CN108038540A - A kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network - Google Patents
A kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network Download PDFInfo
- Publication number
- CN108038540A CN108038540A CN201711093435.5A CN201711093435A CN108038540A CN 108038540 A CN108038540 A CN 108038540A CN 201711093435 A CN201711093435 A CN 201711093435A CN 108038540 A CN108038540 A CN 108038540A
- Authority
- CN
- China
- Prior art keywords
- image
- layer
- neutral net
- marker
- convolutional layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
-
- 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
Abstract
The invention discloses a kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network.The multiple dimensioned neutral net includes：First order neutral net, first order neutral net are used for the coordinate information and object global characteristics that the image-region of whole object is obtained based on the input picture including at least whole object；Second level neutral net, the object global image that the coordinate information that second level neutral net is used for the image-region based on the whole object according to acquired in first order neutral net is reduced obtain the coordinate information and object part feature of the image-region of object part；Third level neutral net, the global image of object part that the coordinate information that third level neutral net is used for the image-region based on the object part according to acquired in the neutral net of the second level is reduced obtains on object part or the coordinate information and marker feature of the image-region of the marker on its periphery.The more object features that can be used in object identification or retrieval can be extracted.
Description
Technical field
The present invention relates to neutral net and image characteristics extraction field.A kind of more particularly it relates to multiple dimensioned god
Image characteristic extracting method through network and based on the network.
Background technology
In recent years, people have been combined nerual network technique and image processing techniques for pedestrian, object be identified into
Substantial amounts of research work is gone.In terms of object identification, current research is concentrated mainly on species or subdivision kind for object
Class is identified.For example, the species of object is identified including identifying vehicle, aircraft, train, apple from different objects
The specific type of items such as fruit.For example, when object to be identified is vehicle, to belonging to vehicle type (including car, truck,
Car etc.) it is identified.
When people want to identify some specific vehicle (not being type of vehicle), usually all use be based on Car license recognition at present
Method.However, it is contemplated that there are vehicle is unlicensed, shielding automobile number plate, can not obtain number plate of vehicle image or number plate of vehicle
The actual conditions that image is not known so that the method based on Car license recognition vehicle can not or be difficult to apply.
Therefore, at least need to propose the neutral net and method for extracting more vehicle characteristics (being not limited to license plate number), with
The identification or retrieval of vehicle are preferably carried out based on these features.
The content of the invention
The purpose of the present invention is what is be achieved through the following technical solutions.
Multiple dimensioned neutral net according to the present invention, including：
First order neutral net, the first order neutral net are used to obtain based on the input picture including at least whole object
Take the coordinate information and object global characteristics of the image-region of whole object, and including：
First global convolutional layer and the first global pool layer, the described first global convolutional layer and the first global pool layer use
The network of first convolutional layer of VGG16 neutral nets and the first maximum pond layer；
Second global convolutional layer and the second global pool layer, the described second global convolutional layer and the second global pool layer use
The network of second convolutional layer of VGG16 neutral nets and the second maximum pond layer；
3rd global convolutional layer and the 3rd global pool layer, the described 3rd global convolutional layer and the 3rd global pool layer use
The network of 3rd convolutional layer of VGG16 neutral nets and the 3rd maximum pond layer；
4th global convolutional layer and the 4th global pool layer, the described 4th global convolutional layer and the 4th global pool layer use
The network of the Volume Four lamination of VGG16 neutral nets and the 4th maximum pond layer；
Second level neutral net, the second level neutral net are used for based on whole according to acquired in first order neutral net
The object global image that the coordinate information of the image-region of a object is reduced obtains the coordinate of the image-region of object part
Information and object part feature, and including：
First component convolutional layer and first component pond layer, the first component convolutional layer and first component pond layer use
The network of second convolutional layer of VGG16 neutral nets and the second maximum pond layer；
Second component convolutional layer and second component pond layer, the second component convolutional layer and second component pond layer use
The network of 3rd convolutional layer of VGG16 neutral nets and the 3rd maximum pond layer；
Third member convolutional layer and third member pond layer, the third member convolutional layer and third member pond layer use
The network of the Volume Four lamination of VGG16 neutral nets and the 4th maximum pond layer；
Third level neutral net, the third level neutral net are used for based on the thing according to acquired in the neutral net of the second level
The global image for the object part that the coordinate information of the image-region of body component is reduced obtains the marker on object part
Image-region coordinate information and marker feature, and including：
First marker convolutional layer and the first marker pond layer, the first marker convolutional layer and the first marker pond
Change layer using the 3rd convolutional layer of VGG16 neutral nets and the network of the 3rd maximum pond layer；
Second marker convolutional layer and the second marker pond layer, the second marker convolutional layer and the second marker pond
Change layer using the Volume Four lamination of VGG16 neutral nets and the network of the 4th maximum pond layer；
3rd marker convolutional layer and the 3rd marker pond layer, the 3rd marker convolutional layer and the 3rd marker pond
Change layer using the 5th convolutional layer of VGG16 neutral nets and the network of the 5th maximum pond layer.
Multiple dimensioned neutral net according to the present invention, when the object is vehicle, the object part includes at least down
At least one of row：Car plate, logo, car light, tire, front windshield, rear windshield, drive window, copilot window, rear passenger window,
Rearview mirror, skylight.
Multiple dimensioned neutral net according to the present invention, when the object part is front windshield, on the object part or
The marker on its periphery includes at least at least one of following：Annual test mark, pass, paper-extracting box, pendant, goods of furniture for display rather than for use.
Image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, including：
Step 1：Structure is according to multiple dimensioned neutral net described above；
Step 2：Multiple dimensioned neutral net is trained；
Step 3：Coordinate information and the image spy of image-regions at different levels are obtained using trained multiple dimensioned neutral net
Sign.
Image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, the step 2 include：
Step 2-1：Sample image is selected, sample image is demarcated, builds proven sample image storehouse；
Step 2-2：Using proven sample image storehouse to first order neutral net, second level neutral net and the third level
Neutral net is trained step by step.
Image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, the step 3 include：
Step 3-1：To be training including at least the size scaling of the input picture of the pending feature extraction of whole object
When the used input picture including at least whole object size；
Step 3-2：Successively using the first order neutral net in the trained multiple dimensioned neutral net, the second level
Neutral net and third level neutral net obtain the coordinate information and characteristics of image of image-regions at different levels.
Image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, the step 3 further include：
Step 3-3：To neutral nets at different levels, each acquired multiple images region carries out non-maxima suppression, obtains warp
Cross the coordinate information of one or more image-regions of optimization.
Image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, uses in step 2 or step 3
First order neutral net, second level neutral net, third level neutral net obtain the coordinate of the image-region of whole object respectively
Information, the coordinate information of image-region of object part, on object part or its periphery marker image-region coordinate
During information, comprise the following steps respectively：
Numerical value in the object global characteristics of 4th global pool layer output is more than corresponding to the feature of the first given threshold
Region be determined as the image-region of whole object and provide coordinate information；
Numerical value in the object part feature of third member pond layer output is more than corresponding to the feature of the second given threshold
Region be determined as the image-region of object part and provide coordinate information；Or
Numerical value in the marker feature of 3rd marker pond layer output is more than corresponding to the feature of the 3rd given threshold
Region be determined as on object part or its periphery marker image-region and provide coordinate information.
The advantage of the invention is that：Can extract can be used in object identification or retrieval more object features (for example,
When object is vehicle, the vehicle characteristics that vehicle is identified or retrieved that are used for extracted are not limited only to license plate number).
Brief description of the drawings
By reading the detailed description of following detailed description, it is various other the advantages of and benefit it is common for this area
Technical staff will be clear understanding.Attached drawing is only used for showing the purpose of embodiment, and is not considered as to the present invention
Limitation.And in whole attached drawing, identical component is denoted by the same reference numerals.In the accompanying drawings：
Fig. 1 shows the schematic diagram of the multiple dimensioned neutral net of embodiment according to the present invention.
Embodiment
The illustrative embodiments of the disclosure are more fully described below with reference to accompanying drawings.Although this public affairs is shown in attached drawing
The illustrative embodiments opened, it being understood, however, that may be realized in various forms the disclosure without the reality that should be illustrated here
The mode of applying is limited.Conversely, there is provided these embodiments are to be able to be best understood from the disclosure, and can be by this public affairs
The scope opened completely is communicated to those skilled in the art.
Fig. 1 shows the schematic diagram of the multiple dimensioned neural network 1 00 of embodiment according to the present invention.
As shown in Figure 1, multiple dimensioned neural network 1 00 (or multiple dimensioned nerve network system 100) includes first order nerve net
Network 101, second level neural network 1 03 and third level neural network 1 05.
First order neural network 1 01 is used for based on input picture (that is, " the input figure in Fig. 1 including at least whole object
Picture ") obtain whole object image-region coordinate information (that is, " coordinate of the image-region of object " in Fig. 1) and object
Global characteristics (that is, " object global characteristics " in Fig. 1).
Moreover, as shown in Figure 1, first order neural network 1 01 includes the first global convolutional layer and the first global pool layer, the
Two global convolutional layers and the second global pool layer, the 3rd global convolutional layer and the 3rd global pool layer, the 4th global convolutional layer and
4th global pool layer.
(that is, pooling1 layers) of first global convolutional layer (that is, Conv1 layers) and the first global pool layer are using VGG16 god
The network of the first convolutional layer and the first maximum pond layer through network.
(that is, pooling2 layers) of second global convolutional layer (that is, Conv2 layers) and the second global pool layer are using VGG16 god
The network of the second convolutional layer and the second maximum pond layer through network.
(that is, pooling3 layers) of 3rd global convolutional layer (that is, Conv3 layers) and the 3rd global pool layer are using VGG16 god
The network of the 3rd convolutional layer and the 3rd maximum pond layer through network.
(that is, pooling4 layers) of 4th global convolutional layer (that is, Conv4 layers) and the 4th global pool layer are using VGG16 god
The network of Volume Four lamination and the 4th maximum pond layer through network.
For example, the specifying information on VGG16 neutral nets may be referred to Karen Simonyan and Andrew
Zisserman was in the article delivered on ICLR (meeting of world study characterization) in 2015《Very Deep Convolutional
Networks For Large-Scale Image Recognition》.Cited by specifically may be referred in its table 1
ConvNet configures C or D (that is, VGG16) and related text description.
Second level neural network 1 03 is used for the image based on the whole object according to acquired in first order neural network 1 01
The coordinate information (that is, " coordinate of the image-region of object " in Fig. 1) in region reduces obtained object global image (that is, Fig. 1
In " object global image ") obtain object part image-region coordinate information (that is, " image of object part in Fig. 1
The coordinate in region ") and object part feature (that is, " object part feature " in Fig. 1).
Moreover, as shown in Figure 1, second level neural network 1 03 includes first component convolutional layer and first component pond layer, the
Two component convolutional layers and second component pond layer, third member convolutional layer and third member pond layer.
(that is, the pooling2_1 layers) use of first component convolutional layer (that is, Conv2_1 layers) and first component pond layer
The network of second convolutional layer of VGG16 neutral nets and the second maximum pond layer.
(that is, the pooling3_1 layers) use of second component convolutional layer (that is, Conv3_1 layers) and second component pond layer
The network of 3rd convolutional layer of VGG16 neutral nets and the 3rd maximum pond layer.
(that is, the pooling4_1 layers) use of third member convolutional layer (that is, Conv4_1 layers) and third member pond layer
The network of the Volume Four lamination of VGG16 neutral nets and the 4th maximum pond layer.
Third level neural network 1 05 is used for the image based on the object part according to acquired in second level neural network 1 03
The coordinate information (that is, " coordinate of the image-region of object part " in Fig. 1) in region reduces the overall situation of obtained object part
Image (that is, " object part image " in Fig. 1) obtain object part on or its periphery marker image-region coordinate
Information (that is, " coordinate of the image-region of the marker on object part or its periphery " in Fig. 1) and marker feature is (i.e.,
" marker feature on object part or its periphery " in Fig. 1).
Moreover, as shown in Figure 1, third level neural network 1 05 includes the first marker convolutional layer and the first marker pond
Layer, the second marker convolutional layer and the second marker pond layer, the 3rd marker convolutional layer and the 3rd marker pond layer.
(that is, the pooling2_2 layers) use of first marker convolutional layer (that is, Conv2_2 layers) and the first marker pond layer
The network of 3rd convolutional layer of VGG16 neutral nets and the 3rd maximum pond layer.
(that is, the pooling3_2 layers) use of second marker convolutional layer (that is, Conv3_2 layers) and the second marker pond layer
The network of the Volume Four lamination of VGG16 neutral nets and the 4th maximum pond layer.
(that is, the pooling4_2 layers) use of 3rd marker convolutional layer (that is, Conv4_2 layers) and the 3rd marker pond layer
The network of 5th convolutional layer of VGG16 neutral nets and the 5th maximum pond layer.
Alternatively, when object is vehicle, object part includes at least at least one of following：Car plate, logo, car
Lamp, tire, front windshield, rear windshield, drive window, copilot window, rear passenger window, rearview mirror, skylight.
For example, when object to be detected is vehicle, first has to detect and export the image district that only includes whole vehicle
Domain, then carries out the image characteristics extraction and detection of above-mentioned critical component for the image-region, this is because these components
Information (including unit type and position coordinates) is particularly useful for correct identification or retrieval vehicle.
Alternatively, when object part is front windshield, the marker on the object part or its periphery includes at least down
At least one of row：Annual test mark, pass, paper-extracting box, pendant, goods of furniture for display rather than for use.
That is, because at front vehicle windshield it can be seen that vehicle interior marker (that is, interior trim information) at most, so can
To select front windshield image as an object portion for needing to carry out it further image characteristics extraction and marker detection
The image of part.Same way, it is also possible to consider using rear windshield, drive window, copilot window, rear passenger window image as needing to it
Carry out the image of the object part of further image characteristics extraction and marker detection.
Although being not shown in Fig. 1, however, those skilled in the art in second level neural network 1 03 it is contemplated that add
Enter the classification results that full articulamentum and soft-max classification layer to obtain object part according to object part feature, and by object portion
The classification results (that is, the specific category of object part) of part and its coordinate information are exported to third level neural network 1 05.In addition,
Full articulamentum and soft-max classification layer can also be added in third level neural network 1 05 to obtain mark according to marker feature
The classification results (that is, the specific category of marker) of will thing simultaneously export its coordinate information.
For above-mentioned multiple dimensioned neural network 1 00 according to the present invention, it is also proposed that one kind is based on the multiple dimensioned nerve net
The image characteristic extracting method of network 100, the described method comprises the following steps：
Step 1：Structure is according to multiple dimensioned neural network 1 00 described above.
Step 2：Multiple dimensioned neural network 1 00 is trained.
Step 3：The coordinate information and image of image-regions at different levels are obtained using trained multiple dimensioned neural network 1 00
Feature.
Alternatively, the image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, the step 2 include
Following steps：
Step 2-1：Sample image is selected, sample image is demarcated, builds proven sample image storehouse.
For example, when being demarcated to sample image, it can be based on video or image carries out object (for example, vehicle, flying
Machine, building etc.) demarcated, can be according to progressively demarcating from top to bottom, from outside to inside, so as at the same time to car during calibration
Itself, before and after vehicle vehicle body everywhere with the marker of visible interior zone (for example, can include vehicle, annual test mark, car plate,
Car light, tire, drive window, rearview mirror, paper-extracting box, pendant, skylight, pass etc.) demarcated, object is obtained (for example, car
, aircraft, building etc.) feature (e.g., including exterior and visible internal feature) set (that is, feature spanning tree), into
And build sample storehouse.
Step 2-2：Using proven sample image storehouse to first order neural network 1 01,03 and of second level neural network 1
Third level neural network 1 05 is trained step by step.
I.e., it is possible to use sample storehouse combination first order neural network 1 01, second level neural network 1 03 and third level nerve
Network 105 (that is, cascading multiple dimensioned network) is layered and (that is, is classified) training, and obtaining Feature Selection Model, (that is, feature detects
Model).
For example, first order neural network 1 01 (that is, coarse grid) (characteristics of image extracted) can be utilized to obtain outside vehicle
Shape information, and then obtain the coordinate information (that is, positional information) of vehicle；Then according to (that is, the middle net of second level neural network 1 03
Network) (characteristics of image extracted) obtain drive window, tire, automobile front lamp, vehicle license, vehicle part and its position such as skylight
Information (that is, coordinate information)；Finally positioned according to third level neural network 1 05 (that is, refined net) (characteristics of image extracted)
Obtain vehicle interior marker and its position (that is, the coordinate informations) such as pendant, annual test marker, paper-extracting box, goods of furniture for display rather than for use.This is also root
Vehicle characteristics are progressively obtained according to the strategy of resolution ratio (that is, the size of input picture from big to small) from big to small, and then obtain car
Detail information.Therefore, above-mentioned neutral net is also referred to as multiple dimensioned neural network 1 00.
Alternatively, the image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, the step 3 include
Following steps：
Step 3-1：To be training including at least the size scaling of the input picture of the pending feature extraction of whole object
When the used input picture including at least whole object size.
That is, first by image according to instruction when carrying out vehicle target detection (that is, carrying out image characteristics extraction for vehicle)
Scale when practicing zooms in and out.
Step 3-2：Successively use the trained multiple dimensioned neural network 1 00 in first order neural network 1 01,
Second level neural network 1 03 and third level neural network 1 05 obtain the coordinate information and characteristics of image of image-regions at different levels.
That is, according to input picture according to first order neural network 1 01, second level neural network 1 03 and third level nerve net
The order of network 105 (that is, according to above-mentioned network structure from outside to inside step by step) output vehicle and its internal information successively.
For example, vehicle position information, Ran Hougen can be obtained according to first order neural network 1 01 (that is, coarse grid) first
Component information (e.g., including component names and its position seat of vehicle are obtained according to second level neural network 1 03 (that is, middle network)
Mark), finally according to the interior trim information of third level neural network 1 05 (that is, refined net) acquisition vehicle.
Alternatively, the image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, the step 3 are also wrapped
Include following steps：
Step 3-3：To neutral nets at different levels, each acquired multiple images region carries out non-maxima suppression, obtains warp
Cross the coordinate information of one or more image-regions of optimization.
For example, vehicle position information is being obtained, according to second level god according to first order neural network 1 01 (that is, coarse grid)
The component information of vehicle is obtained through network 103 (that is, middle network), is obtained according to third level neural network 1 05 (that is, refined net) positioning
During obtaining the vehicle interior markers such as pendant, annual test marker, paper-extracting box, goods of furniture for display rather than for use and its position, it is likely to produce
Some overlapping region frames, therefore to be filtered according to non-maxima suppression, obtain final vehicle, component or marker
Coordinate information is recommended in optimization.
Alternatively, the image characteristic extracting method according to the present invention based on multiple dimensioned neutral net, in step 2 or step
Whole object is obtained respectively using first order neural network 1 01, second level neural network 1 03, third level neural network 1 05 in 3
The coordinate information of image-region, the coordinate information of image-region of object part, the marker on object part or its periphery
Image-region coordinate information when, comprise the following steps respectively：
Numerical value in the object global characteristics of 4th global pool layer output is more than corresponding to the feature of the first given threshold
Region be determined as the image-region of whole object and provide coordinate information；
Numerical value in the object part feature of third member pond layer output is more than corresponding to the feature of the second given threshold
Region be determined as the image-region of object part and provide coordinate information；Or
Numerical value in the marker feature of 3rd marker pond layer output is more than corresponding to the feature of the 3rd given threshold
Region be determined as on object part or its periphery marker image-region and provide coordinate information.
I.e., in this step, can be according to this grade of deep neural network (that is, first order neural network 1 01, second level god
Through network 103, third level neural network 1 05) output convolution characteristic information (that is, object global characteristics, object part feature,
Marker feature), search numerical value is more than the region of (first, second, third) given threshold (for example, half of maximum eigenvalue)
(that is, the higher region of brightness), and respectively by the corresponding original image region in the region be determined as there are object complete image,
There are object part, there are the image-region of marker, provide its coordinate information as there are object complete image, there are object
Component, there are marker image-region final coordinate position.
Above-mentioned technical proposal according to the present invention, the objects such as vehicle, aircraft, building are represented for example, can export
Outside and visible internal feature and its respective positions coordinate, for big data recognizable object (for example, vehicle and the like
Body) information management, comparison, by the vision application such as image retrieval have positive effect.
The above, is only the exemplary embodiment of the present invention, but protection scope of the present invention is not limited to
This, any one skilled in the art the invention discloses technical scope in, the change that can readily occur in or replace
Change, should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection of the claim
Subject to scope.
Claims (8)
- A kind of 1. multiple dimensioned neutral net, it is characterised in that including：First order neutral net, the first order neutral net are used to obtain based on the input picture including at least whole object whole The coordinate information and object global characteristics of the image-region of a object, and including：First global convolutional layer and the first global pool layer, the described first global convolutional layer and the first global pool layer use The network of first convolutional layer of VGG16 neutral nets and the first maximum pond layer；Second global convolutional layer and the second global pool layer, the described second global convolutional layer and the second global pool layer use The network of second convolutional layer of VGG16 neutral nets and the second maximum pond layer；3rd global convolutional layer and the 3rd global pool layer, the described 3rd global convolutional layer and the 3rd global pool layer use The network of 3rd convolutional layer of VGG16 neutral nets and the 3rd maximum pond layer；4th global convolutional layer and the 4th global pool layer, the described 4th global convolutional layer and the 4th global pool layer use The network of the Volume Four lamination of VGG16 neutral nets and the 4th maximum pond layer；Second level neutral net, the second level neutral net are used for based on the whole thing according to acquired in first order neutral net The object global image that the coordinate information of the image-region of body is reduced obtains the coordinate information of the image-region of object part With object part feature, and including：First component convolutional layer and first component pond layer, the first component convolutional layer and first component pond layer use The network of second convolutional layer of VGG16 neutral nets and the second maximum pond layer；Second component convolutional layer and second component pond layer, the second component convolutional layer and second component pond layer use The network of 3rd convolutional layer of VGG16 neutral nets and the 3rd maximum pond layer；Third member convolutional layer and third member pond layer, the third member convolutional layer and third member pond layer use The network of the Volume Four lamination of VGG16 neutral nets and the 4th maximum pond layer；Third level neutral net, the third level neutral net are used for based on the object portion according to acquired in the neutral net of the second level The global image for the object part that the coordinate information of the image-region of part is reduced obtains the mark on object part or its periphery The coordinate information and marker feature of the image-region of will thing, and including：First marker convolutional layer and the first marker pond layer, the first marker convolutional layer and the first marker pond layer Using the 3rd convolutional layer of VGG16 neutral nets and the network of the 3rd maximum pond layer；Second marker convolutional layer and the second marker pond layer, the second marker convolutional layer and the second marker pond layer Using the Volume Four lamination of VGG16 neutral nets and the network of the 4th maximum pond layer；3rd marker convolutional layer and the 3rd marker pond layer, the 3rd marker convolutional layer and the 3rd marker pond layer Using the 5th convolutional layer of VGG16 neutral nets and the network of the 5th maximum pond layer.
- 2. multiple dimensioned neutral net according to claim 1, it is characterised in that when the object is vehicle, the thing Body component includes at least at least one of following：Car plate, logo, car light, tire, front windshield, rear windshield, drive window, the passenger side Sail window, rear passenger window, rearview mirror, skylight.
- 3. multiple dimensioned neutral net according to claim 2, it is characterised in that when the object part is front windshield, The marker on the object part or its periphery includes at least at least one of following：Annual test mark, pass, paper-extracting box, Pendant, goods of furniture for display rather than for use.
- A kind of 4. image characteristic extracting method based on multiple dimensioned neutral net, it is characterised in that including：Step 1：Build multiple dimensioned neutral net according to any one of claim 1 to 3；Step 2：Multiple dimensioned neutral net is trained；Step 3：The coordinate information and characteristics of image of image-regions at different levels are obtained using trained multiple dimensioned neutral net.
- 5. the image characteristic extracting method according to claim 4 based on multiple dimensioned neutral net, it is characterised in that described Step 2 includes：Step 2-1：Sample image is selected, sample image is demarcated, builds proven sample image storehouse；Step 2-2：Using proven sample image storehouse to first order neutral net, second level neutral net and third level nerve Network is trained step by step.
- 6. the image characteristic extracting method according to claim 4 based on multiple dimensioned neutral net, it is characterised in that described Step 3 includes：Step 3-1：To be training when institute including at least the size scaling of the input picture of the pending feature extraction of whole object The size of the input picture including at least whole object used；Step 3-2：Successively using the first order neutral net in the trained multiple dimensioned neutral net, second level nerve Network and third level neutral net obtain the coordinate information and characteristics of image of image-regions at different levels.
- 7. the image characteristic extracting method according to claim 6 based on multiple dimensioned neutral net, it is characterised in that described Step 3 further includes：Step 3-3：To neutral nets at different levels, each acquired multiple images region carries out non-maxima suppression, obtains by excellent The coordinate information of one or more image-regions of change.
- 8. the image characteristic extracting method based on multiple dimensioned neutral net according to any one of claim 4 to 7, its It is characterized in that, is distinguished in step 2 or step 3 using first order neutral net, second level neutral net, third level neutral net Obtain the coordinate information of image-region of whole object, the coordinate information of image-region of object part, on object part or its During the coordinate information of the image-region of the marker on periphery, comprise the following steps respectively：Numerical value in the object global characteristics of 4th global pool floor output is more than the area corresponding to the feature of the first given threshold Domain is determined as the image-region of whole object and provides coordinate information；Numerical value in the object part feature of third member pond floor output is more than the area corresponding to the feature of the second given threshold Domain is determined as the image-region of object part and provides coordinate information；OrNumerical value in the marker feature of 3rd marker pond floor output is more than the area corresponding to the feature of the 3rd given threshold Domain is determined as the image-region of the marker on object part or its periphery and provides coordinate information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711093435.5A CN108038540A (en) | 2017-11-08 | 2017-11-08 | A kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711093435.5A CN108038540A (en) | 2017-11-08 | 2017-11-08 | A kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108038540A true CN108038540A (en) | 2018-05-15 |
Family
ID=62092810
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711093435.5A Pending CN108038540A (en) | 2017-11-08 | 2017-11-08 | A kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108038540A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109815300A (en) * | 2018-12-13 | 2019-05-28 | 北京邮电大学 | A kind of vehicle positioning method |
| CN109903267A (en) * | 2019-01-22 | 2019-06-18 | 江苏恒力化纤股份有限公司 | A method of based on image processing techniques test network wire network degree |
| CN110299028A (en) * | 2019-07-31 | 2019-10-01 | 深圳市捷顺科技实业股份有限公司 | Method, apparatus, equipment and the readable storage medium storing program for executing of line detection are got in a kind of parking |
| WO2020048273A1 (en) * | 2018-09-07 | 2020-03-12 | 阿里巴巴集团控股有限公司 | Neural network system for image matching and location determination, method, and device |
| CN111898502A (en) * | 2020-07-20 | 2020-11-06 | 北京格灵深瞳信息技术有限公司 | Dangerous goods vehicle identification method and device, computer storage medium and electronic equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824049A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded neural network-based face key point detection method |
| CN105046196A (en) * | 2015-06-11 | 2015-11-11 | 西安电子科技大学 | Front vehicle information structured output method base on concatenated convolutional neural networks |
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
-
2017
- 2017-11-08 CN CN201711093435.5A patent/CN108038540A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824049A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded neural network-based face key point detection method |
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
| CN105046196A (en) * | 2015-06-11 | 2015-11-11 | 西安电子科技大学 | Front vehicle information structured output method base on concatenated convolutional neural networks |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020048273A1 (en) * | 2018-09-07 | 2020-03-12 | 阿里巴巴集团控股有限公司 | Neural network system for image matching and location determination, method, and device |
| CN109815300A (en) * | 2018-12-13 | 2019-05-28 | 北京邮电大学 | A kind of vehicle positioning method |
| CN109815300B (en) * | 2018-12-13 | 2021-06-29 | 北京邮电大学 | Vehicle positioning method |
| CN109903267A (en) * | 2019-01-22 | 2019-06-18 | 江苏恒力化纤股份有限公司 | A method of based on image processing techniques test network wire network degree |
| CN109903267B (en) * | 2019-01-22 | 2021-01-05 | 江苏恒力化纤股份有限公司 | Method for testing network wire network degree based on image processing technology |
| CN110299028A (en) * | 2019-07-31 | 2019-10-01 | 深圳市捷顺科技实业股份有限公司 | Method, apparatus, equipment and the readable storage medium storing program for executing of line detection are got in a kind of parking |
| CN110299028B (en) * | 2019-07-31 | 2022-06-14 | 深圳市捷顺科技实业股份有限公司 | Parking line crossing detection method, device, equipment and readable storage medium |
| CN111898502A (en) * | 2020-07-20 | 2020-11-06 | 北京格灵深瞳信息技术有限公司 | Dangerous goods vehicle identification method and device, computer storage medium and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108038540A (en) | A kind of multiple dimensioned neutral net and the image characteristic extracting method based on the network | |
| CN109086668B (en) | Unmanned aerial vehicle remote sensing image road information extraction method based on multi-scale generation countermeasure network | |
| Kong et al. | General road detection from a single image | |
| CN109636795B (en) | Real-time non-tracking monitoring video remnant detection method | |
| CN107633220A (en) | A kind of vehicle front target identification method based on convolutional neural networks | |
| DE102017127346A1 (en) | BRAKE DETECTION | |
| CN108830199A (en) | Identify method, apparatus, readable medium and the electronic equipment of traffic light signals | |
| CN106599773A (en) | Deep learning image identification method and deep learning image identification system used for intelligent driving, and terminal device | |
| CN104463241A (en) | Vehicle type recognition method in intelligent transportation monitoring system | |
| CN109389046B (en) | All-weather object identification and lane line detection method for automatic driving | |
| CN106022300A (en) | Traffic sign identifying method and traffic sign identifying system based on cascading deep learning | |
| CN111767878B (en) | Deep learning-based traffic sign detection method and system in embedded device | |
| KR101845769B1 (en) | Car rear detection system using convolution neural network, and method thereof | |
| CN105809121A (en) | Multi-characteristic synergic traffic sign detection and identification method | |
| CN106250845A (en) | Flame detecting method based on convolutional neural networks and device | |
| CN103902976A (en) | Pedestrian detection method based on infrared image | |
| CN106650731A (en) | Robust license plate and logo recognition method | |
| CN103996198A (en) | Method for detecting region of interest in complicated natural environment | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN107038420A (en) | A kind of traffic lights recognizer based on convolutional network | |
| CN110386065A (en) | Monitoring method, device, computer equipment and the storage medium of vehicle blind zone | |
| CN106897681A (en) | A kind of remote sensing images comparative analysis method and system | |
| CN106407951B (en) | A kind of night front vehicles detection method based on monocular vision | |
| CN106257490A (en) | The method and system of detection driving vehicle information | |
| CN106204509A (en) | Based on region characteristic infrared and visible light image fusion method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180515 |