CN108764459A - Target identification network design method based on semantical definition - Google Patents
Target identification network design method based on semantical definition Download PDFInfo
- Publication number
- CN108764459A CN108764459A CN201810465726.0A CN201810465726A CN108764459A CN 108764459 A CN108764459 A CN 108764459A CN 201810465726 A CN201810465726 A CN 201810465726A CN 108764459 A CN108764459 A CN 108764459A
- Authority
- CN
- China
- Prior art keywords
- semantic
- layer
- image
- primitive
- neuron
- 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
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes a kind of target identification network design method based on semantical definition, mainly solves existing target identification network and big data is needed to drive, and high energy consumption is time-consuming, generalization ability and the weak problem of migration.Its implementation is：1. the Semantic unit for defining semantic hierarchical structure and being felt based on the mankind；2. separately designing the different types of base's semanteme neuron that can execute different function according to the Visual Neuron of mankind's primary visual cortex V1, it is each responsible for detection level line, vertical line, oblique line, straight line, camber line, triangle, quadrangle, polygon and color；3. the level of identification target is semantic as required, semantics recognition network is successively built；4. picture is identified with semantics recognition network.The present invention is inspired based on brain, and big data driving is not needed, and low energy consumption, and has very high generalization ability and migration, can be used for the identification to plurality of target.
Description
Technical field
The invention belongs to artificial intelligence fields, relate generally to a kind of new target identification network, can be used for multi-class targets
It is identified.
Background technology
The huge advance of current information technology pushes the basic theory of artificial intelligence, methods and techniques to develop in depth.
With information technology, big data, deep learning, the huge advance of cranial nerve science pushes artificial intelligence technology again
Leap.Artificial intelligence, which is just being showed, to be mixed based on information and knowledge processing, with human knowledge, can independently complete more awarenesses
The new features such as work.
In recent years, more and more based on deep learning algorithm with the fast development of artificial intelligence field research work
Neural network achieves marvelous results in computer vision, speech recognition, natural language processing etc., especially in computer
The field of target recognition of vision, such as：The networks such as VGGNet, InceptionNet, ResNet, they are all to use smaller volume
Product core and deeper network structure.Although these networks can efficiently extract the significant characteristics of image, in target identification
Also all well and good achievement is achieved in task.But these target identification networks based on deep learning are all by a set of at present
Common working mechanism's operation, and constraint rule can only be portrayed with mathematical function, according to the constraint rule, network could lead to
It crosses propagated forward and reversed adjust weighs progress accurate calculation, the solution of error minimum, the fixed operating mode meeting of this set are found out with fitting
Bring following several deficiencies：
1. the training of network drives by big data, a large amount of labeled data and cumbersome tune is needed to join process, and number
It according to storage and calculates independently of each other, data transmission accounts for significant proportion, causes to store and calculating is complicated, energy consumption is very high.
2. there are the three difficult problems of " hardly possible design, hardly possible is expected, is difficult to resolve and releases " in network.
3. network lacks transfer learning and generalization ability, that is, single task role is can be only done rather than multitask, for same
The sample adaptability of type objects in varied situations is poor, for example does can only just identifying face and being directed to just for recognition of face
The face recognition effect in face is good, if face is partially a little bad with regard to None- identified or recognition effect slightly.
Invention content
It is above-mentioned to solve it is an object of the invention to propose a kind of target identification network design method based on semantical definition
The deficiency of existing target identification network reduces storage, computation complexity and energy consumption, improves transfer learning and generalization ability, realizes
Quickly and effectively target identification.
The present invention technological thought be：A completely new network is constructed by target semanteme knowledge definition, the mankind are obtained
The target information obtained is directly injected into computer help computer identification target.
