CN105404902B - Characteristics of image description and accumulating method based on impulsive neural networks - Google Patents
Characteristics of image description and accumulating method based on impulsive neural networks Download PDFInfo
- Publication number
- CN105404902B CN105404902B CN201510708874.7A CN201510708874A CN105404902B CN 105404902 B CN105404902 B CN 105404902B CN 201510708874 A CN201510708874 A CN 201510708874A CN 105404902 B CN105404902 B CN 105404902B
- Authority
- CN
- China
- Prior art keywords
- neuron
- layer
- granting
- neural networks
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
Abstract
The characteristics of image description and accumulating method that the invention proposes a kind of based on impulsive neural networks, it include: M normalized images of input, the number of plies of impulsive neural networks is determined according to image size, gradient direction at each pixel is obtained to image preprocessing, and its discrete is turned into preset value value, a granting in impulsive neural networks in the every preset value neuron of first layer is determined according to the gradient direction after discretization, the film potential of second layer neuron is calculated according to the granting situation of first layer neuron, to determine the granting situation of second layer neuron, obtain the granting situation of all layers of neuron, the connection weight between each layer of synaptic plasticity STDP rule adjustment impulsive neural networks relied on according to the sequential relationship of all layers of neuron granting and burst length, description and mental picture feature in the form of connection weight.Method of the invention can describe the image of memory plurality of classes, and can completely restore image, while also having the function of image classification.
Description
Technical field
The present invention relates to technical field of computer vision, in particular to a kind of characteristics of image based on impulsive neural networks is retouched
It states and accumulating method.
Background technique
Computer vision is using computer and relevant device simulation biological vision, and final goal in research is to make computer
Visual observation can be passed through as people and understand the world, there is autonomous adaptability to environment.Currently, computer vision is answered extensively
For fields such as industry, military affairs, concrete application includes robot path planning, unmanned plane investigation, from main action etc..However, wanting
Realize above-mentioned application, the most basic and important research contents of one of them is image classification and identification in computer vision.Its
Research Thinking is: firstly, designing description and the accumulating method of a kind of characteristics of image；Then, with this method description and memory training
Image, and record description and memory result；Finally, with the description of same method and recall tests image, and record description and note
Recall as a result, the description of training image and test image and memory result are compared, final realization image classification and identification.It can
To find out, from largely saying, the description of characteristics of image and accumulating method determine the effect of image classification and identification.
It is a series of regional areas that traditional image classification and recognizer, which is by picture breakdown, is extracted in regional area
Characteristics of image, is described and mental picture with all local feature set.In the above-mentioned methods, the selection of feature and extract more
It is cumbersome, and calculate complicated.Later, there is researcher that convolutional neural networks are applied to iamge description and memory, change to a certain extent
It has been apt to the complexity of feature extraction in traditional iamge description and accumulating method.Convolutional neural networks have used for reference human visual system
Hierarchical structure and local receptor field characteristic, each neuron indicates a kind of feature in network, and upper layer feature is by a certain office of lower layer
It is obtained after the linear combination of feature and nonlinear transformation in portion region.By back-propagation algorithm training, the network can mention automatically
It takes on most influential feature of classifying, thus such methods largely improve image classification accuracy rate.
However, either describing the combination of image local feature using collection approach or linear combination method, all can
Lose feature between relative position information, lead to not according to memory image is more completely restored, i.e., it is existing this
A little methods are complete not enough for the description and memory of image.In addition, these above-mentioned methods belong to the scope of supervised learning mostly,
And it is small, although can preferably describe the feature that training sample concentrates image after training, for training sample
The image of other classifications outside this collection generally can not be described preferably, need re -training.
Summary of the invention
The present invention is directed at least solve one of the technical problems existing in the prior art.
In view of this, the present invention need to provide it is a kind of based on impulsive neural networks characteristics of image description and accumulating method,
This method can describe the image of memory plurality of classes simultaneously, and can completely restore image, while also have image classification function
Energy.
According to one embodiment of present invention, a kind of characteristics of image description and memory based on impulsive neural networks is proposed
Method, comprising the following steps:
Input M normalized images；
The number of plies N of impulsive neural networks is determined according to the size of described image, wherein N is the integer greater than 1；
Described image is pre-processed, obtains the gradient direction in described image at each pixel, and by the ladder
Discrete--direction is spent for preset value value；
Determine that first layer is a per the preset value in the impulsive neural networks according to the gradient direction after discretization
A granting in neuron；
The film potential of second layer neuron is calculated, according to the granting situation of the first layer neuron with determination described second
The granting situation of layer neuron, and then obtain the granting situation of all layers of neuron；
According to the sequential relationship and STDP (Spike-Timing-Dependent of all layers of neuron granting
Plasticity, the synaptic plasticity that the burst length relies on) connection weight between each layer of impulsive neural networks described in rule adjustment
Weight describes in the form of the connection weight and remembers described image feature.
According to one embodiment of present invention, the pulse nerve net is determined according to the size of described image and the number of plies
The size of each layer receptive field of network, wherein the receptive field size of first layer neuron is a pixel in the impulsive neural networks.
According to one embodiment of present invention, described that described image is pre-processed, obtain each picture in described image
The method of gradient direction at vegetarian refreshments is using one of Prewitt operator, Sobel operator, Kirsch operator and compass operator.
According to one embodiment of present invention, the preset value is 4.
It according to one embodiment of present invention, include the unit networks of N-1 seed type in N layers of the network, wherein institute
Stating unit networks is to be located at the network that the partial nerve member in continuous two layers is constituted by the N layer network.
According to one embodiment of present invention, the granting neuron presses Poisson process granting.
According to one embodiment of present invention, described that second layer mind is calculated according to the granting situation of the first layer neuron
Film potential through member, with the granting situation of the determination second layer neuron, the specific steps are as follows:
Using the random number of independent identically distributed exponential distribution as Time Of Release interval；
It is obtained providing the moment according to the Time Of Release interval；
With one millisecond for time step, by the granting moment discretization；
According to the granting moment after the discretization, the granting pulse y of the first layer neuron is generatedi(t):
According to the yi(t) it is calculated by the following formula the film potential u of the second layer neuronk(t):
Wherein,wkiFor the first layer neuron and second layer mind
Connection weight through member, yiFor the first layer neuron, 1≤i≤n, n are the number of the first layer neuron, 1≤k≤
K, K are the number of the second layer neuron, and σ is the yiThe unit pulse issued when granting to corresponding second layer neuron
Width；
By for the first layer neuron provide corresponding second layer neuron add WTA (Winner-Take-All,
Winner-take-all) constraint, select the neuron that the highest second layer neuron of film potential is the second layer granting.
According to one embodiment of present invention, described to be advised according to the sequential relationship and STDP of all layers of neuron granting
Then adjust the connection weight between each layer of the impulsive neural networks, the specific steps are as follows:
When a upper layer neuron is provided, the granting moment is denoted as tf, according to all lower layer's neurons of STDP rule adjustment
With the connection weight of upper layer neuron:
Judge upper layer neuron in [tf-σ,tf] in whether have granting；
If so, increasing the connection weight, if it is not, then reducing the connection weight.
The characteristics of image based on impulsive neural networks of the embodiment of the present invention describes and accumulating method, according to input picture
Size determines the number of plies of impulsive neural networks, and is determined first in impulsive neural networks according to the image gradient direction after discretization
A granting in the every preset value neuron of layer, the film of second layer neuron is calculated according to the granting situation of first layer neuron
Current potential obtains the granting situation of all layers of neuron to determine the granting situation of second layer neuron, according to all layers of neuron
Connection weight between the sequential relationship and each layer of STDP rule adjustment impulsive neural networks of granting, is retouched in the form of connection weight
It states and mental picture feature.Method of the invention can describe the image of memory plurality of classes simultaneously, and can completely restore to scheme
Picture, while also having the function of image classification.
Additional aspect and advantage of the invention will be set forth in part in the description, and will partially become from the following description
Obviously, or practice through the invention is recognized.
Detailed description of the invention
Above-mentioned and/or additional aspect of the invention and advantage will become from the description of the embodiment in conjunction with the following figures
Obviously and it is readily appreciated that, in which:
Fig. 1 is to be described and accumulating method according to the characteristics of image based on impulsive neural networks of one embodiment of the invention
Flow chart；
Fig. 2 is the unit networks structural schematic diagram according to one embodiment of the invention；
Fig. 3 is the hair that lower layer's neuron is determined according to the granting situation of upper layer neuron according to one embodiment of the invention
The to one's heart's content flow chart of the method for condition；
Fig. 4 is the description schematic diagram according to the impulsive neural networks third layer of one embodiment of the invention to input picture；
Fig. 5 is the flow chart according to the method for the adjustment link weight of one embodiment of the invention；
Fig. 6 is the connection and granting schematic diagram according to neuron in the unit networks of one embodiment of the invention；
Fig. 7 is the impulsive neural networks structural representation for describing and remembering according to the characteristics of image of one embodiment of the invention
Figure；
Fig. 8 is the memory schematic diagram according to the impulsive neural networks of one embodiment of the invention to facial image.
Specific embodiment
Below with reference to the accompanying drawings characteristics of image description and note according to an embodiment of the present invention based on impulsive neural networks is described
Method is recalled, wherein same or similar label indicates same or similar element or with the same or similar functions from beginning to end
Element.The embodiments described below with reference to the accompanying drawings are exemplary, for explaining only the invention, and should not be understood as to this
The limitation of invention.
The embodiment of the present invention propose it is a kind of based on impulsive neural networks characteristics of image description and accumulating method.
Fig. 1 is to be described and accumulating method according to the characteristics of image based on impulsive neural networks of one embodiment of the invention
Flow chart.
As shown in Figure 1, the characteristics of image based on impulsive neural networks of the embodiment of the present invention describes and accumulating method, including
Following steps:
S101, M normalized images of input.
Wherein, M is integer, can be 40.Image category can be a variety of, such as face, landscape, animal etc..But image
Size is identical, such as can be 112 × 96.
S102 determines the number of plies N of impulsive neural networks according to the size of image.
Wherein, N is the integer greater than 1, such as can be 4.
It in an embodiment of the present invention, include the unit networks of N-1 seed type in N layers of network, wherein unit networks are
The network that partial nerve member in continuous two layers is constituted is located at by N layer network.
Further, in one embodiment of the invention, pulse mind can also be determined according to the size and the number of plies of image
Size through each layer receptive field of network, wherein the receptive field size of first layer neuron is a pixel in impulsive neural networks.
Specifically, the facial image of 40 112 × 96 sizes can be chosen as input, four layers of arteries and veins can be used at this time
Neural network is rushed to be described and remember, the receptive field size of bottom-up each layer neuron is successively in network are as follows: 1 × 1,7
×6、28×24、112×96。
S103 pre-processes image, obtains the gradient direction in image at each pixel, and by gradient direction from
Dispersion is preset value value.
Wherein, preset value can be 4.
In an embodiment of the present invention, the receptive field of first layer (bottom) neuron is single picture in impulsive neural networks
Element, they are sensitive to the edge feature of different directions, can be by carrying out pretreatment realization to input picture.
It specifically, can be using one of Prewitt operator, Sobel operator, Kirsch operator and compass operator to figure
As being pre-processed, for example, the gradient direction in image at each pixel first can be obtained with Sobel operator, further according to needs
By gradient direction discretization.It is sensitive if necessary to edge feature of the first layer neuron to four kinds of different directions, so that it may will be each
Gradient direction at pixel is discrete to turn to four values.
S104 is determined in impulsive neural networks in the every preset value neuron of first layer according to the gradient direction after discretization
One granting.
Wherein, it provides neuron and presses Poisson process granting.
In one particular embodiment of the present invention, each pixel can correspond to four first layer neurons in image,
According to the gradient direction after discretization at each pixel, one in every four neurons can be correspondingly set to provide.In order to convex
Marginal information in aobvious image can use the edge of Canny operator extraction image, and then only allow image edge pixels point corresponding
Neuron has granting, and other corresponding neurons of pixel for being not belonging to image border are not provided.
S105 calculates the film potential of second layer neuron, according to the granting situation of first layer neuron to determine the second layer
The granting situation of neuron, and then obtain the granting situation of all layers of neuron.
In an embodiment of the present invention, impulsive neural networks can actually be that cascade structure is combined by a series of unit networks
At, as shown in Figure 2.
It should be noted that sharing the unit networks of three types in four layers of pulse network structure, they distinguish
It is the unit networks being made of 7 × 6 × 4 first layer neurons and 40 second layer neurons, by 4 × 4 × 40 second layers
Unit networks that neuron and 40 third layer neurons are constituted, by 4 × 4 × 40 third layer neurons and 40 the 4th layer of minds
The unit networks constituted through member.Their implementation method be it is similar, can be said by taking first kind unit networks as an example below
It is bright:
In first kind unit networks, second layer neuron indicates that some specific position of input picture, size are 7 × 6
Regional area in the feature that is likely to occur, these are characterized in the feature indicated by first layer neuron by certain positional relationship
What combination was formed.If totally 40 images, the feature that each regional area is likely to occur is no more than 40 kinds, with 40 second layers
Neuron is enough.
As shown in figure 3, calculating the film potential of second layer neuron, according to the granting situation of first layer neuron to determine the
The granting situation of two layers of neuron, the specific steps are as follows:
S201, using the random number of independent identically distributed exponential distribution as Time Of Release interval.
In an embodiment of the present invention, every input picture can all maintain a period of time, sufficiently to remember.Scheme in input
As that can determine which neuron of first layer can be provided by aforementioned preprocess method in constant this period, make these nerves
Member presses Poisson process granting.Specifically, the independent random number with exponential distribution can be generated as Time Of Release interval, wherein
The parameter of exponential distribution can be 0.08.
S202 obtains providing the moment according to Time Of Release interval.
S203 will provide moment discretization with one millisecond for time step.
S204 generates the granting pulse y of first layer neuron according to the granting moment after discretizationi(t)。
S205, according to yi(t) the film potential u of second layer neuron is calculated by formula (1)k(t):
Wherein,wkiFor the company of first layer neuron and second layer neuron
Meet weight, yiFor first layer neuron, 1≤i≤n, n are the number of the first layer neuron, and 1≤k≤K, K are second layer mind
Number through member, σ yiThe unit pulse width issued when granting to corresponding second layer neuron can be 25.At the beginning of weight
Small random number can be used in value, such as equally distributed random number in [0,1].
S206 adds WTA constraint by providing corresponding second layer neuron for first layer neuron, selects film potential most
High second layer neuron is the neuron of second layer granting.
In an embodiment of the present invention, it is not necessary to which one millisecond of every mistake just calculates a uk(t) value, because of second layer neuron
Group total granting rate be it is constant, and the short period in only one second layer neuron granting, therefore can by preceding method elder generation
The granting moment for calculating second layer neuron pool calculates the film potential of each neuron of the second layer at these moment, allows film potential
Highest neuron issues a pulse at the corresponding moment.
For example, Fig. 4 is the granting of third layer neuron in the impulsive neural networks of a specific embodiment of the invention
Situation schematic diagram, abscissa indicate the time, and ordinate indicates the number of neuron.Here 40 input pictures, every figure are shared
As 5000 milliseconds, third layer neuron totally 60 of study.Through overfitting, every image can be by the hair of single neuron in third layer
It puts to describe, this is it is also assumed that be a kind of classification to input picture, that 20 neurons without granting are remained to new
Input picture descriptive power.The scale of neuron is bigger, and the ability for describing a large amount of multi-class images is stronger.
S106, according between the sequential relationship and each layer of STDP rule adjustment impulsive neural networks of all layers of neuron granting
Connection weight, in the form of connection weight description and mental picture feature.
In an embodiment of the present invention, as shown in figure 5, according to the sequential relationship and STDP rule of all layers of neuron granting
Adjust the connection weight between each layer of impulsive neural networks, the specific steps are as follows:
S301, when a upper layer neuron is provided, the granting moment is denoted as tf, according to all lower layer's minds of STDP rule adjustment
Connection weight through member with upper layer neuron.
S302 judges upper layer neuron in [tf-σ,tf] in whether have granting.
S303, if so, increasing connection weight, if it is not, then reducing connection weight.
In an embodiment of the present invention, after carrying out above-mentioned unsupervised learning to connection weight, each output neuron can
It is gradually sensitive to the certain specific response modes of input neuron pool.
Specifically, Fig. 6 is the connection relationship diagram that neuron and output neuron are inputted in a unit networks.Fig. 6
Middle left side is input neuron, and the neuron in each frame is one group (receptive field i.e. having the same), and right side is output nerve
Member, each are connected with all input neurons, and receptive field is spliced by the receptive field of each group input neuron.Assuming that
When inputting an image, z2The film potential highest of neuron then only has z whithin a period of time in output neuron2It provides and (utilizes
WTA mechanism), the connection weight that heavy black line in figure marks need to be adjusted by STDP rule, institute can be adjusted by formula (2)
There is associated connection weight wki(1≤i≤n=7 × 6 × 4):
wki=wki+ηki·Δwki (2)
Wherein, Δ wki=yi(tf)·exp(-wki) -1, ηkiIt can use lesser constant, such as 0.01, it can also be according to wkiChange
Change and is adaptively adjusted by formula (3), such as
Weight adjustment the result is that break up second layer neuron, i.e., each second layer in the unit networks is refreshing
Through member can different characteristic to same image-region it is sensitive.
The granting moment of all second layer neurons is recorded, so that it may realize more top nerve with similar method
The granting of member and weight study.Finally, entire impulsive neural networks describe input picture by the granting of each layer neuron, and
The feature of each scale of image has been remembered in the form of connection weight.
The characteristics of image based on impulsive neural networks of the embodiment of the present invention describes and accumulating method, according to input picture
Size determines the number of plies N of impulsive neural networks, and is determined in impulsive neural networks according to the image gradient direction after discretization
A granting in one layer of every preset value neuron, calculates second layer neuron according to the granting situation of first layer neuron
Film potential obtains the granting situation of all layers of neuron to determine the granting situation of second layer neuron, according to all layers of nerve
Connection weight between the sequential relationship that member is provided and each layer of STDP rule adjustment impulsive neural networks, in the form of connection weight
Description and mental picture feature.Method of the invention can describe the image of memory plurality of classes simultaneously, and can completely restore
Image, while also having the function of image classification.
It, can be with for convenience of the characteristics of image description based on impulsive neural networks and accumulating method for understanding the embodiment of the present invention
By Fig. 7, present invention is described and is described in detail.
As shown in fig. 7, dot indicates neuron, they have the structure of stratification.In each layer, neuron is according to it
The difference of receptive field is divided into several groups, and the receptive field of each group neuron can cover entire image, and mind without overlapping
Arrangement through member is corresponding with the position of its receptive field.Each layer of receptive field is divided with dotted line in figure, it should be noted that
The receptive field of top neuron covers entire image, so not dividing.
Smaller in the receptive field of the bottom of impulsive neural networks, neuron, these neurons are special to some simple parts
It is more sensitive to levy (such as edge feature)；It is every to rise one layer, if the receptive field of neuron is by being located below one layer of close position
The receptive field of dry group neuron is spliced.As shown in fig. 7, have 2 × 2 groups of neurons, every to rise one layer, the receptive field of neuron
Expand four times greater (in fact, the expansion multiple of receptive field can be different in different levels).At this point, 2 × 2 groups of neurons of lower layer
Full connection is established between one group of neuron on upper layer, and is not connected between other each group neurons of upper layer.
By adding WTA constraint, when input picture can be made constant, only one neuron in every group for every group of neuron
Response is more strong, and the neuron of such upper layer response can know exactly which the position between several neurons of lower layer's response
Relationship is set, so as to naturally enough represent the complex characteristic that the sensitive local feature of lower layer's neuron is spliced to form.
It further, can be non-supervisoryly to interneuronal connection weight according to STDP rule according to the response of neuron
It is adjusted again, finally makes each neuron only to certain feature-sensitive of image in its receptive field, and with minds different in group
It is sensitive to different characteristics of image through member.Neuron is embodied in its connection weight the memory of characteristics of image, connects in adjustment
When weight, similar feature (feature responded by the same neuron) is classified as one kind automatically.Again since network is in each layer
Location information is all remained, so the network can carry out input picture (cat in such as Fig. 7) on each abstraction hierarchy
Whole description.Moreover, using the memory of the network, i.e., connection relationship between each layer neuron can also generate new
Significant image.As shown in figure 8, to be reconstructed using the connection weight after the study of the impulsive neural networks of the embodiment of the present invention
The partial face image come illustrates that the network has more complete and clearly memory capability.
It should be noted that the response due to output neuron in unit networks is similar with the input response of neuron, i.e.,
At most there is a neuron granting in the wild identical neuron pool of a period of time enteroception, so unit networks combining cascade shape
Be at hierarchical structure as shown in Figure 7 it is reasonable and natural, only need to be by the output of several unit networks of lower layer mind in cascade
Input neuron through member as upper one layer of certain unit networks.Within the same layer, the receptive field of different units network is not mutually
It is overlapping, the different piece of their independent process input pictures, thus it is whole that network can be substantially improved by the way of parallel computation
The processing speed of body.Although the various spies being likely to occur in a receptive field are indicated in impulsive neural networks with a large amount of neurons
Sign causes computation complexity high to guarantee there is stronger descriptive power for various types of other image, but since weight learns
Used STDP rule is a kind of local learning strategy, and can carry out parallel computation, so when describing image in processing speed
Larger sacrifice is not had on degree.
Any process described otherwise above or method description are construed as in flow chart or herein, and expression includes
It is one or more for realizing specific logical function or process the step of executable instruction code module, segment or portion
Point, and the range of the preferred embodiment of the present invention includes other realization, wherein can not press shown or discussed suitable
Sequence, including according to related function by it is basic simultaneously in the way of or in the opposite order, to execute function, this should be of the invention
Embodiment person of ordinary skill in the field understood.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
Centainly refer to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be any
One or more embodiment or examples in can be combined in any suitable manner.
Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that: not
A variety of change, modification, replacement and modification can be carried out to these embodiments in the case where being detached from the principle of the present invention and objective, this
The range of invention is by claim and its equivalent limits.
Claims (8)
1. a kind of characteristics of image description and accumulating method based on impulsive neural networks, which comprises the following steps:
Input M normalized images；
The number of plies N of impulsive neural networks is determined according to the size of described image, wherein N is the integer greater than 1；
Described image is pre-processed, obtains the gradient direction in described image at each pixel, and by the gradient side
Preset value value is turned to discrete；
Determine that first layer is per the preset value nerve in the impulsive neural networks according to the gradient direction after discretization
A granting in member；
The film potential of second layer neuron is calculated, according to the granting situation of the first layer neuron with the determination second layer mind
Granting situation through member, and then obtain the granting situation of all layers of neuron；
According between each layer of impulsive neural networks described in the sequential relationship of all layers of neuron granting and STDP rule adjustment
Connection weight, in the form of the connection weight describe and remember described image feature.
2. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that root
The size of each layer receptive field of the impulsive neural networks is determined according to the size and the number of plies of described image, wherein the pulse
The receptive field size of first layer neuron is a pixel in neural network.
3. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that institute
It states and described image is pre-processed, the method for obtaining the gradient direction in described image at each pixel uses Prewitt
One of operator, Sobel operator, Kirsch operator and compass operator.
4. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that institute
Stating preset value is 4.
5. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that institute
State the unit networks in N layers of network comprising N-1 seed type, wherein the unit networks are to be located at continuously by the N layer network
Two layers in partial nerve member constitute network.
6. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that institute
It states and provides neuron by Poisson process granting.
7. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that institute
The film potential that second layer neuron is calculated according to the granting situation of the first layer neuron is stated, with the determination second layer nerve
The granting situation of member, the specific steps are as follows:
Using the random number of independent identically distributed exponential distribution as Time Of Release interval；
It is obtained providing the moment according to the Time Of Release interval；
With one millisecond for time step, by the granting moment discretization；
According to the granting moment after the discretization, the granting pulse y of the first layer neuron is generatedi(t):
According to the yi(t) it is calculated by the following formula the film potential u of the second layer neuronk(t):
Wherein,wkiFor the first layer neuron and the second layer neuron
Connection weight, yiFor the first layer neuron, 1≤i≤n, n are the number of the first layer neuron, and 1≤k≤K, K are
The number of the second layer neuron, σ are the yiThe unit pulse width issued when granting to corresponding second layer neuron；
WTA constraint is added by providing corresponding second layer neuron for the first layer neuron, it is highest to select film potential
The second layer neuron is the neuron of the second layer granting.
8. the characteristics of image description based on impulsive neural networks and accumulating method as described in claim 1, which is characterized in that institute
It states between each layer of impulsive neural networks according to the sequential relationship of all layers of neuron granting and STDP rule adjustment
Connection weight, the specific steps are as follows:
When a upper layer neuron is provided, the granting moment is denoted as tf, according to all lower layer's neurons of STDP rule adjustment and the hair
The connection weight for the upper layer neuron put:
To each lower layer's neuron, judge it in [tf-σ,tf] in whether have granting；
If so, increasing the connection weight of the upper layer neuron of lower layer's neuron and the granting, if it is not, then reducing the lower layer
The connection weight of neuron and the upper layer neuron of the granting.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510708874.7A CN105404902B (en) | 2015-10-27 | 2015-10-27 | Characteristics of image description and accumulating method based on impulsive neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510708874.7A CN105404902B (en) | 2015-10-27 | 2015-10-27 | Characteristics of image description and accumulating method based on impulsive neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105404902A CN105404902A (en) | 2016-03-16 |
| CN105404902B true CN105404902B (en) | 2019-02-05 |
Family
ID=55470377
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510708874.7A Active CN105404902B (en) | 2015-10-27 | 2015-10-27 | Characteristics of image description and accumulating method based on impulsive neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105404902B (en) |
Families Citing this family (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3537344B1 (en) | 2016-11-28 | 2024-01-24 | Huawei Technologies Co., Ltd. | Signal processing method and device based on spiking neural network |
| CN106845541A (en) * | 2017-01-17 | 2017-06-13 | 杭州电子科技大学 | A kind of image-recognizing method based on biological vision and precision pulse driving neutral net |
| CN107256554B (en) * | 2017-05-18 | 2020-08-04 | 西安电子科技大学 | Single-layer pulse neural network structure for image segmentation |
| CN107194426A (en) * | 2017-05-23 | 2017-09-22 | 电子科技大学 | A kind of image-recognizing method based on Spiking neutral nets |
| CN108496188A (en) * | 2017-05-31 | 2018-09-04 | 深圳市大疆创新科技有限公司 | Method, apparatus, computer system and the movable equipment of neural metwork training |
| TWI636404B (en) * | 2017-07-31 | 2018-09-21 | 財團法人工業技術研究院 | Deep neural network and method for using the same and computer readable media |
| CN108304767B (en) * | 2017-12-12 | 2020-06-30 | 中国科学院自动化研究所 | Human action intention recognition training method based on multi-brain-area collaborative computing |
| CN108446757B (en) * | 2017-12-29 | 2020-09-01 | 北京理工大学 | Multi-stage delay cascade speed identification system and method based on impulse neural network |
| CN108364262A (en) * | 2018-01-11 | 2018-08-03 | 深圳大学 | A kind of restored method of blurred picture, device, equipment and storage medium |
| CN110443363B (en) * | 2018-05-04 | 2022-06-07 | 北京市商汤科技开发有限公司 | Image feature learning method and device |
| CN108846349A (en) * | 2018-06-08 | 2018-11-20 | 电子科技大学 | A kind of face identification method based on dynamic Spiking neural network |
| CN109214395A (en) * | 2018-08-21 | 2019-01-15 | 电子科技大学 | A kind of new image representation method based on impulsive neural networks |
| CN109142358B (en) * | 2018-08-22 | 2020-09-29 | 王天骄 | Metal material mechanical property testing method based on neural network |
| WO2020048814A1 (en) * | 2018-09-05 | 2020-03-12 | Daimler Ag | Method for identifying a hand pose in a vehicle |
| CN110059812B (en) * | 2019-01-26 | 2021-09-14 | 中国科学院计算技术研究所 | Pulse neural network operation chip and related operation method |
| CN110111234B (en) * | 2019-04-11 | 2023-12-15 | 上海集成电路研发中心有限公司 | Image processing system architecture based on neural network |
| CN110210563B (en) * | 2019-06-04 | 2021-04-30 | 北京大学 | Image pulse data space-time information learning and identification method based on Spike cube SNN |
| CN110826602B (en) * | 2019-10-23 | 2022-04-26 | 中国科学院自动化研究所 | Image classification method and system based on membrane potential regulation and control pulse neural network |
| CN110781968B (en) * | 2019-10-29 | 2022-04-19 | 中国人民解放军国防科技大学 | Extensible class image identification method based on plastic convolution neural network |
| CN111723726B (en) * | 2020-06-17 | 2023-09-19 | 苏州大学 | Identification method of pulse neural membrane system drying agent packaging bag identification model |
| CN111811532B (en) * | 2020-07-02 | 2022-03-25 | 浙江大学 | Path planning method and device based on impulse neural network |
| CN112232440B (en) * | 2020-11-10 | 2022-11-11 | 北京理工大学 | Method for realizing information memory and distinction of impulse neural network by using specific neuron groups |
| CN113673555B (en) * | 2021-07-09 | 2023-12-12 | 浙江大学 | Memory-based unsupervised domain adaptive picture classification method |
| CN113642535B (en) * | 2021-10-13 | 2022-01-25 | 聊城高新生物技术有限公司 | Biological branch detection method and device and electronic equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101546430A (en) * | 2009-04-30 | 2009-09-30 | 上海大学 | Edge extracting method based on simplified pulse coupled neural network |
| CN102867296A (en) * | 2012-08-20 | 2013-01-09 | 华中科技大学 | Medical image fusion method based on pulsed cortical model |
-
2015
- 2015-10-27 CN CN201510708874.7A patent/CN105404902B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101546430A (en) * | 2009-04-30 | 2009-09-30 | 上海大学 | Edge extracting method based on simplified pulse coupled neural network |
| CN102867296A (en) * | 2012-08-20 | 2013-01-09 | 华中科技大学 | Medical image fusion method based on pulsed cortical model |
Non-Patent Citations (3)
| Title |
|---|
| Image segmentation with pulse-coupled neural network and Canny operators;Wen Jiang et al;《Computers and Electrical Engineering》;20150415;第528-538页 |
| NSCT-based multimodal medical image fusion using pulse-coupled neural network and modified spatial frequency;Sudeb Das et al;《Med Biol Eng Comput》;20120724;第1105-1114页 |
| 基于脉冲神经网络边缘提取的虹膜定位;黄敏 等;《电子测量技术》;20120831;第35卷(第8期);第49-52页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105404902A (en) | 2016-03-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105404902B (en) | Characteristics of image description and accumulating method based on impulsive neural networks | |
| Jozwik et al. | Deep convolutional neural networks outperform feature-based but not categorical models in explaining object similarity judgments | |
| Majdi et al. | Drive-net: Convolutional network for driver distraction detection | |
| CN109558942B (en) | Neural network migration method based on shallow learning | |
| CN106202997B (en) | A kind of cell division detection method based on deep learning | |
| CN109583322B (en) | Face recognition deep network training method and system | |
| Dewan et al. | A deep learning approach to detecting engagement of online learners | |
| US7711663B2 (en) | Multi-layer development network having in-place learning | |
| Masood et al. | Prediction of human ethnicity from facial images using neural networks | |
| CN108491858A (en) | Method for detecting fatigue driving based on convolutional neural networks and system | |
| CN106803069A (en) | Crowd's level of happiness recognition methods based on deep learning | |
| Cadena et al. | How well do deep neural networks trained on object recognition characterize the mouse visual system? | |
| CN110490136A (en) | A kind of human body behavior prediction method of knowledge based distillation | |
| Yamashita et al. | Hand posture recognition based on bottom-up structured deep convolutional neural network with curriculum learning | |
| CN106951923A (en) | A kind of robot three-dimensional shape recognition process based on multi-camera Vision Fusion | |
| Ji et al. | Where-what network 1:“Where” and “What” assist each other through top-down connections | |
| Mermillod et al. | Neural computation as a tool to differentiate perceptual from emotional processes: The case of anger superiority effect | |
| Ou et al. | Design of an end-to-end dual mode driver distraction detection system | |
| CN109508640A (en) | A kind of crowd's sentiment analysis method, apparatus and storage medium | |
| Berger et al. | Visual tracking with vg-ram weightless neural networks | |
| Guo et al. | WWN-9: Cross-domain synaptic maintenance and its application to object groups recognition | |
| CN110009038A (en) | Training method, device and the storage medium of screening model | |
| Gaddam et al. | On sudoku problem using deep learning and image processing technique | |
| Zhang et al. | An interpretable image classification model Combining a fuzzy neural network with a variational autoencoder inspired by the human brain | |
| Romano et al. | An improved pooling scheme for convolutional neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |