of N kind input area one respectively
It answers.Cascade convolutional neural networks have a network structure not at the same level, it is not at the same level between input data attribute type it is (such as different
The input data of grade respectively corresponds the data of different size of input area) it is different, to realize the number for being directed to same learning region
According to study, need respectively to handle the image data of the learning region (processing means may include but be not limited to compression, sampling etc.)
At the data size for meeting neural network inputs at different levels.In the present embodiment, therefore sharing N grades of convolutional neural networks cascades needs
There is the image data of N kind input area, every level-one convolutional neural networks correspond to a kind of input area.In preferred embodiment
In, the size of i-stage convolutional neural networks input area is less than the size of j-th stage convolutional neural networks input area, wherein i
It is positive integer, and 1≤i < j≤N with j.It is hereafter illustrated by taking N=3 as an example, the size of the 1st grade of input area is such as 12*
12, the size of the 2nd grade of input area is such as 24*24, the size of 3rd level input area is such as 48*48, needs to illustrate
It is in the present embodiment, to be not limiting as meeting certain multiple proportion between the size of input area not at the same level.It can will be by step
The size for the predeterminable area that S110 is obtained is tuned into respectively as 12*12,24*24 and 48*48, using as above-mentioned three-stage cascade mind
Input area through network.In the present embodiment, the size of the input area of prime convolutional neural networks is reduced, before can be improved
The operational efficiency of grade neural network, and becoming large-sized the input area of rear class convolutional neural networks, to meet rear class network
The demand for judging fineness enables convolutional neural networks at different levels to weigh efficiency and sentence using various sizes of input area
Disconnected fineness.
Step S140 is trained N grades of cascade convolutional neural networks.It in a particular embodiment, can be rule of thumb
Or it needs to determine neural network structures at different levels.Such as: in the case where ensuring recall rate, convolutional layer can be suitably reduced, with
Improve calculating speed.When the size of input area is bigger, then trains, fineness is higher, and corresponding computational complexity also mentions
Height, at this point it is possible to suitably reduce convolutional layer.It is illustrated by taking N=3 as an example: the 1st grade of neural network (such as 12net nerve net
Network) size of input area is, for example, 12*12, including such as 3 convolutional layers, 2 pond layers, wherein and convolutional layer exports image
Parameter for the convolution of input picture and convolution kernel, convolution kernel can be adjusted by training；Each pixel of pond layer output image
For the average value in some region of input picture.The 1st grade of mind is calculated by the output of the last one convolutional layer in the output of 12net
The same level through network loses loss0.The size of 2nd grade of neural network (such as 24net neural network) input area is, for example, 24*
24, including for example 1 convolutional layer, 1 pond layer and 1 full articulamentum, the output of 24net are exported by full articulamentum, are calculated
The same level to the 2nd grade of neural network loses loss1.The size of 3rd level neural network (such as 48net neural network) input area
For example, 48*48, including for example 2 convolutional layers, 2 pond layers and 1 full articulamentum, the output of 48net are defeated by full articulamentum
Out, the same level loss loss2 of 3rd level neural network is calculated.In a preferred embodiment, neural networks at different levels can be with benefit
The same level input area, which is calculated, with the loss function for returning bounding box true value returns the loss of bounding box true value, such as the recurrence of 12net
The recurrence bounding box true value of recurrence bounding box true value the bbox loss1 and 48net of bounding box true value bbox loss0,24net
bbox loss2。
Step S150 adjusts the parameter of neural networks at different levels according to training result.Specifically, by convolutional neural networks at different levels
At least training result exported respectively is associated, and by the training result after association be back to convolutional neural networks at different levels with
Adjust the parameter of neural networks at different levels.In the present embodiment, in training result, at least one parameter phase of convolutional neural networks at different levels
Association, by the parameter of the adjustable each layer network of the training result of neural network, thus, so that the neural network to it is global most
It is excellent close.In an alternate embodiment of the invention, which is that the same level of convolutional neural networks at different levels is lost, convolution minds at different levels
The same level loss function for being exported respectively through network and/or return the loss of bounding box true value and be weighted summation, obtain N grades it is cascade
The global loss loss of convolutional neural networks, specifically, the global loss loss of N grades of cascade convolutional neural networks is at least by each
The same level loss weighted sum of grade convolutional neural networks obtains, optionally, can be to convolutional Neural nets at different levels in weighted sum
The same level loss (loss0, loss1 and loss2) of network and the loss of recurrence bounding box true value at different levels (bbox loss0,
Bbox loss1 and bbox loss2) it carries out.In a particular embodiment, specific weighting coefficient can be rule of thumb and/or real
It tests to determine.
In order to realize that convolutional neural networks loss functions at different levels are associated, in an alternate embodiment of the invention, step is being executed
It when S140 is trained N grades of cascade convolutional neural networks, may further include: most according to level 1 volume product neural network
The output of later layer obtains the feature vector of level 1 volume product neural network；According to n-th grade of convolutional neural networks the last layer
The feature vector of output and (n-1)th grade of convolutional neural networks, obtains the feature vector of n-th grade of convolutional neural networks, wherein n is
Positive integer, and 1 < n≤N.In a particular embodiment, the feature vector according to obtained the same level convolutional neural networks is available
Prediction label then at least seeks the same level loss of convolutional neural networks at different levels respectively, specifically, can use neural network
Loss function and the loss function for returning bounding box true value are based on prediction label and true tag calculates the same level loss.By the calculating
Obtained loss propagates back to the corresponding position of neural network, and adjustment can be optimized to Parameters of Neural Network Structure.Tool
Body, it being illustrated by taking 3 cascade neural networks as an example, wherein the last layer of level 1 volume product neural network is convolutional layer,
The output of the convolutional layer is the feature vector A of level 1 volume product neural network；Level 2 volume accumulates neural network, including what is be sequentially connected in series
Convolutional layer, pond layer and full articulamentum, the last layer of the 2nd grade of neural network are full articulamentum, which exports feature
Vector B is then connected to obtain the feature vector A-B of the 2nd grade of neural network with the feature vector A of the 1st grade of neural network；3rd level
The last layer of neural network is full articulamentum, which exports feature vector C then feature with the 2nd grade of neural network
Vector A-B, which is connected, obtains the feature vector A-B-C of 3rd level neural network.In the present embodiment, level 1 volume product neural network is most
Later layer is convolutional layer, and the feature vector of level 1 volume product neural network is the last layer convolutional layer in this grade of convolutional neural networks
Output, the last layer of every grade of convolutional neural networks is full articulamentum in the 2nd to N grade of convolutional neural networks, is being embodied
In, the feature vector that convolutional layer exports can be adjusted to meet the data format of the feature vector of articulamentum connection.By not
The feature vector of convolutional neural networks at the same level is attached, and is associated to realize neural network parameters at different levels.At this
In embodiment, for convolutional neural networks not at the same level, level 1 volume product neural network is calculated using full convolutional layer, can be with
Save calculation amount；And the 2nd to N grade of convolutional neural networks the last layer is full articulamentum, and the differentiation of positive negative sample may be implemented.
The present embodiment also discloses a kind of concatenated convolutional neural metwork training device, referring to FIG. 2, for concatenated convolutional mind
Through network training apparatus structure schematic block diagram, which includes: learning data acquiring unit 110, training unit 120 and returns
Leaflet member 130, in which:
Learning data acquiring unit 110 is used to the image data of image to be learned at least regional area being processed into N respectively
The image data of the different size of input area of kind, N are the integer more than or equal to 2；Training unit 120 is used for respectively by N kind
Input of the image data of input area as convolutional neural networks at different levels in N grades of cascade convolutional neural networks, to volumes at different levels
Product neural network is trained, wherein every level-one convolutional neural networks in N grades of cascade convolutional neural networks respectively with N kind
A kind of input area of input area is corresponding；Back propagation unit 130 be used for convolutional neural networks at different levels are exported respectively at least one
Training result is associated, and the training result after association is back to convolutional neural networks at different levels to adjust neural networks at different levels
Parameter.
In an alternate embodiment of the invention, training unit 120 includes: primary vector unit, for according to level 1 volume product nerve net
The output of network the last layer obtains the feature vector of level 1 volume product neural network；Secondary vector unit, for being rolled up according to n-th grade
The output of product neural network the last layer and the feature vector of (n-1)th grade of convolutional neural networks, obtain n-th grade of convolutional neural networks
Feature vector, wherein n is positive integer, and 1 < n≤N.
In an alternate embodiment of the invention, training unit 120 includes: that unit is sought in the same level loss, each at least seeking respectively
The same level loss of grade convolutional neural networks；Unit is sought in overall situation loss, at least exporting respectively to convolutional neural networks at different levels
The same level loss be weighted summation, obtain N grade cascade convolutional neural networks the overall situation lose.
The present embodiment also discloses a kind of concatenated convolutional neural metwork training system, comprising: photographic device, memory and place
Manage device, in which:
Image acquiring device, for obtaining the image data of image to be learned；Memory, for storing program；Processor,
The image data for receiving image to be learned realizes the behaviour in above-mentioned concatenated convolutional neural network training method for executing program
Make.
Concatenated convolutional neural network training method provided in this embodiment, apparatus and system, by image to be learned at least office
The image data in portion region is processed into the image data of N kind input area respectively, and is separately input into N grades of cascade convolutional Neurals
Network is then trained N grades of cascade convolutional neural networks, therefore the association of an at least training result, will train knot
When fruit propagates to convolutional neural networks at different levels, the parameter of neural networks at different levels can be adjusted, so that concatenated convolutional neural network exists
Neural network parameter global optimization can be reached when training.
Embodiment 2
Present embodiment discloses a kind of image detecting methods based on concatenated convolutional neural network, referring to FIG. 3, being the base
In the image detecting method flow chart of concatenated convolutional neural network, which includes the following steps:
Step S210, training concatenated convolutional neural network model.It, can the grade according to disclosed in embodiment 1 in the present embodiment
Connection convolutional neural networks training method training neural network obtains concatenated convolutional neural network model.It should be noted that this reality
It applies in example, step S10 is that execution can no longer execute the step after the completion of neural metwork training in training neural network.
Step S220 obtains the image data of image to be detected.In a particular embodiment, can in advance to image data into
Row pretreatment, obtains the image data of pretreated image to be detected, it is alternatively possible to which the pixel value of image data is subtracted
A certain numerical value, the numerical value can be the mean value of ImageNet data set, be also possible to be closed according to training pictures calculated equal
Value.Certainly, in other alternative embodiments, the pretreatment of binaryzation can also be carried out to image data in advance.
Step S230 detects the image data of image to be detected, obtains testing result.It, can be in the present embodiment
Neural network model is established according to the training method of above-described embodiment, then, after the image data for obtaining image to be detected,
Image to be detected is detected using the image data of the image to be detected as the input of the neural network model, is obtained to be checked
The testing result of altimetric image.In an alternate embodiment of the invention, when detecting to image to be detected, it can detecte image to be detected
The location information (such as coordinate position) of middle target object, the location information of target object can be used as testing result output.
It should be noted that in a particular embodiment, carrying out image detection for the first time, executes step S210 training and establish
To after concatenated convolutional neural network, in subsequent detection process, it may not be necessary to execute step S210 again.
It should be noted that the image data of image to be detected that step S220 is obtained can be big for preset standard size
It is small, such as 48*48.
Certainly, the image data of image to be detected that step S220 is obtained is also possible to other sizes, at this point, executing step
After rapid S220, image to be detected data of the acquisition can be divided into multiple regions, to obtain the picture number of each region
According to then being detected respectively to each region, i.e., successively regard the image data of each region as neural network model respectively
Input the image data of each region is detected, obtain the testing result of each region.In an alternate embodiment of the invention, exist
When the image data of the image to be detected that will acquire is divided into multiple regions, each region can be tuned into as preset gauge
Very little size, such as 48*48.It in a particular embodiment, should be respectively by the image data of each region when executing step S230
Neural network model is sequentially input, successively the image data to each region should carry out the inspection that detection respectively obtains each region
Survey result.
In a particular embodiment, when being detected, can be to detect image to be detected by way of score of classifying
No includes target object, specifically, in at least one level neural network of the N grade concatenated convolutional neural network of neural network model
Classification score is carried out to image data to calculate, the calculated value of the classification score of at least one level convolutional neural networks is pre- at least one
Determine score value to be compared, and determines whether image data includes target object according to an at least comparison result.As an example, if
Score of classifying is less than predetermined score value, then judgement is characterized combination not comprising target object.Specifically, referring to FIG. 4, executing step
After rapid S220, further includes:
Step S250, by the size compression of image data at the size for meeting n-th grade of convolutional neural networks input area,
In, n is positive integer, and 1≤n≤N-1.For example, by the Image Data Compression of step S220 image to be detected obtained at meeting
The size (such as 12*12) of level 1 volume product neural network input area.
When executing step S230, may include:
Compressed image data is inputted n-th grade of convolutional neural networks by step S231.For example, step S250 is compressed
It is input at the image data of the image to be detected for the size (such as 12*12) for meeting level 1 volume product neural network input area
Level 1 volume accumulates neural network.
Step S241 carries out the first classification score to compressed image data by n-th grade of convolutional neural networks and calculates.
For example, after compressed image data to be input to level 1 volume product neural network, available first classification score, at this
In embodiment, classifying for this grade of convolutional neural networks can be exported by the convolutional layer of the 1st grade of neural network the last layer
Divide cls score 1.In a particular embodiment, a certain element in the feature vector of neural networks at different levels characterizes this grade of nerve
A certain element in the classification score of network, such as the feature vector A of the 1st grade of neural network characterizes classification score cls
score 1.It should be noted that in a particular embodiment, when n is such as 2, then the first classification is scored at the 2nd grade of nerve net
The classification score cls score 2 that network calculates.
Whether step S242 judges the first classification score less than the first predetermined score value.It in a particular embodiment, can basis
Experience determines the first predetermined score value.If step S242 judges that the first classification score less than the first predetermined score value, judges to tie
Fruit is that image data does not include target object.
In an alternate embodiment of the invention, if step S242 judges that the first classification score is greater than or equal to the first predetermined score value,
Then further include:
Step S243, by the size compression of image data at the size for meeting (n+1)th grade of convolutional neural networks input area.
For example, as n=1, if the 1st grade of classification score cls score 1 is less than this grade of predetermined value, by the size pressure of image data
Shorten the size (such as 24*24) for meeting level 2 volume product neural network input area into；As n=2, if the 2nd grade of classification score
Cls score 2 is less than this grade of predetermined value, then by the size compression of image data at meeting 3rd level convolutional neural networks input area
The size (such as 48*48) in domain.
Compressed image data is inputted (n+1)th grade of convolutional neural networks by step S244.
Step S245 carries out the second classification score meter to compressed image data by (n+1)th grade of convolutional neural networks
It calculates.After compressed image data is input to convolutional neural networks at different levels, available the same level neural network classification score,
As n=1, then the second classification is scored at the classification score cls score 2 of the 2nd grade of neural computing；As n=2, then
Two classification are scored at the classification score cls score 3 of 3rd level neural computing.
Whether step S246 judges the second classification score less than the second predetermined score value.It in a particular embodiment, can basis
Experience determines the second predetermined score value.If step S242 judges that the second classification score less than the second predetermined score value, judges to tie
Fruit is that image data does not include target object.
In a preferred embodiment, when executing step S246, if the second classification score is greater than or equal to second predetermined point
Value, and when n=N-1, then it is judged as that image data includes target object.
In a preferred embodiment, it after judging that image data includes target object, can further include: output
The characteristic information of image data comprising target object.In a particular embodiment, it can be finely tuned according to the adjustment amount of prediction to be checked
Survey position, the size of zone boundary frame.After detecting all area to be tested of a picture, selecting prediction includes pair
The region of elephant, for example non-maximum value of use inhibit method to remove the higher region of registration.As target in the detection method prognostic chart
The position of object and size.
The present embodiment also discloses a kind of image detection device based on concatenated convolutional neural network, referring to FIG. 5, for should
Image detection device structural schematic block diagram based on concatenated convolutional neural network, the detection device include: the grade of above-described embodiment
Join convolutional neural networks training device 1, image data acquisition unit 2 and detection unit 3, in which:
Image data acquisition unit 2 is used to obtain the image data of image to be detected；Detection unit is used for mapping to be checked
The image data of picture detects image to be detected as the input for the neural network model that above-mentioned training device is established, and obtains
To the testing result of image to be detected.
In an alternate embodiment of the invention, it is somebody's turn to do the image detection device based on concatenated convolutional neural network further include: region division
Unit obtains the image data of each region for partitioning image data into multiple regions；Detection unit is used for respectively will be each
The image data in a region is successively used as the input of neural network model to detect the image data of each region, obtains each
The testing result in a region.
In a preferred embodiment, detection unit includes: sub-unit of classifying to obtain, for the N grade grade in neural network model
Join in at least one level neural network of convolutional neural networks and classification score calculating is carried out to image data；Comparing unit, being used for will
The calculated value of the classification score of at least one level convolutional neural networks is compared at least one predetermined score value, and according at least one ratio
Relatively result determines whether image data includes target object.
In a preferred embodiment, it is somebody's turn to do the image detection device based on concatenated convolutional neural network further include: compression unit,
For by the size compression of image data at the size for meeting n-th grade of convolutional neural networks input area, wherein n is positive integer,
And 1≤n≤N-1；Detection unit is used to compressed image data inputting n-th grade of convolutional neural networks；First classification score
Unit is calculated for carrying out the first classification score to compressed image data by n-th grade of convolutional neural networks；Judgement is single
Member, if being judged as that image data does not include target object for the first classification score less than the first predetermined score value.
The present embodiment also discloses a kind of image detecting system based on concatenated convolutional neural network, comprising:
Image acquiring device, for obtaining the image data of image to be learned；Memory, for storing program；Processor,
The image data for receiving image to be learned realizes the operation in above-mentioned detection method for executing program.
The present embodiment also discloses a kind of cascade convolutional neural networks, comprising: cascade N grades of convolutional neural networks, N grades
The image data for being used to receive the different size of input area of N kind of convolutional neural networks, for different size of to N kind respectively
The image data of input area is trained/detects, and N is the integer more than or equal to 2；Convolutional neural networks at different levels export respectively
An at least training result it is associated, the training result after association is back to convolutional neural networks at different levels.
Image detecting method provided in this embodiment based on concatenated convolutional neural network, apparatus and system, in nerve net
When network training, training result is propagated into convolutional neural networks at different levels, adjusts the parameter of neural networks at different levels, so that concatenated convolutional
Neural network can reach the global optimization of neural network parameter in training.Thus when being detected using the network, it can
To obtain better detection effect while improving and detecting speed.
Below with reference to Fig. 6, it illustrates the calculating of the terminal device or server that are suitable for being used to realize the embodiment of the present application
The structural schematic diagram of machine system 600.
As shown in fig. 6, computer system 600 includes central processing unit (CPU) 601, it can be read-only according to being stored in
Executable instruction in memory (ROM) 602 is loaded into random access storage device (RAM) 603 from storage section 608
Executable instruction and execute various movements appropriate and processing.CPU601 can also be stored with system 600 and operate required various journeys
Sequence and data.CPU601, ROM602 and RAM603 are connected with each other by bus 604.Input/output (I/O) interface 605 also connects
It is connected to bus 604.
I/O interface 605 is connected to lower component: the importation 606 including keyboard, mouse etc.；It is penetrated including such as cathode
The output par, c 607 of spool (CRT), liquid crystal display (LCD) etc. and loudspeaker etc.；Storage section 608 including hard disk etc.；
And the communications portion 609 of the network interface card including LAN card, modem etc..Communications portion 609 via such as because
The network of spy's net executes communication process.Driver 610 is also connected to I/O interface 605 as needed.Detachable media 611, such as
Disk, CD, magneto-optic disk, semiconductor memory etc. are mounted on as needed on driver 610, in order to read from thereon
Computer program be mounted into storage section 608 as needed.
Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description
Software program.For example, embodiment of the disclosure includes a kind of computer program product comprising be tangibly embodied in machine respectively into
Row image block is divided to obtain the instruction of image block set；According to each image block in image block set generate at least one first
The instruction of image block group；According to the instruction of at least one the first image block group training convolutional neural networks；Based on the first convolution mind
Classify to each image block in image block set through network to obtain the instruction ... of at least one the second image block group.?
In such embodiment, which can be downloaded and installed from network by communications portion 609, and/or from can
Medium 611 is dismantled to be mounted.When the computer program is executed by central processing unit (CPU) 601, the present processes are executed
The above-mentioned function of middle restriction.
It should be understood by those skilled in the art that, the embodiment of the present invention can provide as method, system or computer program
Product.Therefore, complete hardware embodiment, complete software embodiment or reality combining software and hardware aspects can be used in the present invention
Apply the form of example.Moreover, it wherein includes the computer of computer usable program code that the present invention, which can be used in one or more,
The computer program implemented in usable storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.) produces
The form of product.
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that every one stream in flowchart and/or the block diagram can be realized by computer program instructions
The combination of process and/or box in journey and/or box and flowchart and/or the block diagram.It can provide these computer programs
Instruct the processor of general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices to produce
A raw machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute for real
The device for the function of being specified in present one or more flows of the flowchart and/or one or more blocks of the block diagram.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that it includes referring to that instruction stored in the computer readable memory, which generates,
Enable the manufacture of device, the command device realize in one box of one or more flows of the flowchart and/or block diagram or
The function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that counting
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, thus in computer or
The instruction executed on other programmable devices is provided for realizing in one or more flows of the flowchart and/or block diagram one
The step of function of being specified in a box or multiple boxes.
Obviously, the above embodiments are merely examples for clarifying the description, and does not limit the embodiments.It is right
For those of ordinary skill in the art, can also make on the basis of the above description it is other it is various forms of variation or
It changes.There is no necessity and possibility to exhaust all the enbodiments.And it is extended from this it is obvious variation or
It changes still within the protection scope of the invention.
Claims (22)
1. a kind of concatenated convolutional neural network training method characterized by comprising
The image data of image to be learned at least regional area is processed into the picture number of the different size of input area of N kind respectively
According to the N is the integer more than or equal to 2；
Respectively using the image data of the N kind input area as convolutional Neural nets at different levels in N grades of cascade convolutional neural networks
The input of network is trained convolutional neural networks at different levels, wherein every level-one in the N grades of cascade convolutional neural networks
Convolutional neural networks are corresponding with a kind of input area of the N kind input area respectively；
At least training result that convolutional neural networks at different levels are exported respectively is associated, and the training result after association is returned
Convolutional neural networks at different levels are reached to adjust the parameter of neural networks at different levels.
2. concatenated convolutional neural network training method as described in claim 1, which is characterized in that described to convolutional Neurals at different levels
Network is trained, comprising:
According to the output of level 1 volume product neural network the last layer, the feature vector of the level 1 volume product neural network is obtained；
According to the output of n-th grade of convolutional neural networks the last layer and the feature vector of (n-1)th grade of convolutional neural networks, the is obtained
The feature vector of n grades of convolutional neural networks, wherein the n is positive integer, and 1 < n≤N.
3. concatenated convolutional neural network training method as claimed in claim 2, which is characterized in that
The last layer of level 1 volume product neural network is convolutional layer；
The last layer of convolutional neural networks at different levels is full articulamentum in 2nd to N grade of convolutional neural networks.
4. concatenated convolutional neural network training method as claimed in any one of claims 1-3, which is characterized in that
It is described that convolutional neural networks at different levels are trained, comprising: at least to seek the same level damage of convolutional neural networks at different levels respectively
It loses；
At least training result that convolutional neural networks at different levels are exported respectively is associated, comprising: at least to convolution minds at different levels
The same level loss exported respectively through network is weighted summation, obtains the global loss of the N grades of cascade convolutional neural networks.
5. concatenated convolutional neural network training method as claimed in claim 4, which is characterized in that
The same level loss at least seeking convolutional neural networks at different levels respectively, comprising: seek convolutional neural networks at different levels respectively
The same level loss function and return bounding box true value loss；
Summation at least is weighted to the same level loss that convolutional neural networks at different levels export respectively, obtains the N grades of cascade volume
The global loss of product neural network, comprising: the same level loss function exported respectively to convolutional neural networks at different levels and/or recurrence side
Frame true value loss in boundary's is weighted summation, obtains the global loss of the N grades of cascade convolutional neural networks.
6. concatenated convolutional neural network training method as claimed in any one of claims 1-3, which is characterized in that i-stage volume
The size of product neural network input area is less than the size of j-th stage convolutional neural networks input area, wherein the i and j
For positive integer, and 1≤i < j≤N.
7. a kind of image detecting method based on concatenated convolutional neural network characterized by comprising
Obtain the image data of image to be detected；
The mind that the image data of described image to be detected is established as training method as claimed in any one of claims 1 to 6
Input through network model detects described image to be detected, obtains the testing result of described image to be detected.
8. the image detecting method as claimed in claim 7 based on concatenated convolutional neural network, which is characterized in that obtained described
After taking the image data of image to be detected, further includes:
Described image data are divided into multiple regions and obtain the image data of each region；
Described image to be detected is detected using the image data of described image to be detected as the input of neural network model,
Obtain the testing result of described image to be detected, comprising: successively regard the image data of described each region as the mind respectively
Input through network model detects the image data of described each region, obtains the testing result of described each region.
9. the image detecting method as claimed in claim 7 or 8 based on concatenated convolutional neural network, which is characterized in that described
Carrying out detection to described image to be detected includes:
To described image data in at least one level neural network of the N grade concatenated convolutional neural network of the neural network model
Classification score is carried out to calculate；
The calculated value of the classification score of at least one level convolutional neural networks is compared at least one predetermined score value, and according to extremely
A few comparison result determines whether described image data include target object.
10. the image detecting method as claimed in claim 8 based on concatenated convolutional neural network, which is characterized in that described
After the image data for obtaining image to be detected, further includes:
By the size compression of described image data at the size for meeting m grades of convolutional neural networks input areas, wherein m is positive
Integer, and 1≤m≤N-1；
Described image to be detected is detected using the image data of described image to be detected as the input of neural network model,
Obtain the testing result of described image to be detected, comprising: compressed described image data are inputted into the m grades of convolutional Neurals
Network；
The first classification score is carried out to compressed described image data by the m grades of convolutional neural networks to calculate；
If the first classification score is judged as that described image data do not include target object less than the first predetermined score value.
11. the image detecting method as claimed in claim 10 based on concatenated convolutional neural network, which is characterized in that if institute
State the first classification score more than or equal to the first predetermined score value, then further include:
By the size compression of described image data at the size for meeting m+1 grades of convolutional neural networks input areas；
Compressed described image data are inputted into the m+1 grades of convolutional neural networks；
The second classification score is carried out to compressed described image data by the m+1 grades of convolutional neural networks to calculate；
If the second classification score is judged as that described image data do not include target object less than the second predetermined score value.
12. the image detecting method as claimed in claim 11 based on concatenated convolutional neural network, which is characterized in that if institute
The second classification score is stated more than or equal to the second predetermined score value, and when m=N-1, is then judged as that described image data include target
Object.
13. the image detecting method as claimed in claim 12 based on concatenated convolutional neural network, which is characterized in that described
Judge described image data comprising after target object, further includes:
The characteristic information of described image data of the output comprising target object.
14. a kind of concatenated convolutional neural metwork training device characterized by comprising
Learning data acquiring unit is different for the image data of image to be learned at least regional area to be processed into N kind respectively
The image data of the input area of size, the N are the integer more than or equal to 2；
Training unit, for respectively using the image data of the N kind input area as each in N grades of cascade convolutional neural networks
The input of grade convolutional neural networks, is trained convolutional neural networks at different levels, wherein the N grades of cascade convolutional Neural net
Every level-one convolutional neural networks in network are corresponding with a kind of input area of the N kind input area respectively；
Back propagation unit, at least training result for exporting convolutional neural networks at different levels respectively are associated, and will association
Training result afterwards is back to convolutional neural networks at different levels to adjust the parameter of neural networks at different levels.
15. concatenated convolutional neural metwork training device as claimed in claim 14, which is characterized in that the training unit packet
It includes:
Primary vector unit obtains the level 1 volume product mind for the output according to level 1 volume product neural network the last layer
Feature vector through network；
Secondary vector unit, for according to n-th grade of convolutional neural networks the last layer output and (n-1)th grade of convolutional neural networks
Feature vector, obtain the feature vector of n-th grade of convolutional neural networks, wherein the n is positive integer, and 1 < n≤N.
16. the concatenated convolutional neural metwork training device as described in claims 14 or 15, which is characterized in that the training unit
Include:
Unit is sought in the same level loss, and the same level at least seeking convolutional neural networks at different levels respectively is lost；
Unit is sought in overall situation loss, is asked for being at least weighted to the same level loss that convolutional neural networks at different levels export respectively
With obtain the global loss of the N grades of cascade convolutional neural networks.
17. a kind of image detection device based on concatenated convolutional neural network characterized by comprising
Image data acquisition unit, for obtaining the image data of image to be detected；
Detection unit, for using the image data of described image to be detected as described in claim 14-16 any one
The input for the neural network model that training device is established detects described image to be detected, obtains described image to be detected
Testing result.
18. the image detection device as claimed in claim 17 based on concatenated convolutional neural network, which is characterized in that also wrap
It includes:
Area division unit obtains the image data of each region for described image data to be divided into multiple regions；
The detection unit is for successively regarding the image data of described each region as the defeated of the neural network model respectively
Enter and the image data of described each region is detected, obtains the testing result of described each region.
19. the image detection device based on concatenated convolutional neural network as described in claim 17 or 18, which is characterized in that institute
Stating detection unit includes:
Classify to obtain sub-unit, at least one level nerve net for the N grade concatenated convolutional neural network in the neural network model
Classification score is carried out to described image data in network to calculate；
Comparing unit, for carrying out the calculated value of the classification score of at least one level convolutional neural networks and at least one predetermined score value
Compare, and determines whether described image data include target object according to an at least comparison result.
20. the image detection device as claimed in claim 17 based on concatenated convolutional neural network, which is characterized in that also wrap
It includes:
Compression unit, for by the size compression of described image data at the ruler for meeting m grades of convolutional neural networks input areas
It is very little, wherein m is positive integer, and 1≤m≤N-1；
The detection unit is used to compressed described image data inputting the m grades of convolutional neural networks；
First classifies to obtain sub-unit, for being carried out by the m grades of convolutional neural networks to compressed described image data
First classification score calculates；
Judging unit, if being judged as described image data not less than the first predetermined score value for the first classification score
Include target object.
21. a kind of concatenated convolutional neural metwork training system characterized by comprising
Image acquiring device, for obtaining the image data of image to be learned；
Memory, for storing program；
Processor receives the image data of the image to be learned, for executing described program to realize that claim 1-6 such as appoints
Operation in the method for anticipating.
22. a kind of image detecting system based on concatenated convolutional neural network characterized by comprising
Image acquiring device, for obtaining the image data of image to be detected；
Memory, for storing program；
Processor receives the image data of described image to be detected, for executing described program to realize such as claim 7-13
Operation in any one the method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610439342.2A CN106096670B (en) | 2016-06-17 | 2016-06-17 | Concatenated convolutional neural metwork training and image detecting method, apparatus and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610439342.2A CN106096670B (en) | 2016-06-17 | 2016-06-17 | Concatenated convolutional neural metwork training and image detecting method, apparatus and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106096670A CN106096670A (en) | 2016-11-09 |
| CN106096670B true CN106096670B (en) | 2019-07-30 |
Family
ID=57236547
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610439342.2A Active CN106096670B (en) | 2016-06-17 | 2016-06-17 | Concatenated convolutional neural metwork training and image detecting method, apparatus and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106096670B (en) |
Families Citing this family (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108122234B (en) * | 2016-11-29 | 2021-05-04 | 北京市商汤科技开发有限公司 | Convolutional neural network training and video processing method and device and electronic equipment |
| US10331975B2 (en) * | 2016-11-29 | 2019-06-25 | Google Llc | Training and/or using neural network models to generate intermediary output of a spectral image |
| CN106651877B (en) * | 2016-12-20 | 2020-06-02 | 北京旷视科技有限公司 | Instance partitioning method and device |
| CN108242046B (en) * | 2016-12-27 | 2022-02-18 | 阿里巴巴集团控股有限公司 | Picture processing method and related equipment |
| CN106845406A (en) * | 2017-01-20 | 2017-06-13 | 深圳英飞拓科技股份有限公司 | Head and shoulder detection method and device based on multitask concatenated convolutional neutral net |
| CN108629354B (en) * | 2017-03-17 | 2020-08-04 | 杭州海康威视数字技术股份有限公司 | Target detection method and device |
| CN108305240B (en) * | 2017-05-22 | 2020-04-28 | 腾讯科技（深圳）有限公司 | Image quality detection method and device |
| CN108229284B (en) * | 2017-05-26 | 2021-04-09 | 北京市商汤科技开发有限公司 | Sight tracking and training method and device, system, electronic equipment and storage medium |
| US11741354B2 (en) | 2017-08-25 | 2023-08-29 | Ford Global Technologies, Llc | Shared processing with deep neural networks |
| CN107578453B (en) * | 2017-10-18 | 2019-11-01 | 北京旷视科技有限公司 | Compressed image processing method, apparatus, electronic equipment and computer-readable medium |
| CN109903350B (en) * | 2017-12-07 | 2021-08-06 | 上海寒武纪信息科技有限公司 | Image compression method and related device |
| US10599978B2 (en) | 2017-11-03 | 2020-03-24 | International Business Machines Corporation | Weighted cascading convolutional neural networks |
| CN110310287B (en) * | 2018-03-22 | 2022-04-19 | 北京连心医疗科技有限公司 | Automatic organ-at-risk delineation method, equipment and storage medium based on neural network |
| CN108510083B (en) * | 2018-03-29 | 2021-05-14 | 国信优易数据股份有限公司 | Neural network model compression method and device |
| CN108830209B (en) * | 2018-06-08 | 2021-12-17 | 西安电子科技大学 | Remote sensing image road extraction method based on generation countermeasure network |
| CN109460704B (en) * | 2018-09-18 | 2020-09-15 | 厦门瑞为信息技术有限公司 | Fatigue detection method and system based on deep learning and computer equipment |
| CN109583501B (en) * | 2018-11-30 | 2021-05-07 | 广州市百果园信息技术有限公司 | Method, device, equipment and medium for generating image classification and classification recognition model |
| CN109829371B (en) * | 2018-12-26 | 2022-04-26 | 深圳云天励飞技术有限公司 | Face detection method and device |
| CN109620269B (en) * | 2019-01-28 | 2021-10-22 | 锦图计算技术(深圳)有限公司 | Fatigue detection method, device, equipment and readable storage medium |
| CN109871791A (en) * | 2019-01-31 | 2019-06-11 | 北京字节跳动网络技术有限公司 | Image processing method and device |
| US10885625B2 (en) | 2019-05-10 | 2021-01-05 | Advanced New Technologies Co., Ltd. | Recognizing damage through image analysis |
| CN110569703B (en) * | 2019-05-10 | 2020-09-01 | 阿里巴巴集团控股有限公司 | Computer-implemented method and device for identifying damage from picture |
| CN110276411B (en) * | 2019-06-28 | 2022-11-18 | 腾讯科技（深圳）有限公司 | Image classification method, device, equipment, storage medium and medical electronic equipment |
| CN111291716B (en) * | 2020-02-28 | 2024-01-05 | 深圳市瑞图生物技术有限公司 | Sperm cell identification method, sperm cell identification device, computer equipment and storage medium |
| CN111507408B (en) * | 2020-04-17 | 2022-11-04 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN112784985A (en) * | 2021-01-29 | 2021-05-11 | 北京百度网讯科技有限公司 | Training method and device of neural network model, and image recognition method and device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102946613A (en) * | 2012-10-10 | 2013-02-27 | 北京邮电大学 | Method for measuring QoE |
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
| CN104657748A (en) * | 2015-02-06 | 2015-05-27 | 中国石油大学(华东) | Vehicle type recognition method based on convolutional neural network |
| CN104951668A (en) * | 2015-04-07 | 2015-09-30 | 上海大学 | Method for predicting protein association graphs on basis of cascade neural network structures |
| CN105404865A (en) * | 2015-11-16 | 2016-03-16 | 杭州电子科技大学 | Probability state restricted Boltzmann machine cascade based face detection method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
-
2016
- 2016-06-17 CN CN201610439342.2A patent/CN106096670B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102946613A (en) * | 2012-10-10 | 2013-02-27 | 北京邮电大学 | Method for measuring QoE |
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
| CN104657748A (en) * | 2015-02-06 | 2015-05-27 | 中国石油大学(华东) | Vehicle type recognition method based on convolutional neural network |
| CN104951668A (en) * | 2015-04-07 | 2015-09-30 | 上海大学 | Method for predicting protein association graphs on basis of cascade neural network structures |
| CN105404865A (en) * | 2015-11-16 | 2016-03-16 | 杭州电子科技大学 | Probability state restricted Boltzmann machine cascade based face detection method |
Non-Patent Citations (2)
| Title |
|---|
| A convolutional neural network cascade for face detection;HaoXiang Li et al.;《2015 IEEE Conference on Computer Vision and Pattern Recognition》;20150630;第5325-5334页 |
| 基于多层卷积神经网络学习的单帧图像超分辨率重建方法;刘娜 李翠华;《中国科技论文》;20150131;第10卷(第2期);第201-206页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106096670A (en) | 2016-11-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106096670B (en) | Concatenated convolutional neural metwork training and image detecting method, apparatus and system | |
| CN109685152B (en) | Image target detection method based on DC-SPP-YOLO | |
| CN111914944B (en) | Object detection method and system based on dynamic sample selection and loss consistency | |
| CN111126472A (en) | Improved target detection method based on SSD | |
| CN109635685A (en) | Target object 3D detection method, device, medium and equipment | |
| CN108898610A (en) | A kind of object contour extraction method based on mask-RCNN | |
| CN106599789A (en) | Video class identification method and device, data processing device and electronic device | |
| CN107481188A (en) | A kind of image super-resolution reconstructing method | |
| CN112084869B (en) | Compact quadrilateral representation-based building target detection method | |
| CN109101897A (en) | Object detection method, system and the relevant device of underwater robot | |
| CN109543701A (en) | Vision significance method for detecting area and device | |
| CN107529650A (en) | Network model construction and closed loop detection method, corresponding device and computer equipment | |
| CN111640089A (en) | Defect detection method and device based on feature map center point | |
| CN111126278A (en) | Target detection model optimization and acceleration method for few-category scene | |
| JP2020098587A (en) | Object Shape Regression Using Wasserstein Distance | |
| CN111274981B (en) | Target detection network construction method and device and target detection method | |
| CN114782311B (en) | CENTERNET improvement-based multi-scale defect target detection method and system | |
| CN108875587A (en) | Target distribution detection method and equipment | |
| CN113160117A (en) | Three-dimensional point cloud target detection method under automatic driving scene | |
| CN117788957B (en) | Deep learning-based qualification image classification method and system | |
| CN115587964A (en) | Entropy screening-based pseudo label cross consistency change detection method | |
| CN110472673B (en) | Parameter adjustment method, fundus image processing device, fundus image processing medium and fundus image processing apparatus | |
| CN110189318A (en) | Pulmonary nodule detection method and system with semantic feature score | |
| CN111445025B (en) | Method and device for determining hyper-parameters of business model | |
| CN115953330B (en) | Texture optimization method, device, equipment and storage medium for virtual scene image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20190710 Address after: 518000 Room 201, building A, No. 1, Qian Wan Road, Qianhai Shenzhen Hong Kong cooperation zone, Shenzhen, Guangdong (Shenzhen Qianhai business secretary Co., Ltd.) Applicant after: SHENZHEN SHANGTANG SCIENCE & TECHNOLOGY CO., LTD. Address before: Room 710-712, 7th floor, No. 1 Courtyard, Zhongguancun East Road, Haidian District, Beijing Applicant before: Shang Tang Science and Technology Development Co., Ltd. of Beijing |