According to above-mentioned thought, the present invention is based on the target identification network design methods of semantical definition, it is characterised in that：Packet
It includes：
1) definition semanteme and Semantic unit：
Semanteme be by multiple sub- semantic combinations, it is semantic again by sub- semantic combination per height, and so on, up to the bottom
Semantic unit and the hierarchical structure that constitutes；
Semantic unit is the most basic sensory information of finger vision, the sense of hearing, smell, the sense of taste, tactile these types mankind；It is different
Semantic unit between there are relevances；
2) all kinds of primitive filters are designed：
Different type 2a) is separately designed according to the neuron of mankind's primary visual cortex v1, and the base of different function can be executed
First filter is filtered for the pretreatment image to input, detects different vision Semantic units, the vision semanteme base
Member includes：Point, horizontal line, vertical line, oblique line, camber line, triangle, four side types, polygon, circle, circular arc, color；
2b) according to the matching degree of primitive filter and pretreatment image block so that the output of primitive filter is 0 or 1,
Or some decimal between 0 to 1, and the pulse inputs of respective magnitudes is converted into next layer；
3) semantics recognition network is built：
3a) pretreatment image of input is filtered with different primitive filters, obtains the image containing Semantic unit,
Corresponding Semantic unit is extracted from the image again, to form the bottom of semantics recognition network；
The different Semantic unit of bottom 3b) is subjected to longitudinal combination according to the target information of required identification, is abstracted into and is higher than
The semanteme of bottom forms the semantics recognition network second layer, then from the second layer, is abstracted into the semanteme higher than the second layer, is formed
Semantics recognition network third layer, and so on, it is further combined to the higher layer of semantics recognition network, until top layer forms mesh
Poster justice；
It 3c) is added in bottom according to the relative tertiary location relationship between Semantic unit laterally attached and each in network
Parallel link is carried out between layer, the wherein spatial relation between Semantic unit includes：Up and down, it controls, intersect, containing and is preceding
Afterwards；
Each classification target semantic information carrier in network 3d) is defined as semantic neuron, it in hierarchical fashion will be every
The semantic neuron parallel arranged of a kind of target forms the semantics recognition network that can identify multi-class targets；
4) identification process of semantics recognition network：
Pretreated image is input to semantics recognition network, is used in a networkIndicate i-th layer of t-th of semanteme
Neuron, the then weighted type that the output of each semantic neuron is inputted by its last layer semanteme neuron are expressed as：
Wherein, i=1,2 ..., n, n are total numbers of plies of network, K (i-1) be all semantic neurons in (i-1)-th layer collection
It closes,For the weight coefficient corresponding to (i-1)-th layer of t-th of semantic neuron；
The first layer of network, pretreatment image of this layer semanteme neuron, that is, primitive filter to input are indicated as i=1
It is filtered and extracts Semantic unit and be input to the second layer, the relevant semantic neuron of the activation second layer, then extract the second layer
Corresponding semantic information is input to third layer, activates the relevant semantic neuron of third layer, and so on, until successively activating institute
There is relevant semantic neuron；
As i=n, the result of identification is exported on top layer accordingly semantic neuron.
Compared with the prior art, the present invention has the following advantages：
1. semantics recognition network proposed by the present invention is inspired based on brain, total is to copy the objective generation of human cognitive
The Process Design on boundary, it is proposed that semantic hierarchical structure, and being identified using the priori of the mankind for many years allows base
In the Semantic unit alternate data bit of sense organ, artificial intelligence truly is realized, reach the target that every people can identify
Computer can identify.
2. the present invention does not need a large amount of training set and test set, completely by Knowledge driving, obtained using low volume data
While good recognition effect, storage overhead, training energy consumption and cumbersome tune ginseng that traditional big data driving is brought are avoided
Process.
3. the present invention compared with tional identification network, by data set training limited, can to semantic hierarchical structure into
Row increase and decrease achievees the purpose that identify multiple target, realizes multitask classification, have good migration.
4. the present invention is due to the relevance between flexibly defining each semantic neuron so that the network has very strong general
Change ability.
Description of the drawings
Fig. 1 is the implementation flow chart of the present invention；
Fig. 2 is the semantic hierarchical structure schematic diagram that the present invention defines；
Fig. 3 is the vision semanteme neuron schematic diagram that four classes that the present invention designs detect lines；
Fig. 4 is that the present invention is based on the semantics recognition network diagrams that semantic hierarchical structure is built；
Fig. 5 is the instance graph with semantics recognition Network Recognition handwritten word " 7 " proposed by the present invention.
Specific implementation mode
The present invention is described in detail with example below in conjunction with the accompanying drawings.
Referring to Fig.1, steps are as follows for realization of the invention：
Step 1, definition semanteme and Semantic unit.
1a) definition is semantic：
Traditional semanteme refer to transmission signal data in for the meaning that is more readily understood, and the semanteme that the present invention defines be by
Q sub- semantic combinations, it is semantic again by p sub- semantic combination per height, and so on, up to the bottom Semantic unit and structure
At hierarchical structure, wherein q >=1, p >=1.As commaterial is by molecular composition, molecule is made of atom and atom is by electricity
Son and atomic nucleus combine as, in image data target object by its each component form and each component by junior unit group
At, such as a little white rabbit can define in this way：Long long ear, red eyes, the mouth of three petaloids, white, it is fluffy
Body, wherein ear, eyes, mouth can use more specifical concept, be defined such as outline, color, such as Fig. 2 institutes
Show.
1b) define Semantic unit：
By describing fundamental sensation to recognize things be the first step that the mankind initially recognize, for example, the shape of things, color,
Temperature, taste and sense of touch, and these be used to describe fundamental sensation vocabulary can't because of people stock of knowledge it is different and
Difference, the mankind have vision, the sense of hearing, smell, the sense of taste, tactile this five kinds of fundamental sensations, the fundamental sensation that these mankind share
It is the basis for constituting Semantic unit.Therefore, each Semantic unit that the present invention defines represents a kind of fundamental sensation, that is, is formed
Vision Semantic unit, sense of hearing Semantic unit, smell Semantic unit, sense of taste Semantic unit and tactile Semantic unit, wherein：
Vision Semantic unit, including：Shape, contrast, saturation degree and color, the shape again include point, line, plane and
Solid, the color include red, orange, yellow, green, blue, blue, purple, black and white again；
Sense of hearing Semantic unit, including：Tone, volume and tone color；
Smell Semantic unit, including：Perfume odor, rosin, fruit aroma, putrefactive odor, empyreumatic taste；
Sense of taste Semantic unit, including：Sweet tea, acid, hardship and salty；
Tactile Semantic unit, including：Temperature, humidity, pain, pressure and vibration；
1c) define between different classes of Semantic unit that there are relevances, such as one section of voice can be with vision or the sense of taste
Or the Semantic unit in tactile or smell is interrelated, that is, organic combination is carried out to various feelings similar to human perception,
It is significant entirety by a variety of attribute synthesis of things, for example, one people of identification can pass through the sound of the silhouette and he of this people
It is associated to obtain a result.This pattern of being mutually related is the important foundation that semantics recognition network forms complicated semanteme so that net
The information of network being capable of multimode.
Step 2, all kinds of underlying semantics neurons are designed.
The process that information is extracted by imitating human perception objective world can design the underlying semantics god of different sensory systems
Corresponding Semantic unit information is extracted through member, underlying semantics neuron with different feelings can use different physical component or electronics
Device realization, such as：Underlying semantics neuron based on tactile can be designed as temperature, humidity and pressure sensor.
In the present invention, it is mainly illustrated by taking the design of vision system underlying semantics neuron as an example.
Different type 2a) is separately designed according to the neuron of mankind's primary visual cortex v1, and the base of different function can be executed
Plinth semanteme neuron, is denoted as primitive filter, is filtered for the pretreatment image to input, detects different vision languages
Adopted primitive, the vision Semantic unit include：Point, horizontal line, vertical line, oblique line, camber line, triangle, four side types, polygon, circle,
Circular arc and color, actually each primitive filter is exactly the spatial filter for having specific shape or attribute, for detecting
And only detect some corresponding vision Semantic unit；
2b) according to the matching degree of primitive filter and pretreatment image block so that the output of primitive filter is 0 or 1,
Or some decimal between 0 to 1：
When the target in primitive filter and image block exactly matches, i.e., primitive filter extracts Semantic unit and figure
As information in the block is completely the same, then the output of primitive filter is 1；
When the target in primitive filter and image block mismatches completely, i.e., the Semantic unit of primitive filter extraction with
Image information in the block is completely inconsistent, then the output of primitive filter is 0；
When the target Incomplete matching in primitive filter and image block, i.e., the Semantic unit of primitive filter extraction with
Image INFORMATION OF INCOMPLETE in the block is consistent, then the output of primitive filter is the number between 0 to 1, such as：When image mesh in the block
When there is graticule item the deflection of certain angle to be matched with the primitive filter for being responsible for detection vertical line specially, primitive filter
The number between one 0 to 1 can be exported；
2c) by 2b) the obtained output of primitive filter is converted into the pulse inputs of respective magnitudes to next layer, then pulse
Amplitude will characterize more advanced semanteme as the weight of each input picture block；
Above-mentioned 2a) to 2c) this process is equivalent to 2a) and described in the primitive filter of detection vision Semantic unit constitute
One optic nerve tuple completes the vision Semantic unit in perceptual image and converts thereof into the process of nerve impulse.
Step 3, semantics recognition network is built.
Complicated semanteme can be formed by sub- semantic combination, and a combination thereof process is the process of high-grade intelligent, the identification of the mankind
A kind of Knowledge based engineering composition identification is can be regarded as, i.e., is felt to carry out different combinations by different bases to complete, this
Kind of recognition mode is built upon on the neural net base of people, for example, human vision signal from retina to foreign journals again to
Visual cortex, including primary visual cortex V1 and line exodermis, the visual information in pathways for vision is a kind of transmission of rank character, herein
In the process, more abstract concept is will produce in the information input of the innervation of a level to next layer, is inspired by this, this
Hierarchical structure of the invention based on target semanteme, one semantics recognition different from conventional target identification network concept and structure of structure
Network, construction step are as follows：
3a) two-value contour edge image after the pretreatment of input is filtered with different primitive filters, is contained
The image of Semantic unit, then corresponding Semantic unit is extracted from the image, to form the bottom of semantics recognition network：
Target image pretreatment after M × M two-value contour edge images, 3a1) is further partitioned into the image of N × N at size
The size of block, each block isWherein M >=1, N >=1；
3a2) separately design can detection level line, vertical line, oblique line, arc spatial filter, constitute four class detection lines
The primitive filter of item represents it and is exported to the responsiveness of specific position as shown in figure 3, the color of primitive filter is deeper
The amplitude of pulse is bigger；
Square formation 3a3) will be lined up in rows and columns respectively per a kind of primitive filter, being sized makes itself and 3a1) obtain
Image block matches；
3a4) traversal whole image is filtered all matched image blocks, obtains the figure for containing only corresponding Semantic unit
Picture；
It includes detection 3a5) to design more rich primitive filter：Triangle, quadrangle, polygon and circle, for right
The space filtering of complicated image, filtered on relatively more complicated image by these primitive filter obtain it is corresponding more rich
Semantic unit image.For example, the image containing bicycle is carried out with a variety of different primitive filters after a width is pre-processed
Filtering, then being responsible for the circular primitive filter of detection will propose to contain only circular Semantic unit in the place of two wheels
Circle in image, that is, input picture can and can only activate responsible circular primitive filter；It is responsible for the base of triangle
First filter will propose the only Semantic unit image containing triangle in the position of bicycle beam, such, pass through primitive
All Semantic units contained by target in input picture can be obtained after the filtering operation of filter；
The different Semantic unit of bottom 3b) is subjected to longitudinal combination according to the target information of required identification, is abstracted into and is higher than
The semanteme of bottom forms the semantics recognition network second layer, then from the second layer, is abstracted into the semanteme higher than the second layer, is formed
Semantics recognition network third layer, and so on, it is further combined to the higher layer of semantics recognition network, until top layer forms mesh
Poster justice, as shown in Figure 4；
3c) in addition to 3b) other than longitudinal passage semantic in step, according to the relative tertiary location relationship sheet between Semantic unit
Invention is laterally attached in bottom addition, i.e., it is general that the new semanteme of composition one can be connected between relevant underlying semantics neuron
It reads, such as：The camber line of different angle can connect into the shapes such as a circle, semicircle, s types, ellipse, the wherein sky between Semantic unit
Between position relationship include：Up and down, it controls, intersect, containing and is front and back；In addition, carrying out parallel link, class between each layer of network
The throwback relationship being similar between ancestor node and descendant nodes, high-level semantics can also be with alternating floors or every several layers of rudimentary semantemes even
It connects and forms new semanteme.For example, after high level obtains classification semantic concept " apple ", then the semanteme " green color " with low layer
New semanteme " granny smith " can be formed；
Each classification target semantic information carrier in network 3d) is defined as semantic neuron, it in hierarchical fashion will be every
The semantic neuron parallel arranged of a kind of target forms the semantics recognition network that can identify multi-class targets.
Step 4, the identification process of semantics recognition network.
Pretreated image is input to semantics recognition network and carries out target identification, uses m in a networkt (i)Indicate i-th
The semantic neuron of t-th of layer, then the weighting shape that the output of each semantic neuron is inputted by its last layer semanteme neuron
Formula is expressed as：
Wherein, i=1,2 ..., n, n are total number of plies of network, K(i-1)For the collection of all semantic neurons in (i-1)-th layer
It closes,For the weight coefficient corresponding to (i-1)-th layer of t-th of semantic neuron；
It is different according to the value of i, successively activate semantic neuron：
The first layer of network, pretreatment image of this layer semanteme neuron, that is, primitive filter to input are indicated as i=1
It is filtered and extracts Semantic unit and be input to the second layer, the relevant semantic neuron of the activation second layer, then extract the second layer
Corresponding semantic information is input to third layer, activates the relevant semantic neuron of third layer, and so on, until successively activating institute
There is relevant semantic neuron；
As i=n, exported on top layer accordingly semantic neuron identifying as a result, for example, being counted in handwritten word data set
It is divided into 3 × 3 after the image input network of word " 7 " totally 9 blocks, each block is connected to 3a2) the primitive filters of four kinds of detections lines
Wave device is detected, i.e., corresponding four semantic output, these outputs activate the god of each level according to corresponding semantic path successively
Pulse can be exported on the high-level semantics neuron that top layer represents " 7 " through member, i.e. high-level semantics can be expressed as these activation
Neuron weighted sum, thus achieve the purpose that identification is digital " 7 ", as shown in Figure 5.
Above description is only example of the present invention, does not constitute any limitation of the invention, it is clear that for
It, all may be without departing substantially from the principle of the invention, structure after having understood the content of present invention and principle for one of skill in the art
In the case of, carry out various modifications in form and details and change, but these modifications and variations based on inventive concept
Still within the claims of the present invention.
Claims (7)
1. the target identification network design method based on semantical definition, which is characterized in that including：
1) definition semanteme and Semantic unit：
Semanteme be by multiple sub- semantic combinations, it is semantic again by sub- semantic combination per height, and so on, up to the language of the bottom
Adopted primitive and the hierarchical structure constituted；
Semantic unit is the most basic sensory information of finger vision, the sense of hearing, smell, the sense of taste, tactile these types mankind；Different languages
There are relevances between adopted primitive；
2) all kinds of primitive filters are designed：
Different type 2a) is separately designed according to the neuron of mankind's primary visual cortex v1, and the primitive filter of different function can be executed
Wave device is filtered for the pretreatment image to input, detects different vision Semantic units, the vision Semantic unit packet
It includes：Point, horizontal line, vertical line, oblique line, camber line, triangle, four side types, polygon, circle, circular arc and color；
2b) according to the matching degree of primitive filter and pretreatment image block so that the output of primitive filter is 0 or 1 or 0
Some decimal between to 1, and the pulse inputs of respective magnitudes is converted into next layer；
3) semantics recognition network is built：
3a) pretreatment image of input is filtered with different primitive filters, obtains the image containing Semantic unit, then from
Corresponding Semantic unit is extracted in the image, to form the bottom of semantics recognition network；
The different Semantic unit of bottom 3b) is subjected to longitudinal combination according to the target information of required identification, is abstracted into higher than bottom
Semanteme, form the semantics recognition network second layer, then from the second layer, be abstracted into the semanteme higher than the second layer, formed semantic
Identify network third layer, and so on, it is further combined to the higher layer of semantics recognition network, until top layer forms target language
Justice；
3c) according to the relative tertiary location relationship between Semantic unit bottom be added it is laterally attached, and each layer of network it
Between carry out parallel link, the wherein spatial relation between Semantic unit includes：Up and down, it controls, intersect, containing and is front and back；
Each classification target semantic information carrier in network 3d) is defined as semantic neuron, it in hierarchical fashion will be per a kind of
The semantic neuron parallel arranged of target forms the semantics recognition network that can identify multi-class targets；
4) identification process of semantics recognition network：
Pretreated image is input to semantics recognition network, is used in a networkIndicate i-th layer of t-th of semantic nerve
Member, the then weighted type that the output of each semantic neuron is inputted by its last layer semanteme neuron are expressed as：
Wherein, i=1,2 ..., n, n are total number of plies of network, K(i-1)For the set of all semantic neurons in (i-1)-th layer,For the weight coefficient corresponding to (i-1)-th layer of t-th of semantic neuron；
Indicate that the first layer of network, this layer semanteme neuron, that is, primitive filter carry out the pretreatment image of input as i=1
Filtering extracts Semantic unit and is input to the second layer, the relevant semantic neuron of the activation second layer, then to extract the second layer corresponding
Semantic information be input to third layer, the relevant semantic neuron of activation third layer, and so on, until successively activating all phases
The semantic neuron of pass；
As i=n, the result of identification is exported on top layer accordingly semantic neuron.
2. according to the method described in claim 1, the basic sense of the vision, the sense of hearing, smell, the sense of taste, tactile wherein in step 1)
Feel information, is described as follows respectively：
The basic visual information, including：Shape, contrast, saturation degree and color；
The basic auditory information, including：Tone, volume and tone color；
The basic olfactory information, including：Perfume odor, rosin taste, fruit aroma, putrefactive odor and empyreumatic taste；
The basic taste sensation information, including：Sweet tea, acid, hardship and salty；
The basic tactile data, including：Temperature, humidity, pain, pressure and vibration.
3. according to the method described in claim 2, the shape, including：Point, line, surface and stereochemical structure.
4. according to the method described in claim 2, the color, including：It is red, orange, yellow, green, blue, blue, purple, black and white.
5. according to the method described in claim 1, wherein step 3a) in different primitive filters to the pretreatment figure of input
As being filtered, the image containing Semantic unit is obtained, is carried out as follows：
3a1) image that input size is M × M is divided into N × N's after pretreatment is at two-value contour edge image
The size of image block, each block isWherein M >=1, N >=1；
3a2) separately design can detection level line, vertical line, oblique line, arc spatial filter, constitute four classes detection lines
Primitive filter；
Square formation 3a3) will be lined up in rows and columns respectively per a kind of primitive filter, being sized makes itself and 3a1) image that obtains
Block matches；
3a4) traversal whole image is filtered all matched image blocks, obtains the image for containing only corresponding Semantic unit.
6. according to the method described in claim 4, wherein step 3a2) in the primitive filter that designs, further include triangle, four
Side shape, polygon and circle, for the space filtering to complicated image.
7. according to the method described in claim 1, wherein step 2b) according to of primitive filter and pretreated image block
With degree so that some decimal of the output of primitive filter between 0 or 1 or 0 to 1 carries out according to the following rules：
When the target in primitive filter and image block exactly matches, i.e., primitive filter extracts Semantic unit and image block
In information it is completely the same, then the output of primitive filter be 1；
When the target in primitive filter and image block mismatches completely, i.e., primitive filter extracts Semantic unit and image
Information in the block is completely inconsistent, then the output of primitive filter is 0；
When the target Incomplete matching in primitive filter and image block, i.e., primitive filter extracts Semantic unit and image
INFORMATION OF INCOMPLETE in the block is consistent, then the output of primitive filter is the number between 0 to 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810465726.0A CN108764459B (en) | 2018-05-16 | 2018-05-16 | Target recognition network design method based on semantic definition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810465726.0A CN108764459B (en) | 2018-05-16 | 2018-05-16 | Target recognition network design method based on semantic definition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108764459A true CN108764459A (en) | 2018-11-06 |
| CN108764459B CN108764459B (en) | 2020-06-16 |
Family
ID=64008031
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810465726.0A Active CN108764459B (en) | 2018-05-16 | 2018-05-16 | Target recognition network design method based on semantic definition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108764459B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112078593A (en) * | 2020-07-24 | 2020-12-15 | 西安电子科技大学 | Automatic driving system and method based on multiple network collaborative models |
| CN113705245A (en) * | 2021-09-01 | 2021-11-26 | 北京邮电大学 | Semantic communication method, device, system, computer equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1658234A (en) * | 2004-02-18 | 2005-08-24 | 国际商业机器公司 | Hierchical visual structure of generating semantic network |
| CN106408030A (en) * | 2016-09-28 | 2017-02-15 | 武汉大学 | SAR image classification method based on middle lamella semantic attribute and convolution neural network |
| CN106599883A (en) * | 2017-03-08 | 2017-04-26 | 王华锋 | Face recognition method capable of extracting multi-level image semantics based on CNN (convolutional neural network) |
| CN106803256A (en) * | 2017-01-13 | 2017-06-06 | 深圳市唯特视科技有限公司 | A kind of 3D shape based on projection convolutional network is split and semantic marker method |
| US20180121799A1 (en) * | 2016-11-03 | 2018-05-03 | Salesforce.Com, Inc. | Training a Joint Many-Task Neural Network Model using Successive Regularization |
| CN108132968A (en) * | 2017-12-01 | 2018-06-08 | 西安交通大学 | Network text is associated with the Weakly supervised learning method of Semantic unit with image |
-
2018
- 2018-05-16 CN CN201810465726.0A patent/CN108764459B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1658234A (en) * | 2004-02-18 | 2005-08-24 | 国际商业机器公司 | Hierchical visual structure of generating semantic network |
| CN106408030A (en) * | 2016-09-28 | 2017-02-15 | 武汉大学 | SAR image classification method based on middle lamella semantic attribute and convolution neural network |
| US20180121799A1 (en) * | 2016-11-03 | 2018-05-03 | Salesforce.Com, Inc. | Training a Joint Many-Task Neural Network Model using Successive Regularization |
| CN106803256A (en) * | 2017-01-13 | 2017-06-06 | 深圳市唯特视科技有限公司 | A kind of 3D shape based on projection convolutional network is split and semantic marker method |
| CN106599883A (en) * | 2017-03-08 | 2017-04-26 | 王华锋 | Face recognition method capable of extracting multi-level image semantics based on CNN (convolutional neural network) |
| CN108132968A (en) * | 2017-12-01 | 2018-06-08 | 西安交通大学 | Network text is associated with the Weakly supervised learning method of Semantic unit with image |
Non-Patent Citations (4)
| Title |
|---|
| NICHOLLS J G ETAL.: "From Neuron to Brain:A Cellular and Molecular Approach to the Function of Nervous System", 《SINAUER ASSOCIATES》 * |
| 刘剑 等: "面向智能搜索的动态知识网络建模", 《电信科学》 * |
| 石光明 等: "语义通讯———智能时代的产物", 《模式识别与人工智能》 * |
| 谭阳: "基于脑波的图像情感特征映射研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112078593A (en) * | 2020-07-24 | 2020-12-15 | 西安电子科技大学 | Automatic driving system and method based on multiple network collaborative models |
| CN112078593B (en) * | 2020-07-24 | 2021-12-21 | 西安电子科技大学 | Automatic driving system and method based on multiple network collaborative models |
| CN113705245A (en) * | 2021-09-01 | 2021-11-26 | 北京邮电大学 | Semantic communication method, device, system, computer equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108764459B (en) | 2020-06-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378985B (en) | Animation drawing auxiliary creation method based on GAN | |
| CN106650806B (en) | A kind of cooperating type depth net model methodology for pedestrian detection | |
| CN106778506A (en) | A kind of expression recognition method for merging depth image and multi-channel feature | |
| CN105426919B (en) | The image classification method of non-supervisory feature learning is instructed based on conspicuousness | |
| CN108985377A (en) | A kind of image high-level semantics recognition methods of the multiple features fusion based on deep layer network | |
| CN109299396A (en) | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model | |
| CN106504064A (en) | Clothes classification based on depth convolutional neural networks recommends method and system with collocation | |
| CN110263912A (en) | A kind of image answering method based on multiple target association depth reasoning | |
| CN110399821A (en) | Customer satisfaction acquisition methods based on facial expression recognition | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| CN108595602A (en) | The question sentence file classification method combined with depth model based on shallow Model | |
| CN107679462A (en) | A kind of depth multiple features fusion sorting technique based on small echo | |
| CN109299701A (en) | Expand the face age estimation method that more ethnic group features cooperate with selection based on GAN | |
| CN108596051A (en) | A kind of intelligent identification Method towards product style image | |
| CN108932517A (en) | A kind of multi-tag clothes analytic method based on fining network model | |
| Spratt et al. | Computational beauty: Aesthetic judgment at the intersection of art and science | |
| CN108921047A (en) | A kind of multi-model ballot mean value action identification method based on cross-layer fusion | |
| CN107423727A (en) | Face complex expression recognition methods based on neutral net | |
| CN108416397A (en) | A kind of Image emotional semantic classification method based on ResNet-GCN networks | |
| CN106777402A (en) | A kind of image retrieval text method based on sparse neural network | |
| CN107066979A (en) | A kind of human motion recognition method based on depth information and various dimensions convolutional neural networks | |
| Zhai et al. | Asian female facial beauty prediction using deep neural networks via transfer learning and multi-channel feature fusion | |
| CN107492038A (en) | A kind of community discovery method based on neutral net | |
| CN106980831A (en) | Based on self-encoding encoder from affiliation recognition methods | |
| Seeley | Attentional engines: A perceptual theory of the arts |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |