CN108304821A - Image-recognizing method and device, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing - Google Patents
Image-recognizing method and device, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing Download PDFInfo
- Publication number
- CN108304821A CN108304821A CN201810151420.8A CN201810151420A CN108304821A CN 108304821 A CN108304821 A CN 108304821A CN 201810151420 A CN201810151420 A CN 201810151420A CN 108304821 A CN108304821 A CN 108304821A
- Authority
- CN
- China
- Prior art keywords
- image
- layer
- scene
- feature
- neural networks
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 61
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 112
- 238000012549 training Methods 0.000 claims abstract description 89
- 238000012360 testing method Methods 0.000 claims abstract description 73
- 238000003475 lamination Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 45
- 238000012545 processing Methods 0.000 claims description 24
- 238000003909 pattern recognition Methods 0.000 claims description 21
- 238000007781 pre-processing Methods 0.000 claims description 15
- 238000012790 confirmation Methods 0.000 claims description 14
- 230000035945 sensitivity Effects 0.000 claims description 11
- 230000007797 corrosion Effects 0.000 claims description 7
- 238000005260 corrosion Methods 0.000 claims description 7
- 230000008447 perception Effects 0.000 claims description 7
- 238000012986 modification Methods 0.000 claims description 4
- 230000004048 modification Effects 0.000 claims description 4
- 238000003062 neural network model Methods 0.000 claims description 4
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 238000004364 calculation method Methods 0.000 abstract description 5
- 230000006870 function Effects 0.000 description 20
- 238000010586 diagram Methods 0.000 description 13
- 238000003384 imaging method Methods 0.000 description 6
- 230000001537 neural effect Effects 0.000 description 6
- 238000010276 construction Methods 0.000 description 5
- 230000004913 activation Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- 238000005070 sampling Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012886 linear function Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000006641 stabilisation Effects 0.000 description 2
- 238000011105 stabilization Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 239000000872 buffer Substances 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 235000012054 meals Nutrition 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of image-recognizing method based on multilayer convolutional neural networks and device, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing.The image-recognizing method and device based on multilayer convolutional neural networks of embodiment of the present invention, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing structure three-layer coil lamination add the multilayer convolutional neural networks model of two layers of pond layer, and the training image that first resolution is normalized to using resolution ratio trains multilayer convolutional neural networks model, the test image that second resolution is normalized to using resolution ratio tests multilayer convolutional neural networks model, the identification to image scene can be realized without using full articulamentum, reduce the complexity of scene Recognition algorithm, the calculation amount of scene Recognition is smaller, it calculates time-consuming shorter.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of image based on multilayer convolutional neural networks model
Recognition methods, the pattern recognition device based on multilayer convolutional neural networks model, image acquiring method, image acquisition equipment, meter
Calculate machine equipment and non-volatile computer readable storage medium storing program for executing.
Background technology
The existing feature using engineer is long, poor robustness there are the design cycle to identify the method for image scene
Disadvantage, and it is poor for the recognition capability of complicated image scene.And the scene based on convolutional neural networks is recognition methods need
Full articulamentum is used, there are defects that is computationally intensive, calculating time length.
Invention content
The image-recognizing method that the embodiment provides a kind of based on multilayer convolutional neural networks model, based on more
The layer pattern recognition device of convolutional neural networks model, image acquiring method, image acquisition equipment, computer equipment and it is non-easily
The property lost computer readable storage medium.
The present invention provides a kind of image-recognizing method based on multilayer convolutional neural networks model, described image recognition methods
Including:
Target category marked to every width training image gathered in advance, and to training image described in every width pre-processed with
Obtain the training image of several first resolutions;
The initial configuration of the multilayer convolutional neural networks model is set, the initial configuration is tactic first layer
Convolutional layer, second layer pond layer, third layer convolutional layer, the 4th layer of pond layer and layer 5 convolutional layer;
An at least width is calculated according to the first parameter of the training image of first resolution and first layer convolutional layer
Fisrt feature image;
The fisrt feature image is inputted into second layer pond layer to be calculated and the fisrt feature image one
One corresponding second feature image;
An at least width third is calculated according to the third parameter of the second feature image and the third layer convolutional layer
Characteristic image;
The third feature image is inputted into the 4th layer of pond layer to be calculated and the third feature image one
One corresponding fourth feature image;
It is calculated at least one the 5th according to the 5th parameter of the fourth feature image and layer 5 pond layer
Characteristic image;
The scene Recognition result of training image described in every width is confirmed according to the fifth feature image;
The loss value of the multilayer convolutional neural networks is calculated according to the target category and the scene Recognition result;
The multilayer convolutional neural networks model convergence is confirmed when the loss value is less than default loss value;
The test image of acquisition is pre-processed to obtain the test image of several second resolutions, described second
Resolution ratio is more than the first resolution;
The test image is inputted to the convergent multilayer convolutional neural networks model to test the convergent multilayer
Convolutional neural networks model;With
Using the scene type in the multilayer convolutional neural networks Model Identification scene image after test.
The present invention provides a kind of image acquiring method, and described image acquisition methods include:
Obtain scene image;
Using the scene type in scene image described in above-mentioned multilayer convolutional neural networks Model Identification;
The acquisition parameters of camera are adjusted to obtain new scene corresponding with the scene image according to the scene type
Image, the acquisition parameters include at least one of colour temperature, time for exposure, sensitivity and exposure compensating.
The present invention provides a kind of pattern recognition device based on multilayer convolutional neural networks model.Described image identification device
Mould is calculated including the first preprocessing module, setting module, the first computing module, the second computing module, third computing module, the 4th
Block, the 5th computing module, the first confirmation module, the 6th computing module, second the second preprocessing module of confirmation module, test module
And identification module.First preprocessing module is used to mark target category to every width training image gathered in advance, and to every
Training image described in width is pre-processed to obtain the training image of several first resolutions.The setting module is for setting
The initial configuration of the fixed multilayer convolutional neural networks model, the initial configuration are tactic first layer convolutional layer, the
Two layers of pond layer, third layer convolutional layer, the 4th layer of pond layer and layer 5 convolutional layer.First computing module is used for according to the
An at least width fisrt feature image is calculated with the first parameter of first layer convolutional layer in the training image of one resolution ratio.Institute
The second computing module is stated for the fisrt feature image to be inputted second layer pond layer to be calculated and described first
The one-to-one second feature image of characteristic image.The third computing module be used for according to the second feature image with it is described
An at least width third feature image is calculated in the third parameter of third layer convolutional layer.4th computing module is used for will be described
It is special with the third feature image the one-to-one 4th to be calculated that third feature image inputs the 4th layer of pond layer
Levy image.5th computing module is based on the 5th parameter according to the fourth feature image and layer 5 pond layer
Calculation obtains an at least width fifth feature image.First confirmation module is used to confirm every width institute according to the fifth feature image
State the scene Recognition result of training image.6th computing module is used for according to the target category and the scene Recognition knot
Fruit calculates the loss value of the multilayer convolutional neural networks.Second confirmation module is used to be less than default damage in the loss value
The multilayer convolutional neural networks model convergence is confirmed when consumption value.Second preprocessing module is used for the test image to acquisition
It is pre-processed to obtain the test image of several second resolutions, the second resolution is more than described first and differentiates
Rate.The test module is for inputting the test image to the convergent multilayer convolutional neural networks model to test convergence
The multilayer convolutional neural networks model.The identification module is used for using the multilayer convolutional neural networks mould after test
Type identifies the scene type in scene image.
The present invention provides a kind of image acquisition equipment, and it includes acquiring unit and image recognition dress that described image, which obtains equipment,
It sets.The acquiring unit is for obtaining scene image.Described image identification device is used to use above-mentioned multilayer convolutional Neural net
Scene type in scene image described in network Model Identification.The acquiring unit is additionally operable to be adjusted according to the scene type and image
The acquisition parameters of head are to obtain new scene image corresponding with the scene image, when the acquisition parameters include colour temperature, exposure
Between, at least one of sensitivity and exposure compensating.
The present invention provides a kind of computer equipment, including memory and processor, and computer is stored in the memory
Readable instruction, when the computer-readable instruction is executed by the processor so that the processor, which executes above-mentioned image, to be known
Other method and above-mentioned image acquiring method.
The present invention provides the non-volatile computer readable storage medium storing program for executing that one or more includes computer executable instructions,
When the computer executable instructions are executed by one or more processors so that the processor, which executes above-mentioned image, to be known
Other method and above-mentioned image acquiring method.
It the construction method and device, image acquiring method of the multilayer convolutional neural networks model of embodiment of the present invention and sets
Standby, computer equipment and non-volatile computer readable storage medium storing program for executing structure three-layer coil lamination add the multilayer convolution of two layers of pond layer
Neural network model, and using resolution ratio be normalized to first resolution training image train multilayer convolutional neural networks mould
Type, the test image that second resolution is normalized to using resolution ratio tests multilayer convolutional neural networks model, without using complete
The identification to image scene can be realized in articulamentum, reduces the complexity of scene Recognition algorithm, the calculation amount of scene Recognition compared with
It is small, it calculates time-consuming shorter.
The additional aspect and advantage of the present invention will be set forth in part in the description, and will partly become from the following description
Obviously, or practice through the invention is recognized.
Description of the drawings
Above-mentioned and/or additional aspect and advantage of the invention will become from the following description of the accompanying drawings of embodiments
Obviously and it is readily appreciated that, wherein:
Fig. 1 is that the flow of image-recognizing method of the certain embodiments of the present invention based on multilayer convolutional neural networks model is shown
It is intended to.
Fig. 2 is that the module of pattern recognition device of the certain embodiments of the present invention based on multilayer convolutional neural networks model is shown
It is intended to.
Fig. 3 is the module diagram of the computer equipment of certain embodiments of the present invention.
Fig. 4 is the scene of the image-recognizing method based on multilayer convolutional neural networks model of certain embodiments of the present invention
Schematic diagram.
Fig. 5 is the flow of the image-recognizing method based on multilayer convolutional neural networks model of certain embodiments of the present invention
Schematic diagram.
Fig. 6 is the module of the pattern recognition device based on multilayer convolutional neural networks model of certain embodiments of the present invention
Schematic diagram.
Fig. 7 is the flow of the image-recognizing method based on multilayer convolutional neural networks model of certain embodiments of the present invention
Schematic diagram.
Fig. 8 is the module of the pattern recognition device based on multilayer convolutional neural networks model of certain embodiments of the present invention
Schematic diagram.
Fig. 9 is the flow diagram of the image acquiring method of certain embodiments of the present invention.
Figure 10 is the module diagram of the image acquisition equipment of certain embodiments of the present invention.
Figure 11 is the module diagram of the computer equipment of certain embodiments of the present invention.
The module diagram of the image processing circuit of Figure 12 certain embodiments of the present invention.
Specific implementation mode
The embodiment of the present invention is described below in detail, examples of the embodiments are shown in the accompanying drawings, wherein from beginning to end
Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached
The embodiment of figure description is exemplary, it is intended to for explaining the present invention, and is not considered as limiting the invention.
Referring to Fig. 1, the present invention provides a kind of image-recognizing method based on multilayer convolutional neural networks model.Image is known
Other method includes:
00:Target category marked to every width training image gathered in advance, and to every width training image pre-processed with
Obtain the training image of several first resolutions;
01:The initial configuration of multilayer convolutional neural networks model is set, initial configuration is tactic first layer convolution
Layer, second layer pond layer, third layer convolutional layer, the 4th layer of pond layer and layer 5 convolutional layer;
02:It is calculated at least one the according to the first parameter of the training image of first resolution and first layer convolutional layer
One characteristic image;
03:Fisrt feature image is inputted into second layer pond layer to be calculated and fisrt feature image one-to-one the
Two characteristic images;
04:An at least width third feature figure is calculated according to the third parameter of second feature image and third layer convolutional layer
Picture;
05:Third feature image is inputted into the 4th layer of pond layer to be calculated and third feature image one-to-one
Four characteristic images;
06:An at least width fifth feature figure is calculated according to the 5th parameter of fourth feature image and layer 5 pond layer
Picture;
07:The scene Recognition result of every width training image is confirmed according to fifth feature image;
08:The loss value of multilayer convolutional neural networks is calculated according to target category and scene Recognition result;
09:The convergence of multilayer convolutional neural networks model is confirmed when loss value is less than default loss value;
011:The test image of acquisition is pre-processed to obtain the test image of several second resolutions, second differentiates
Rate is more than the first resolution;
012:Input test image tests convergent multilayer convolutional Neural to convergent multilayer convolutional neural networks model
Network model;With
013:Using the scene type in the multilayer convolutional neural networks Model Identification scene image after test.
Referring to Fig. 2, the present invention also provides a kind of pattern recognition devices 100 for rolling up neural network model based on multilayer.This
The image-recognizing method based on multilayer convolutional neural networks of invention embodiment can be by embodiment of the present invention based on more
The pattern recognition device 100 of layer convolutional neural networks is realized.Pattern recognition device 100 includes the first preprocessing module 30, setting
Module 31, the first computing module 32, the second computing module 33, third computing module 34, the 4th computing module the 35, the 5th calculate mould
Block 36, the first confirmation module 37, the 6th computing module 38, the second confirmation module 39, the second preprocessing module 41, test module 42
With identification module 43.Step 00 can be realized by the first preprocessing module 30.Step 01 can be realized by setting module 31.Step
02 can be realized by the first computing module 32.Step 03 can be realized by the second computing module 33.Step 04 can be by third meter
Module 34 is calculated to realize.Step 05 can be realized by the 4th computing module 35.Step 06 can be realized by the 5th computing module 36.Step
Rapid 07 can be realized by the first confirmation module 37.Step 08 can be realized by the 6th computing module 38.Step 09 can be by second
Confirmation module 39 is realized.Step 011 can be realized by the second preprocessing module 41.Step 012 can be real by test module 42
It is existing.Step 013 can be realized by identification module 43.
In other words, the first preprocessing module 30 can be used for marking target category to every width training image gathered in advance,
And every width training image is pre-processed to obtain the training image of several first resolutions.Setting module 31 can be used for setting
The initial configuration of multilayer convolutional neural networks model, initial configuration be tactic first layer convolutional layer, second layer pond layer,
Third layer convolutional layer, the 4th layer of pond layer and layer 5 convolutional layer.First computing module 32 can be used for according to first resolution
An at least width fisrt feature image is calculated in first parameter of training image and first layer convolutional layer.Second computing module 33 can
For fisrt feature image to be inputted second layer pond layer to be calculated and the one-to-one second feature of fisrt feature image
Image.Third computing module 34 can be used for being calculated at least according to the third parameter of second feature image and third layer convolutional layer
One width third feature image.4th computing module 35 can be used for third feature image inputting the 4th layer of pond layer to calculate
It obtains and the one-to-one fourth feature image of third feature image.5th computing module 36 can be used for according to fourth feature image
An at least width fifth feature image is calculated with the 5th parameter of layer 5 pond layer.First confirmation module 37 can be used for basis
Fifth feature image confirms the scene Recognition result of every width training image.6th computing module 38 can be used for according to target category and
Scene Recognition result calculates the loss value of multilayer convolutional neural networks.Second confirmation module 39 can be used for being less than in loss value default
The convergence of multilayer convolutional neural networks model is confirmed when loss value.Second preprocessing module 41 can be used for the test image of acquisition into
To obtain the test image of several second resolutions, second resolution is more than the first resolution for row pretreatment.Test module
42 can be used for input test image to convergent multilayer convolutional neural networks model to test convergent multilayer convolutional neural networks
Model.Identification module 43 can be used for using the scene class in the multilayer convolutional neural networks Model Identification scene image after test
Not.
Referring to Fig. 3, the present invention provides a kind of computer equipment 1000.Computer equipment 1000 includes memory 61 and place
Manage device 62.Computer-readable instruction 611 is stored in memory 61.When computer-readable instruction 611 is executed by processor 62, make
It obtains processor 62 and executes following operation:Target category is marked to every width training image gathered in advance, and to every width training image
It is pre-processed to obtain the training image of several first resolutions;The initial configuration of multilayer convolutional neural networks model is set,
Initial configuration is tactic first layer convolutional layer, second layer pond layer, third layer convolutional layer, the 4th layer of pond layer and the 5th
Layer convolutional layer;It is calculated at least one first according to the first parameter of the training image of first resolution and first layer convolutional layer
Characteristic image;Fisrt feature image is inputted into second layer pond layer to be calculated and fisrt feature image one-to-one second
Characteristic image;An at least width third feature figure is calculated according to the third parameter of second feature image and third layer convolutional layer
Picture;Third feature image is inputted into the 4th layer of pond layer to be calculated and the one-to-one fourth feature figure of third feature image
Picture;An at least width fifth feature image is calculated according to the 5th parameter of fourth feature image and layer 5 pond layer;According to
Fifth feature image confirms the scene Recognition result of every width training image;Multilayer is calculated according to target category and scene Recognition result
The loss value of convolutional neural networks;The convergence of multilayer convolutional neural networks model is confirmed when loss value is less than default loss value;It is right
The test image of acquisition is pre-processed to obtain the test image of several second resolutions, and second resolution is more than described first
Resolution ratio;Input test image tests convergent multilayer convolutional neural networks mould to convergent multilayer convolutional neural networks model
Type;Using the scene type in the multilayer convolutional neural networks Model Identification scene image after test.Wherein loss value is less than pre-
If damage value illustrates that the recognition accuracy of multilayer convolutional neural networks model is higher.
The multilayer convolutional neural networks model of embodiment of the present invention is used for scene Recognition.Wherein, multilayer convolutional Neural net
The first parameter in network model includes fisrt feature matrix and the first bias term, and third parameter includes third feature matrix and third
Bias term.5th parameter includes fifth feature matrix and the 5th bias term.Fisrt feature matrix, third feature matrix and the 5th are special
The number for levying matrix all can be multiple.Multiple eigenmatrixes is used to extract feature in image, with according to feature to image
Classify.More feature is conducive to the classification of image.
Incorporated by reference to Fig. 4, specifically, acquisition first includes a large amount of training images of scene, and training image can derive from micro-
The new medias platforms such as rich, wechat include various common scenes in training image, for example, sky, seashore, meadow, forest, meal
Room etc..May include one or more scenes in every width training image, but there are one home court scape, home courts for palpus in every width training image
Accounting of the scape in training image is big relative to accounting of other scenes in training image.Assuming that every width training image is X,
The home court scape in every width training image is then labeled as target category Y.
Secondly, every width training image X is pre-processed.Wherein, pretreated operation includes all training figures of normalization
As the resolution ratio of X.It is appreciated that the resolution ratio of the training picture X obtained from various channels greatly may be different,
First resolution ratio is normalized the training that can facilitate multilayer convolutional neural networks model before the scene of recognition training image X,
Accelerate the convergence of the training of multilayer convolutional neural networks model.The normalization of resolution ratio be specifically to every width training image X into
Row down-sampling.It is in a specific embodiment of the present invention, unified that the resolution ratio of all training image X is normalized into 64x64,
I other words first resolution 64x64.
Then, the training image X of 64x64 is input to first layer convolutional layer.Fisrt feature square in first layer convolutional layer
Battle array is Wlayer1, wherein Wlayer1Number N1Can be it is multiple, in a specific embodiment of the present invention, fisrt feature matrix Wlayer1
Number N1Value be 32.Each fisrt feature matrix Wlayer1Perception domain size be k1×k1, in the specific implementation of the present invention
In example, k1Value be 3.The fisrt feature image for defining the output of first layer convolutional layer is Flayer1, then Flayer1=δ (X*Wlayer1+
blayer1), whereinFor the first bias term, RN1Reference and N1Relevant real number space;δ (x)=max (x, 0) is sharp
Function living.Fisrt feature matrix Wlayer1Effect be fitting function so that the function can divide the scene in image
Class.Specifically, fisrt feature matrix Wlayer1The feature in training image X, multiple fisrt feature matrix Ws can be extractedlayer1It can carry
Multiple features in training image X are taken, in this way, multiple features can be used for scene classification, promote the accuracy of scene classification.First
Bias term blayer1The similar linear function of effect in intercept, the accuracy of the function category scene of fitting can be promoted.Activate letter
Number δ (x)=max (x, 0) is used to increase the non-linear of the function of fitting, can further promote the essence of the function category scene of fitting
True property.In a specific embodiment of the present invention, training image X and fisrt feature matrix Wlayer1When doing convolution algorithm, the cunning of window
Dynamic step-length is 2.Fisrt feature matrix W in first layer convolutional layerlayer1Number N1It it is 32, then first layer convolutional layer is exportable
32 width fisrt feature image Flayer1, for the i-th width fisrt feature image
Wherein the value range of i is [1,32], and i is positive integer, and the value range of j is [1,32], and j is positive integer, and each first special
Levy image Flayer1Resolution ratio be 31x31.
Then, 32 width fisrt feature image F first layer convolutional layer exportedlayer1Second layer pond layer is input to carry out
Chi Hua.In a specific embodiment of the present invention, using the method in maximum pond to each width fisrt feature image Flayer1Carry out pond
Change operation.Specifically, the kernel function size of second layer pond layer is 3x3, and window sliding step-length is 2, then each width fisrt feature
Image Flayer1Output and each width fisrt feature image F after the operation of second layer pond Hua Ceng pondizationslayer1Corresponding second feature
Image Flayer2, each width second feature image Flayer2Resolution ratio be 15x15.The second of second layer pond layer output is special
Levy image Flayer2Quantity be 32 width.
Then, by 32 width second feature image Flayer2It is input to third layer convolutional layer.Third in third layer convolutional layer is special
Sign matrix is Wlayer3, wherein Wlayer3Number N3Can be it is multiple, in a specific embodiment of the present invention, third feature matrix
Wlayer3Number N3Value be 32.Each third feature matrix Wlayer3Perception domain size be k3×k3, in the specific of the present invention
In embodiment, k3Value be 3.The third feature image for defining the output of third layer convolutional layer is Flayer3, then Flayer3=δ
(Flayer2*Wlayer3+blayer3), whereinFor third bias term,Reference and N3Relevant real number space;δ(x)
=max (x, 0) is activation primitive.Third feature matrix Wlayer3Effect be fitting function so that the function can be in image
Scene classify.Third bias term blayer3The similar linear function of effect in intercept, the function point of fitting can be promoted
The accuracy of class scene.Activation primitive δ (x)=max (x, 0) is used to increase the non-linear of the function of fitting, can further be promoted
The accuracy of the function category scene of fitting.In a specific embodiment of the present invention, second feature image Flayer2With third feature
Matrix Wlayer3When doing convolution algorithm, the sliding step of window is 2.Third feature matrix W in third layer convolutional layerlayer3Number
N3Be 32, then the exportable 32 width third feature image F of first layer convolutional layerlayer3, for the i-th width third feature imageIn other words, each width third of third layer convolutional layer output
Characteristic image Flayer3It is several second feature images Flayer2With the same third feature matrix Wlayer3It is added and obtains after convolution
, wherein the value range of i is [1,32], and i is positive integer, and the value range of j is [1,32], and j is positive integer, each the
Three characteristic image Flayer3Resolution ratio be 7x7.
Then, 32 width third feature image F third layer convolutional layer exportedlayer3The 4th layer of pond layer is input to carry out
Chi Hua.In a specific embodiment of the present invention, using the method in maximum pond to each width third feature image Flayer3Carry out pond
Change operation.Specifically, the kernel function size of the 4th layer of pond layer is 3x3, and window sliding step-length is 2, then each width third feature
Image Flayer3Output and each width third feature image F after the 4th layer of pond layerlayer3Corresponding fourth feature image
Flayer4, each width fourth feature image Flayer4Resolution ratio be 3x3.The fourth feature image of 4th layer of pond layer output
Flayer4Quantity be 32 width.
Then, by 32 width fourth feature image Flayer4It is input to layer 5 convolutional layer.The 5th in layer 5 convolutional layer is special
Sign matrix is Wlayer5, wherein Wlayer5Number N5It can be multiple, each fifth feature matrix Wlayer5A corresponding scene
Classification, in a specific embodiment of the present invention, fifth feature matrix Wlayer5Number N5Value be 10.Each fifth feature square
The perception domain size of battle array is k5×k5, in a specific embodiment of the present invention, k5Value be 3.Define the output of layer 5 convolutional layer
Fifth feature image is Flayer5, then Flayer5=δ (Flayer4*Wlayer5+blayer5), whereinFor the 5th bias term,Reference and N5Relevant real number space;δ (x)=max (x, 0) is activation primitive.Fifth feature matrix Wlayer5Effect be
Fitting function so that the function can classify to the scene in image.5th bias term blayer5Effect similar to linear letter
Intercept in number can promote the accuracy of the function category scene of fitting.Activation primitive δ (x)=max (x, 0) is for increasing
The function of fitting it is non-linear, can further promote the accuracy of the function category scene of fitting.In the specific implementation of the present invention
In example, fourth feature image Flayer4With fifth feature matrix Wlayer5When doing convolution algorithm, the sliding step of window is 1.5th
Fifth feature matrix W in layer convolutional layerlayer5Number N5Be 10, then the exportable 10 width fifth feature figure of layer 5 convolutional layer
As Flayer5, for the i-th width fifth feature image In other words,
Each width fifth feature image F of five layers of convolutional layer outputlayer5It is several fourth feature images Flayer4It is special with the same 5th
Levy matrix Wlayer5Addition obtains after convolution, and wherein the value range of i is [1,32], and i is positive integer, and the value range of j is
[1,32], j are positive integer, each width fifth feature image Flayer5Resolution ratio be 1x1.
Then, from the 10 obtained width fifth feature image Flayer5Middle selection Flayer5It is worth maximum width fifth feature
Image is then used to that F to be calculatedlayer5It is worth the fifth feature matrix W of maximum fifth feature imagelayer5Corresponding classification is
The scene type identified
The identification that scene can be carried out by above-mentioned mode for each width training image X, obtains its unique field
Scape classificationTherefore, the multilayer convolutional neural networks model stated in use acquires the scene class of each width training image X
NotAfterwards, according to target category Y and the classification identifiedCalculate the loss value Loss of multilayer convolutional neural networks model:Wherein, N is the quantity of training image X, and k is kth width training image X, and k is just whole
Number.The convergence of multilayer convolutional neural networks model is confirmed when loss value Loss is less than default loss value.Default loss value characterization is more
Layer convolutional neural networks model is used for identification error rate when scene Recognition.When loss value Loss is less than default loss value, say
The identification error rate of bright multilayer convolutional neural networks Model Identification scene type is relatively low, in other words, multilayer convolutional neural networks mould
Type identifies that the recognition accuracy of scene type is higher.So far, structure and the training of multilayer convolutional neural networks model are completed.
Great amount of images need to be acquired in advance at model construction initial stage, be different between the arbitrary two images in these images
's.These images are according to 4:1 ratio is divided into training image and test image, for example, the image of acquisition has 3000,
In 2400 be used as training image, for training multilayer convolutional neural networks model, 600 images to be used for as test image
Test multilayer convolutional neural networks model.Training image and test image are with 4:1 is relatively mild for ratio, on the one hand can meet
The demand of recognition accuracy after the training of multilayer convolutional neural networks model, while multilayer convolutional neural networks model construction
Time complexity is relatively low.
Therefore, after multilayer convolutional neural networks model training, further, using test image to trained
Multilayer convolutional neural networks model is tested.Specifically, the normalized of resolution ratio first is carried out to obtain to test image
The test image of second resolution can specifically carry out test image by way of down-sampling the normalized of resolution ratio.
Then, the test image of second resolution convergent multilayer convolutional neural networks model is input to roll up to test convergent multilayer
Product neural network model.It can be to avoid multilayer convolution god using the multilayer convolutional neural networks model after test image test convergence
Overfitting through network model.And multilayer convolutional Neural net is trained using the lower first resolution of resolution ratio in the training stage
Network model can reduce the calculation amount in multilayer convolutional neural networks model construction process, higher using resolution ratio in test phase
Second resolution test multilayer convolutional neural networks model, then layer 5 convolutional layer output fifth feature image resolution ratio
It is higher, it can be used for subsequent scene layout's structure.
The image-recognizing method based on multilayer convolutional neural networks model, the pattern recognition device of embodiment of the present invention
100 and computer equipment 1000 build three-layer coil lamination and add the multilayer convolutional neural networks model of two layers of pond layer, and using dividing
Resolution is normalized to the training image training multilayer convolutional neural networks model of first resolution, and the is normalized to using resolution ratio
The test image of two resolution ratio tests multilayer convolutional neural networks model, can be realized to image scene without using full articulamentum
Identification, reduce the complexity of scene Recognition algorithm, the calculation amount of scene Recognition is smaller, calculate take it is shorter.
Referring to Fig. 5, in some embodiments, the image-recognizing method based on multilayer convolutional neural networks model also wraps
It includes:
010:When loss value is greater than or equal to default loss value, the first parameter of modification, third parameter and the 5th parameter;
It is calculated back to step 02 according to the training image of first resolution and the first parameter of first layer convolutional layer
An at least width fisrt feature image.
Referring to Fig. 6, in some embodiments, pattern recognition device 100 further includes modified module 40.Step 010 can
To be realized by modified module 40.In other words, modified module 40 can be used for, when loss value is greater than or equal to default loss value, repairing
Change the first parameter, third parameter and the 5th parameter, and 02 is entered step after parameter is changed.
Referring again to Fig. 3, in some embodiments, when computer-readable instruction 611 is executed by processor 62, also make
Processor 62 is executed when loss value is greater than or equal to default loss value, the first parameter of modification, third parameter and the 5th parameter, with
And entered step after parameter is changed 02 operation.
Specifically, when loss value is greater than or equal to default loss value, the fisrt feature square in first layer convolutional layer is changed
The 5th in third feature matrix and third bias term, layer 5 convolutional layer in battle array and the first bias term, third layer convolutional layer
Eigenmatrix and the 5th bias term.
It is appreciated that when loss value is larger, illustrate the accurate of the multilayer convolutional neural networks Model Identification image scene
Rate is relatively low, therefore, it should change each layer of eigenmatrix and bias term so that each layer of eigenmatrix and bias term is formed
Fitting function can more accurately identify image scene, further promote entire multilayer convolutional neural networks Model Identification image
The accuracy rate of scene.
Referring to Fig. 7, in some embodiments, the image-recognizing method based on multilayer convolutional neural networks model is in step
Further include after rapid 013:
014:Pair fifth feature image corresponding with scene type carries out expansion or corrosion treatment to obtain the profile of scene.
Referring to Fig. 8, in some embodiments, pattern recognition device 100 further includes processing module 44.Step 014 can
To be realized by processing module 44.In other words, processing module 44 can be used for pair fifth feature image corresponding with scene type into
Row expansion or corrosion treatment are to obtain the profile of scene.
Referring again to Fig. 3, in some embodiments, when computer-readable instruction 611 is executed by processor 62, also make
Processor 62 executes a pair fifth feature image corresponding with scene type and carries out expansion or corrosion treatment to obtain the profile of scene
Operation.
Specifically, after multilayer convolutional neural networks model training, a width images to be recognized, images to be recognized are inputted
Down-sampling can first be done to reduce resolution ratio, but resolution ratio should be greater than 64x64, in this way, may make the fifth feature figure finally selected
The resolution ratio of picture will not be too small, is convenient for the acquisition of scene profile.Input images to be recognized after, will layer 5 convolutional layer export with
The corresponding 10 width fifth feature image of images to be recognized, F is selected from fifth feature imagelayer5Maximum fifth feature image,
The classification that the corresponding eigenmatrix of width fifth feature image is referred to is the classification of the scene recognized.Then, according to choosing
The F gone outlayer5The pixel of maximum 5th width characteristic image determines region corresponding with each pixel in images to be recognized,
And expansion and corrosion treatment are carried out to obtain the profile of scene to these regions.
The present invention also provides the non-volatile computer readable storage mediums that one or more includes computer executable instructions
Matter, when computer executable instructions are executed by one or more processors 62 so that processor 62 executes above-mentioned any one
The construction method of multilayer convolutional neural networks model described in embodiment.
For example, when computer executable instructions are executed by one or more processors 62 so that processor 62 execute with
The operation of lower step:
00:Target category marked to every width training image gathered in advance, and to every width training image pre-processed with
Obtain the training image of several first resolutions;
01:The initial configuration of multilayer convolutional neural networks model is set, initial configuration is tactic first layer convolution
Layer, second layer pond layer, third layer convolutional layer, the 4th layer of pond layer and layer 5 convolutional layer;
02:It is calculated at least one the according to the first parameter of the training image of first resolution and first layer convolutional layer
One characteristic image;
03:Fisrt feature image is inputted into second layer pond layer to be calculated and fisrt feature image one-to-one the
Two characteristic images;
04:An at least width third feature figure is calculated according to the third parameter of second feature image and third layer convolutional layer
Picture;
05:Third feature image is inputted the 4th layer of pond layer to correspond with third feature image to be calculated
Fourth feature image;
06:An at least width fifth feature figure is calculated according to the 5th parameter of fourth feature image and layer 5 pond layer
Picture;
07:The scene Recognition result of every width training image is confirmed according to fifth feature image;
08:The loss value of multilayer convolutional neural networks is calculated according to target category and scene Recognition result;
09:The convergence of multilayer convolutional neural networks model is confirmed when loss value is less than default loss value;
011:The test image of acquisition is pre-processed to obtain the test image of several second resolutions, second differentiates
Rate is more than the first resolution;
012:Input test image tests convergent multilayer convolutional Neural to convergent multilayer convolutional neural networks model
Network model;With
013:Using the scene type in the multilayer convolutional neural networks Model Identification scene image after test.
For another example when computer executable instructions are executed by one or more processors 62 so that processor 62 executes
The operation of following steps:
014:Pair fifth feature image corresponding with scene type carries out expansion or corrosion treatment to obtain the profile of scene.
Referring to Fig. 9, the present invention also provides a kind of image acquiring methods.Image acquiring method includes:
21:Obtain scene image;
22:Using in the multilayer convolutional neural networks Model Identification scene image described in above-mentioned any one embodiment
Scene type;With
23:According to the acquisition parameters of scene type adjustment camera to obtain new scene image corresponding with scene image,
Acquisition parameters include at least one of colour temperature, time for exposure, sensitivity and exposure compensating.
Referring to Fig. 10, the present invention also provides a kind of image acquisition equipments 200.The image acquisition side of embodiment of the present invention
Method can be realized by the image acquisition equipment 200 of embodiment of the present invention.Image acquisition equipment 200 includes acquiring unit 50 and figure
As identification device 100.Step 21 and step 23 can be realized by acquiring unit 50.Step 22 can be by pattern recognition device
100 realize.In other words, acquiring unit 50 can be used for obtaining scene image.Pattern recognition device 100 can be used for using above-mentioned
The scene type in multilayer convolutional neural networks Model Identification scene image described in an embodiment of anticipating.Acquiring unit 50 is also
It can be used for adjusting the acquisition parameters of camera according to scene type to obtain new scene image corresponding with scene image, shooting ginseng
Number includes at least one of colour temperature, time for exposure, sensitivity and exposure compensating.
Referring again to Fig. 3, when computer-readable instruction 611 is executed by processor 62, also so that the execution of processor 62 obtains
Scene image is taken, using in the multilayer convolutional neural networks Model Identification scene image described in above-mentioned any one embodiment
Scene type, and the acquisition parameters of camera are adjusted to obtain new scene image corresponding with scene image according to scene type
Operation.Wherein, scene image is shot by camera 81 (shown in Figure 12), and processor 62 is connect with camera 81 to read this
Scape image.
Acquisition parameters include that at least one of colour temperature, time for exposure, sensitivity and exposure compensating refer to:Acquisition parameters
Can only include colour temperature, time for exposure, sensitivity or exposure compensating.Acquisition parameters also can include colour temperature and time for exposure simultaneously,
Or simultaneously including colour temperature, time for exposure and sensitivity, or include colour temperature, time for exposure, sensitivity and exposure compensating etc. simultaneously.
Specifically, for example, after the one width scene image of shooting of camera 81, processor 62, which executes, uses multilayer convolutional Neural net
The operation of scene type in network Model Identification scene image, if identifying, scene is seashore, since the scene of seashore usually has
More strong sunlight then can suitably reduce the time for exposure to prevent the image overexposure taken, and with the shorter time for exposure
Shoot new scene image etc..
In this way, using the multilayer convolutional neural networks Model Identification scene image of embodiment of the present invention, can must show up
The classification of scape image, the profile of also extractable scene image.Further, the shooting according to the scene identified to camera 81
Parameter is adjusted, and can be improved the quality of the new scene image after shooting, be promoted the usage experience of user.
In addition, in some embodiments, after acquisition finishes new scene image, due to the scene in new scene image
Know, new scene image can also be further processed according to known scene, for example, the scene in new scene image is sea
Bank then matches the audio of the preceding paragraph wave for new scene image, during user browses the new scene image stored, in real time
Play the audio of wave;For another example new scene image Scene is forest, then contain tweedle with the preceding paragraph for new scene image
Audio, user browse stored new scene image during, play audio of this section containing tweedle in real time, such as
This, increases audio-frequency information corresponding with scene to new scene image according to scene, promotes the shooting entertaining of user, improve user's
Usage experience.
The present invention also provides the non-volatile computer readable storage mediums that one or more includes computer executable instructions
Matter, when computer executable instructions are executed by one or more processors 62 so that the execution of processor 62 is obtained with hypograph
Method:
For example, when computer executable instructions are executed by one or more processors 62 so that processor 62 execute with
The operation of lower step:
21:Obtain scene image;
22:Using in the multilayer convolutional neural networks Model Identification scene image described in above-mentioned any one embodiment
Scene type;With
23:According to the acquisition parameters of scene type adjustment camera to obtain new scene image corresponding with scene image,
Acquisition parameters include at least one of colour temperature, time for exposure, sensitivity and exposure compensating.
Figure 11 is the internal module schematic diagram of the computer equipment 1000 in one embodiment.As shown in figure 11, the calculating
Machine equipment 1000 include the processor 62 connected by system bus 66, memory 61 (being, for example, non-volatile memory medium),
Built-in storage 63, display screen 65 and input unit 64.Wherein, the memory 61 of computer equipment 100 be stored with operating system and
Computer-readable instruction 611 (shown in Fig. 3).The computer-readable instruction 611 can be executed by processor 62, above-mentioned arbitrary to realize
Image-recognizing method based on multilayer convolutional neural networks model and above-mentioned any one described in one embodiment are implemented
Image acquiring method described in mode.The processor 62 can be used for providing calculating and control ability, support entire computer equipment
1000 operation.The built-in storage 63 of computer equipment 1000 provides for the operation of computer-readable instruction 611 in memory 61
Environment.The display screen 65 of computer equipment 1000 can be liquid crystal display or electric ink display screen etc., input unit 64
Can be the touch layer covered on display screen 65, can also be the case being arranged on 1000 shell of computer equipment, trace ball or
Trackpad can also be external keyboard, Trackpad or mouse etc..The computer equipment 1000 can be mobile phone, tablet computer,
Laptop, personal digital assistant or Wearable (such as Intelligent bracelet, smartwatch, intelligent helmet, intelligent glasses)
Deng.It will be understood by those skilled in the art that structure shown in Figure 11, only with the relevant part-structure of the present invention program
Schematic diagram, does not constitute the restriction for the computer equipment 1000 being applied thereon to the present invention program, and specific computer is set
Standby 1000 may include either combining certain components or with different components than more or fewer components as shown in the figure
Arrangement.
2 are please referred to Fig.1, the computer equipment 1000 of the embodiment of the present invention includes Image Processor Circuits 80.Image procossing
Circuit 80 can utilize hardware and/or software realization.It may include defining ISP (Image Signal Processing, image letter
Number processing) pipeline various processing units.Figure 12 is the schematic diagram of image processing circuit 80 in one embodiment.Such as Figure 12 institutes
Show, for purposes of illustration only, only showing the various aspects with the relevant image processing techniques of the embodiment of the present invention.
As shown in figure 12, image processing circuit includes that (ISP processors can be processor 62 or processor 62 to ISP processors
A part) and control logic device 84.The image data that camera 81 captures is handled by ISP processors 83 first, ISP processing
Device 83 is analyzed image data to capture the image statistics for the one or more control parameters that can be used for determining camera 81
Information.Camera 81 may include lens 811 and imaging sensor 812.Imaging sensor 812 can obtain each imaging pixel and capture
Luminous intensity and wavelength information, and provide one group of raw image data being handled by ISP processors 83.82 (such as top of sensor
Spiral shell instrument) parameter (such as stabilization parameter) of the image procossing of acquisition can be supplied to ISP processors based on sensor interface type
83.Sensor interface can be SMIA (Standard Mobile Imaging Architecture, Standard Mobile Imager frame
Structure) interface, other serial or parallel camera interfaces or above-mentioned interface combination.
In addition, raw image data can be also sent to sensor 82 by imaging sensor 812, sensor 82 can be based on sensing
Device interface type is supplied to ISP processors 83 or sensor by raw image data storage to storage raw image data
In device 61.
ISP processors 83 handle raw image data pixel by pixel in various formats.For example, each image pixel can have
There are the bit depth of 8,10,12 or 14 bits, ISP processors 83 that can carry out one or more image procossings to raw image data
The statistical information of operation, collection about image data.Wherein, image processing operations can by identical or different bit depth precision into
Row.
ISP processors 83 can also receive image data from memory 61.For example, sensor interface sends out raw image data
Memory 61 is given, it is for processing that the raw image data in memory 61 is available to ISP processors 83.
When receiving the original image from image sensor interface or from 82 interface of sensor or from memory 61
When data, ISP processors 83 can carry out one or more image processing operations, such as time-domain filtering.Treated, and image data can
It is sent to memory 61, to carry out other processing before shown.ISP processors 83 receive processing number from memory 61
According to, and image real time transfer is carried out to the processing data.Treated that image data may be output to display for ISP processors 83
Screen, so that user watches and/or by graphics engine or GPU (Graphics Processing Unit, graphics processor) into one
Step processing.In addition, the output of ISP processors 83 also can be transmitted to memory 61, and display screen 65 can read from memory 61 and scheme
As data.In one embodiment, memory 61 can be configured as realizing one or more frame buffers.In addition, ISP processing
The output of device 83 can be transmitted to encoder/decoder 85, so as to encoding/decoding image data.The image data of coding can be protected
It deposits, and is decompressed before being shown on display screen 65.Encoder/decoder 85 can be realized by CPU or GPU or coprocessor.
The statistical data that ISP processors 83 determine can be transmitted to control logic device unit 84.For example, statistical data may include
The imaging sensors statistical information such as automatic exposure, automatic focusing, flicker detection, black level compensation, correcting lens shadow.Control is patrolled
It collects device 84 and may include the processing element and/or microcontroller that execute one or more routines (such as firmware), one or more routines
According to the statistical data of reception, the control parameter of camera 81 and the control parameter of ISP processor 83 can be determined.For example, camera shooting
First 81 control parameter may include 82 control parameter of sensor (such as gain, the time of integration of spectrum assignment, stabilization parameter etc.),
The combination of camera flash control parameter, lens control parameter (such as focusing or zoom focal length) or these parameters.
One of ordinary skill in the art will appreciate that realizing all or part of flow in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in that a non-volatile computer is readable to be deposited
In storage media, the program is when being executed, it may include such as the flow of the embodiment of above-mentioned each method.Wherein, the storage medium
Can be magnetic disc, CD, read-only memory (Read-Only Memory, ROM) etc..
The several embodiments of the application above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously
Cannot the limitation to the application the scope of the claims therefore be interpreted as.It should be pointed out that for those of ordinary skill in the art
For, under the premise of not departing from the application design, various modifications and improvements can be made, these belong to the guarantor of the application
Protect range.Therefore, the protection domain of the application patent should be determined by the appended claims.
Claims (16)
1. a kind of image-recognizing method based on multilayer convolutional neural networks model, which is characterized in that described image recognition methods
Including:
Target category is marked to every width training image gathered in advance, and training image described in every width is pre-processed to obtain
The training image of several first resolutions;
The initial configuration of the multilayer convolutional neural networks model is set, the initial configuration is tactic first layer convolution
Layer, second layer pond layer, third layer convolutional layer, the 4th layer of pond layer and layer 5 convolutional layer;
It is calculated at least one first according to the first parameter of the training image of first resolution and first layer convolutional layer
Characteristic image;
It is a pair of with the fisrt feature image one to be calculated that the fisrt feature image is inputted second layer pond layer
The second feature image answered;
An at least width third feature is calculated according to the third parameter of the second feature image and the third layer convolutional layer
Image;
It is a pair of with the third feature image one to be calculated that the third feature image is inputted the 4th layer of pond layer
The fourth feature image answered;
An at least width fifth feature is calculated according to the 5th parameter of the fourth feature image and layer 5 pond layer
Image;
The scene Recognition result of training image described in every width is confirmed according to the fifth feature image;
The loss value of the multilayer convolutional neural networks is calculated according to the target category and the scene Recognition result;
The multilayer convolutional neural networks model convergence is confirmed when the loss value is less than default loss value;
The test image of acquisition is pre-processed to obtain the test image of several second resolutions, described second differentiates
Rate is more than the first resolution;
The test image is inputted to the convergent multilayer convolutional neural networks model to test the convergent multilayer convolution
Neural network model;With
Using the scene type in the multilayer convolutional neural networks Model Identification scene image after test.
2. image-recognizing method according to claim 1, which is characterized in that the training image includes several, the survey
Attempt as including several, the ratio of the number of the number of the training image and the test image is 4:1;
It is described to the training image carry out pretreatment include that the training image is normalized;
It includes that the test image is normalized that the test image of described pair of acquisition, which carries out pretreatment,.
3. image-recognizing method according to claim 1, which is characterized in that first parameter includes fisrt feature matrix
With the first bias term;The third parameter includes third feature matrix and third bias term;5th parameter includes the 5th special
Levy matrix and the 5th bias term.
4. image-recognizing method according to claim 3, which is characterized in that the number of the fisrt feature matrix is 32
A, the size in the perception domain of each fisrt feature matrix is 3 × 3;And/or
The number of the third feature matrix is 32, and the size in the perception domain of each third feature matrix is 3 × 3;With/
Or
The number of the fifth feature matrix is 10, and the size in the perception domain of each fifth feature matrix is 3 × 3.
5. image-recognizing method according to claim 3, which is characterized in that described image recognition methods further includes:
When the loss value is greater than or equal to the default loss value, first parameter, the third parameter and institute are changed
State the 5th parameter;With
Back to the training image according to first resolution and first layer convolutional layer the first parameter be calculated to
The step of few width fisrt feature image.
6. image-recognizing method according to claim 1, which is characterized in that described image recognition methods is described using survey
Further include after the step of scene type in the multilayer convolutional neural networks Model Identification scene image after examination:
Pair fifth feature image corresponding with the scene type carries out expansion or corrosion treatment to obtain the scene
Profile.
7. a kind of image acquiring method, which is characterized in that described image acquisition methods include:
Obtain scene image;
Using the field in scene image described in the multilayer convolutional neural networks Model Identification described in claim 1 to 6 any one
Scape classification;
According to the acquisition parameters of scene type adjustment camera to obtain new scene image corresponding with the scene image,
The acquisition parameters include at least one of colour temperature, time for exposure, sensitivity and exposure compensating.
8. a kind of pattern recognition device based on multilayer convolutional neural networks model, which is characterized in that described image identification device
Including:
First preprocessing module, first preprocessing module are used to mark target class to every width training image gathered in advance
Not, and to training image described in every width it is pre-processed to obtain the training image of several first resolutions;
Setting module, the setting module is used to set the initial configuration of the multilayer convolutional neural networks model, described initial
Structure is tactic first layer convolutional layer, second layer pond layer, third layer convolutional layer, the 4th layer of pond layer and layer 5 volume
Lamination;
First computing module, first computing module are used for the training image and first layer convolution according to first resolution
An at least width fisrt feature image is calculated in first parameter of layer;
Second computing module, second computing module be used for by the fisrt feature image input second layer pond layer with
It is calculated and the one-to-one second feature image of the fisrt feature image;
Third computing module, the third computing module are used for according to the second feature image and the third layer convolutional layer
An at least width third feature image is calculated in third parameter;
4th computing module, the 4th computing module be used for by the third feature image input the 4th layer of pond layer with
It is calculated and the one-to-one fourth feature image of the third feature image;
5th computing module, the 5th computing module are used for according to the fourth feature image and layer 5 pond layer
An at least width fifth feature image is calculated in 5th parameter;
First confirmation module, first confirmation module are used to confirm training image described in every width according to the fifth feature image
Scene Recognition result;
6th computing module, the 6th computing module are used to calculate institute according to the target category and the scene Recognition result
State the loss value of multilayer convolutional neural networks;
Second confirmation module, second confirmation module are used to confirm the multilayer when the loss value is less than default loss value
Convolutional neural networks model is restrained;
Second preprocessing module, second preprocessing module is for pre-processing to obtain several test image of acquisition
The test image of second resolution, the second resolution are more than the first resolution;
Test module, the test module is for inputting the test image to the convergent multilayer convolutional neural networks model
To test the convergent multilayer convolutional neural networks model;With
Identification module, the identification module are used for using the multilayer convolutional neural networks Model Identification scene image after test
In scene type.
9. pattern recognition device according to claim 8, which is characterized in that the training image includes several, the survey
Attempt as including several, the ratio of the number of the number of the training image and the test image is 4:1;
It is described to the training image carry out pretreatment include that the training image is normalized;
It includes that the test image is normalized that the test image of described pair of acquisition, which carries out pretreatment,.
10. pattern recognition device according to claim 8, which is characterized in that first parameter includes fisrt feature square
Battle array and the first bias term;The third parameter includes third feature matrix and third bias term;5th parameter includes the 5th
Eigenmatrix and the 5th bias term.
11. pattern recognition device according to claim 10, which is characterized in that the number of the fisrt feature matrix is 32
A, the size in the perception domain of each fisrt feature matrix is 3 × 3;And/or
The number of the third feature matrix is 32, and the size in the perception domain of each third feature matrix is 3 × 3;With/
Or
The number of the fifth feature matrix is 10, and the size in the perception domain of each fifth feature matrix is 3 × 3.
12. pattern recognition device according to claim 10, which is characterized in that described image identification device further includes modification
Module, the modified module are used for:
When the loss value is greater than or equal to the default loss value, first parameter, the third parameter and institute are changed
State the 5th parameter;With
Back to the training image according to first resolution and first layer convolutional layer the first parameter be calculated to
The step of few width fisrt feature image.
13. pattern recognition device according to claim 7, which is characterized in that described image identification device further includes:
Processing module, the processing module for pair fifth feature image corresponding with the scene type carry out expansion or
Corrosion treatment is to obtain the profile of the scene.
14. a kind of image acquisition equipment, which is characterized in that described image obtains equipment and includes:
Acquiring unit, the acquiring unit is for obtaining scene image;
Pattern recognition device, described image identification device are used for using the multilayer convolution god described in claim 1 to 6 any one
The scene type in the scene image is identified through network model;
The acquiring unit is additionally operable to adjust the acquisition parameters of camera according to the scene type to obtain and the scene graph
As corresponding new scene image, the acquisition parameters include at least one in colour temperature, time for exposure, sensitivity and exposure compensating
Kind.
15. a kind of computer equipment, including memory and processor, it is stored with computer-readable instruction in the memory, institute
When stating computer-readable instruction and being executed by the processor so that the processor perform claim requires described in 1 to 6 any one
Image recognition method and claim 7 described in image acquiring method.
16. one or more includes the non-volatile computer readable storage medium storing program for executing of computer executable instructions, when the calculating
When machine executable instruction is executed by one or more processors so that the processor perform claim requires 1 to 6 any one institute
The image acquiring method described in image-recognizing method and claim 7 stated.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810151420.8A CN108304821B (en) | 2018-02-14 | 2018-02-14 | Image recognition method and device, image acquisition method and device, computer device and non-volatile computer-readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810151420.8A CN108304821B (en) | 2018-02-14 | 2018-02-14 | Image recognition method and device, image acquisition method and device, computer device and non-volatile computer-readable storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108304821A true CN108304821A (en) | 2018-07-20 |
CN108304821B CN108304821B (en) | 2020-12-18 |
Family
ID=62865152
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810151420.8A Expired - Fee Related CN108304821B (en) | 2018-02-14 | 2018-02-14 | Image recognition method and device, image acquisition method and device, computer device and non-volatile computer-readable storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108304821B (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109101931A (en) * | 2018-08-20 | 2018-12-28 | Oppo广东移动通信有限公司 | A kind of scene recognition method, scene Recognition device and terminal device |
CN109166141A (en) * | 2018-08-10 | 2019-01-08 | Oppo广东移动通信有限公司 | Dangerous based reminding method, device, storage medium and mobile terminal |
CN109272442A (en) * | 2018-09-27 | 2019-01-25 | 百度在线网络技术(北京)有限公司 | Processing method, device, equipment and the storage medium of panorama spherical surface image |
CN109344715A (en) * | 2018-08-31 | 2019-02-15 | 北京达佳互联信息技术有限公司 | Intelligent composition control method, device, electronic equipment and storage medium |
CN109525859A (en) * | 2018-10-10 | 2019-03-26 | 腾讯科技(深圳)有限公司 | Model training, image transmission, image processing method and relevant apparatus equipment |
CN109583445A (en) * | 2018-11-26 | 2019-04-05 | 平安科技(深圳)有限公司 | Character image correction processing method, device, equipment and storage medium |
CN110232313A (en) * | 2019-04-28 | 2019-09-13 | 南京览视医疗科技有限公司 | A kind of eye recommended method, system, electronic equipment and storage medium |
CN110288030A (en) * | 2019-06-27 | 2019-09-27 | 重庆大学 | Image-recognizing method, device and equipment based on lightweight network model |
CN110348291A (en) * | 2019-05-28 | 2019-10-18 | 华为技术有限公司 | A kind of scene recognition method, a kind of scene Recognition device and a kind of electronic equipment |
CN111027489A (en) * | 2019-12-12 | 2020-04-17 | Oppo广东移动通信有限公司 | Image processing method, terminal and storage medium |
CN111091593A (en) * | 2018-10-24 | 2020-05-01 | 深圳云天励飞技术有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
WO2020098360A1 (en) * | 2018-11-15 | 2020-05-22 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for processing images using cross-stage skip connections |
CN111597476A (en) * | 2020-05-06 | 2020-08-28 | 北京金山云网络技术有限公司 | Image processing method and device |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106845549A (en) * | 2017-01-22 | 2017-06-13 | 珠海习悦信息技术有限公司 | A kind of method and device of the scene based on multi-task learning and target identification |
CN107464216A (en) * | 2017-08-03 | 2017-12-12 | 济南大学 | A kind of medical image ultra-resolution ratio reconstructing method based on multilayer convolutional neural networks |
-
2018
- 2018-02-14 CN CN201810151420.8A patent/CN108304821B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106845549A (en) * | 2017-01-22 | 2017-06-13 | 珠海习悦信息技术有限公司 | A kind of method and device of the scene based on multi-task learning and target identification |
CN107464216A (en) * | 2017-08-03 | 2017-12-12 | 济南大学 | A kind of medical image ultra-resolution ratio reconstructing method based on multilayer convolutional neural networks |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109166141A (en) * | 2018-08-10 | 2019-01-08 | Oppo广东移动通信有限公司 | Dangerous based reminding method, device, storage medium and mobile terminal |
CN109101931A (en) * | 2018-08-20 | 2018-12-28 | Oppo广东移动通信有限公司 | A kind of scene recognition method, scene Recognition device and terminal device |
CN109344715A (en) * | 2018-08-31 | 2019-02-15 | 北京达佳互联信息技术有限公司 | Intelligent composition control method, device, electronic equipment and storage medium |
CN109272442A (en) * | 2018-09-27 | 2019-01-25 | 百度在线网络技术(北京)有限公司 | Processing method, device, equipment and the storage medium of panorama spherical surface image |
CN109525859A (en) * | 2018-10-10 | 2019-03-26 | 腾讯科技(深圳)有限公司 | Model training, image transmission, image processing method and relevant apparatus equipment |
CN109525859B (en) * | 2018-10-10 | 2021-01-15 | 腾讯科技(深圳)有限公司 | Model training method, image sending method, image processing method and related device equipment |
CN111091593B (en) * | 2018-10-24 | 2024-03-22 | 深圳云天励飞技术有限公司 | Image processing method, device, electronic equipment and storage medium |
CN111091593A (en) * | 2018-10-24 | 2020-05-01 | 深圳云天励飞技术有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
WO2020098360A1 (en) * | 2018-11-15 | 2020-05-22 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for processing images using cross-stage skip connections |
CN109583445A (en) * | 2018-11-26 | 2019-04-05 | 平安科技(深圳)有限公司 | Character image correction processing method, device, equipment and storage medium |
CN110232313A (en) * | 2019-04-28 | 2019-09-13 | 南京览视医疗科技有限公司 | A kind of eye recommended method, system, electronic equipment and storage medium |
CN110348291A (en) * | 2019-05-28 | 2019-10-18 | 华为技术有限公司 | A kind of scene recognition method, a kind of scene Recognition device and a kind of electronic equipment |
CN110288030B (en) * | 2019-06-27 | 2023-04-07 | 重庆大学 | Image identification method, device and equipment based on lightweight network model |
CN110288030A (en) * | 2019-06-27 | 2019-09-27 | 重庆大学 | Image-recognizing method, device and equipment based on lightweight network model |
CN111027489A (en) * | 2019-12-12 | 2020-04-17 | Oppo广东移动通信有限公司 | Image processing method, terminal and storage medium |
CN111027489B (en) * | 2019-12-12 | 2023-10-20 | Oppo广东移动通信有限公司 | Image processing method, terminal and storage medium |
CN111597476A (en) * | 2020-05-06 | 2020-08-28 | 北京金山云网络技术有限公司 | Image processing method and device |
CN111597476B (en) * | 2020-05-06 | 2023-08-22 | 北京金山云网络技术有限公司 | Image processing method and device |
Also Published As
Publication number | Publication date |
---|---|
CN108304821B (en) | 2020-12-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108304821A (en) | Image-recognizing method and device, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing | |
Zhang et al. | Single image defogging based on multi-channel convolutional MSRCR | |
CN105323497B (en) | The high dynamic range (cHDR) of constant encirclement operates | |
CN108604369B (en) | Method, device and equipment for removing image noise and convolutional neural network | |
CN108537155B (en) | Image processing method, image processing device, electronic equipment and computer readable storage medium | |
CN108777815A (en) | Method for processing video frequency and device, electronic equipment, computer readable storage medium | |
CN110428366A (en) | Image processing method and device, electronic equipment, computer readable storage medium | |
CN110276767A (en) | Image processing method and device, electronic equipment, computer readable storage medium | |
CN110473185A (en) | Image processing method and device, electronic equipment, computer readable storage medium | |
CN108419028A (en) | Image processing method, device, computer readable storage medium and electronic equipment | |
CN110290289A (en) | Image denoising method, device, electronic equipment and storage medium | |
CN107493432A (en) | Image processing method, device, mobile terminal and computer-readable recording medium | |
CN107958235A (en) | A kind of facial image detection method, device, medium and electronic equipment | |
CN110191291A (en) | Image processing method and device based on multiple image | |
CN108875619A (en) | Method for processing video frequency and device, electronic equipment, computer readable storage medium | |
CN109002843A (en) | Image processing method and device, electronic equipment, computer readable storage medium | |
CN108897786A (en) | Recommended method, device, storage medium and the mobile terminal of application program | |
CN107509044A (en) | Image combining method, device, computer-readable recording medium and computer equipment | |
CN109089046A (en) | Image denoising method, device, computer readable storage medium and electronic equipment | |
CN108875820A (en) | Information processing method and device, electronic equipment, computer readable storage medium | |
CN107172354A (en) | Method for processing video frequency, device, electronic equipment and storage medium | |
Yan et al. | High dynamic range imaging via gradient-aware context aggregation network | |
CN108848306A (en) | Image processing method and device, electronic equipment, computer readable storage medium | |
CN107563979A (en) | Image processing method, device, computer-readable recording medium and computer equipment | |
CN109005343A (en) | Control method, device, imaging device, electronic equipment and readable storage medium storing program for executing |
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 | ||
CB02 | Change of applicant information |
Address after: Changan town in Guangdong province Dongguan 523860 usha Beach Road No. 18 Applicant after: GUANGDONG OPPO MOBILE TELECOMMUNICATIONS Corp.,Ltd. Address before: Changan town in Guangdong province Dongguan 523860 usha Beach Road No. 18 Applicant before: GUANGDONG OPPO MOBILE TELECOMMUNICATIONS Corp.,Ltd. |
|
CB02 | Change of applicant information | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20201218 |
|
CF01 | Termination of patent right due to non-payment of annual fee |