CN105404902A - Impulsive neural network-based image feature describing and memorizing method - Google Patents
Impulsive neural network-based image feature describing and memorizing method Download PDFInfo
- Publication number
- CN105404902A CN105404902A CN201510708874.7A CN201510708874A CN105404902A CN 105404902 A CN105404902 A CN 105404902A CN 201510708874 A CN201510708874 A CN 201510708874A CN 105404902 A CN105404902 A CN 105404902A
- Authority
- CN
- China
- Prior art keywords
- layer
- image
- neuronic
- granting
- 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
- 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 invention provides an impulsive neural network-based image feature describing and memorizing method. the method comprises steps: M normalized images are inputted, the layer number of the impulsive neural network is determined according to the size of the image, a gradient direction at each pixel point is acquired when pretreatment is carried out on the images, the gradient direction is discretized into a preset individual value, distribution of one of each preset value number of neurons in the first layer in the impulsive neural network is determined according to the discretized gradient direction, membrane potential of neurons in the second layer and the distribution condition of the neurons in the second layer are calculated according to the distribution condition of the neurons in the first layer, the distribution conditions of the neurons in all layers are obtained, a connection weight of each layer of the impulsive neural network is adjusted according to a timing relationship for distribution of neurons in all layers and a STDP (Spike Timing-dependent Plasticity) rule, and the image features are described and memorized in a connection weight form. The method of the invention can describe and memorize images of various kinds, can completely restore an image, and also has an image classification function.
Description
Technical field
The present invention relates to technical field of computer vision, particularly a kind of characteristics of image based on impulsive neural networks describes and accumulating method.
Background technology
Computer vision uses computing machine and relevant device simulation biological vision, and its final goal in research makes calculating functional image people equally by Visual Observations Observations and understand the world, has autonomous adaptive faculty to environment.At present, computer vision is widely used in the field such as industry, military affairs, and embody rule comprises robot path planning, unmanned plane is investigated, from main action etc.But will realize above-mentioned application, wherein research contents basic and the most important is Images Classification in computer vision and identification.Its Research Thinking is: first, designs a kind of description and accumulating method of characteristics of image; Then, describe and memory training image by the method, and record description and memory result; Finally, use the same method and describe and recall tests image, and record description and memory result, by the description of training image and test pattern with remember result and compare, finally realize Images Classification and identification.Can find out, to a great extent, the description of characteristics of image and accumulating method determine the effect of Images Classification and identification.
Traditional Images Classification and recognizer are a series of regional area by picture breakdown, extracts the characteristics of image in regional area, describe and mental picture with all local characteristic set.In the above-mentioned methods, the selection of feature is comparatively loaded down with trivial details with extraction, and calculation of complex.Afterwards, there is researcher that convolutional neural networks is applied to iamge description and memory, improved the complexity of feature extraction in traditional iamge description and accumulating method to a certain extent.Convolutional neural networks has used for reference hierarchical structure and the local receptor field characteristic of human visual system, and in network, each neuron represents a kind of feature, and upper strata feature obtains by after the linear combination of feature in a certain regional area of lower floor and nonlinear transformation.Train this network can automatically extract the most influential feature of classification by back-propagation algorithm, thus these class methods improve Images Classification accuracy rate to a great extent.
But, no matter be the combination adopting collection approach or linear combination method to carry out Description Image local feature, the relative position information between feature is lost in capital, cause cannot according to memory more completely recover image, namely these methods existing for image description and memory complete not enough.In addition, these methods above-mentioned belong to the category of supervised learning mostly, and small, although can describe the feature that training sample concentrates image preferably after training, but usually can not describe preferably for the image of other classifications outside training sample set, need re-training.
Summary of the invention
The present invention is intended at least to solve one of technical matters existed in prior art.
In view of this, the present invention needs to provide a kind of characteristics of image based on impulsive neural networks to describe and accumulating method, and the method can describe the image of memory plurality of classes simultaneously, and can intactly Recovery image, also has Images Classification function simultaneously.
According to one embodiment of present invention, propose a kind of characteristics of image based on impulsive neural networks and describe and accumulating method, comprise the following steps:
Input M opens normalized image;
According to the number of plies N of the size determination impulsive neural networks of described image, wherein, N be greater than 1 integer;
Pre-service is carried out to described image, obtains the gradient direction at each pixel place in described image, and turn to a preset value value by discrete for described gradient direction;
A granting in described impulsive neural networks in the every described preset value neuron of ground floor is determined according to the described gradient direction after discretize;
Calculate the neuronic film potential of the second layer according to described ground floor neuronic granting situation, to determine the neuronic granting situation of the described second layer, and then obtain the neuronic granting situation of all layers;
The sequential relationship provided according to described all layer neurons and STDP (Spike-Timing-DependentPlasticity, the synaptic plasticity that burst length relies on) connection weight described in rule adjustment between each layer of impulsive neural networks, describe with the form of described connection weight and remember described characteristics of image.
According to one embodiment of present invention, determine the size of each layer receptive field of described impulsive neural networks according to the size of described image and the described number of plies, wherein, in described impulsive neural networks, the neuronic receptive field size of ground floor is a pixel.
According to one embodiment of present invention, describedly carry out pre-service to described image, the method obtaining the gradient direction at each pixel place in described image adopts the one in Prewitt operator, Sobel operator, Kirsch operator and compass operator.
According to one embodiment of present invention, described preset value is 4.
According to one embodiment of present invention, comprise the unit networks of N-1 type in the network of described N layer, wherein, described unit networks is be arranged in the first network formed of the two-layer partial nerve of continuous print by described N layer network.
According to one embodiment of present invention, described granting neuron is by Poisson process granting.
According to one embodiment of present invention, described according to the neuronic film potential of the described ground floor neuronic granting situation calculating second layer, to determine the neuronic granting situation of the described second layer, concrete steps are as follows:
Using the random number of independent identically distributed exponential distribution as Time Of Release interval;
Obtain providing the moment according to described Time Of Release interval;
With one millisecond for time step, by described granting moment discretize;
According to the granting moment after described discretize, generate described ground floor neuronic granting pulse y
i(t);
According to described y
it () is by the neuronic described film potential u of the second layer described in following formulae discovery
k(t):
Wherein,
w
kifor described ground floor neuron and the neuronic connection weight of the described second layer, y
ifor described ground floor neuron, 1≤i≤n, n is the neuronic number of described ground floor, and 1≤k≤K, K is the neuronic number of the described second layer, and σ is described y
ito the unit pulse width that the second layer neuron of correspondence sends during granting;
By adding WTA (Winner-Take-All, winner-take-all) constraint for the second layer neuron of described ground floor neuron granting correspondence, select the neuron that the highest described second layer neuron of film potential is the granting of the described second layer.
According to one embodiment of present invention, the connection weight described in the described sequential relationship provided according to described all layer neurons and STDP rule adjustment between each layer of impulsive neural networks, concrete steps are as follows:
When a upper strata neuron is provided, the granting moment is denoted as t
f, according to the neuronic connection weight of STDP rule adjustment all lower floors neuron and upper strata:
Judge that upper strata neuron is at [t
f-σ, t
f] in whether have granting;
If so, then increase described connection weight, if not, then reduce described connection weight.
The characteristics of image based on impulsive neural networks of the embodiment of the present invention describes and accumulating method, according to the number of plies of the size determination impulsive neural networks of input picture, and determine a granting in impulsive neural networks in the every preset value of a ground floor neuron according to the image gradient direction after discretize, the neuronic film potential of the second layer is calculated according to ground floor neuronic granting situation, to determine the neuronic granting situation of the second layer, obtain the neuronic granting situation of all layers, connection weight between the sequential relationship provided according to all layer neurons and each layer of STDP rule adjustment impulsive neural networks, describe and mental picture feature with the form of connection weight.Method of the present invention can describe the image of memory plurality of classes simultaneously, and can intactly Recovery image, also has Images Classification function simultaneously.
Additional aspect of the present invention and advantage will part provide in the following description, and part will become obvious from the following description, or be recognized by practice of the present invention.
Accompanying drawing explanation
Above-mentioned and/or additional aspect of the present invention and advantage will become obvious and easy understand from accompanying drawing below combining to the description of embodiment, wherein:
Fig. 1 is according to an embodiment of the invention based on the characteristics of image description of impulsive neural networks and the process flow diagram of accumulating method;
Fig. 2 is unit networks structural representation according to an embodiment of the invention;
Fig. 3 is the process flow diagram of the method determining the neuronic granting situation of lower floor according to an embodiment of the invention according to upper strata neuronic granting situation;
Fig. 4 be according to an embodiment of the invention impulsive neural networks third layer to the description schematic diagram of input picture;
Fig. 5 is the process flow diagram of the method adjusting link weight according to an embodiment of the invention;
Fig. 6 is neuronic connection and provide schematic diagram in unit networks according to an embodiment of the invention;
Fig. 7 is the impulsive neural networks structural representation of characteristics of image description and memory according to an embodiment of the invention;
Fig. 8 be according to an embodiment of the invention impulsive neural networks to the memory schematic diagram of facial image.
Embodiment
Below with reference to the accompanying drawings describe and describe and accumulating method according to the characteristics of image based on impulsive neural networks of the embodiment of the present invention, wherein same or similar label represents same or similar element from start to finish or has element that is identical or similar functions.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.
Embodiments of the invention propose a kind of characteristics of image based on impulsive neural networks and describe and accumulating method.
Fig. 1 is according to an embodiment of the invention based on the characteristics of image description of impulsive neural networks and the process flow diagram of accumulating method.
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, comprises the following steps:
S101, input M opens normalized image.
Wherein, M is integer, can be 40.Image category can be multiple, such as face, landscape, animal etc.But image size is identical, such as, can be 112 × 96.
S102, according to the number of plies N of the size determination impulsive neural networks of image.
Wherein, N be greater than 1 integer, can be such as 4.
In an embodiment of the present invention, comprise the unit networks of N-1 type in the network of N layer, wherein, unit networks is be arranged in the first network formed of the two-layer partial nerve of continuous print by N layer network.
Further, in one embodiment of the invention, can also according to the size of the size of image and each layer receptive field of number of plies determination impulsive neural networks, wherein, in impulsive neural networks, the neuronic receptive field size of ground floor is a pixel.
Particularly, the facial image of 40 112 × 96 sizes can be chosen as input, four layers of impulsive neural networks now can be adopted to be described and to remember, and in network, the neuronic receptive field size of bottom-up each layer is followed successively by: 1 × 1,7 × 6,28 × 24,112 × 96.
S103, carries out pre-service to image, obtains the gradient direction at each pixel place in image, and turns to a preset value value by discrete for gradient direction.
Wherein, preset value can be 4.
In an embodiment of the present invention, in impulsive neural networks, ground floor (bottom) neuronic receptive field is single pixel, and they are responsive to the edge feature of different directions, can by carrying out pre-service realization to input picture.
Particularly, the one in Prewitt operator, Sobel operator, Kirsch operator and compass operator can be adopted to carry out pre-service to image, such as, the gradient direction at each pixel place in image first can be obtained with Sobel operator, more as required by gradient direction discretize.If need ground floor neuron responsive to the edge feature of four kinds of different directions, just four values can be turned to by discrete for the gradient direction at each pixel place.
S104, according to a granting in the every preset value of a ground floor neuron in the gradient direction determination impulsive neural networks after discretize.
Wherein, neuron is provided by Poisson process granting.
In one particular embodiment of the present invention, in image, each pixel can corresponding four ground floor neurons, according to the gradient direction after each pixel place discretize, correspondingly can make a granting in every four neurons.In order to the marginal information in saliency maps picture, can with the edge of Canny operator extraction image, and then only allow neuron corresponding to image edge pixels point have granting, and the neuron that other do not belong to the pixel of image border corresponding not all be provided.
S105, calculates the neuronic film potential of the second layer according to ground floor neuronic granting situation, to determine the neuronic granting situation of the second layer, and then obtains the neuronic granting situation of all layers.
In an embodiment of the present invention, in fact impulsive neural networks can combine cascade by a series of unit networks and form, as shown in Figure 2.
It should be noted that, in four layers of pulse network structure, the unit networks of total three types, they be respectively the unit networks be made up of 7 × 6 × 4 ground floor neurons and 40 second layer neurons, the unit networks be made up of 4 × 4 × 40 second layer neurons and 40 third layer neurons, by 4 × 4 × 40 third layer neurons and 40 the 4th layer of unit networks that neuron is formed.Their implementation method is similar, can be described below for first kind unit networks:
In first kind unit networks, second layer neuron represents certain ad-hoc location of input picture, size be 7 × 6 regional area in the feature that may occur, these features are that the feature represented by ground floor neuron is combined to form by certain position relationship.If totally 40 images, then the feature that each regional area may occur is no more than 40 kinds, is enough with 40 second layer neurons.
As shown in Figure 3, calculate the neuronic film potential of the second layer according to ground floor neuronic granting situation, to determine the neuronic granting situation of the second layer, concrete 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, often open input picture and all can maintain a period of time, fully to remember.Input picture constant during this period of time in, can determine which neuron of ground floor can be provided by aforementioned preprocess method, make these neurons by Poisson process granting.Particularly, can generate the independent random number with exponential distribution 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, with one millisecond for time step, will provide moment discretize.
S204, according to the granting moment after discretize, generates ground floor neuronic granting pulse y
i(t).
S205, according to y
it () through type (1) calculates the neuronic film potential u of the second layer
k(t):
Wherein,
w
kifor ground floor neuron and the neuronic connection weight of the second layer, y
ifor ground floor neuron, 1≤i≤n, n is the neuronic number of described ground floor, and 1≤k≤K, K is the neuronic number of the second layer, and σ is y
ito the unit pulse width that the second layer neuron of correspondence sends during granting, can be 25.Weight initial value can use little random number, as equally distributed random number in [0,1].
S206, by adding WTA constraint for the second layer neuron of ground floor neuron granting correspondence, selects the neuron that the highest second layer neuron of film potential is second layer granting.
In an embodiment of the present invention, need not often cross one millisecond and just calculate a u
kthe value of (t), because total granting rate of second layer neuron pool is constant, and in the short period, only have the granting of a second layer neuron, therefore the granting moment of second layer neuron pool first can be calculated by preceding method, calculate each neuronic film potential of the second layer in these moment, allow the highest neuron of film potential send a pulse in the corresponding moment.
Illustrate, Fig. 4 is third layer neuronic granting situation schematic diagram in the impulsive neural networks of the present invention's specific embodiment, and abscissa representing time, ordinate represents neuronic numbering.Here have 40 input pictures, often open image study 5000 milliseconds, third layer neuron totally 60.Through study, often open image and can be described by single neuronic granting in third layer, this also can think to classify to the one of input picture, and those 20 neurons do not provided remain the descriptive power to new input picture.Neuronic scale is larger, and its ability describing a large amount of multi-class image is stronger.
S106, the connection weight between the sequential relationship provided according to all layer neurons and each layer of STDP rule adjustment impulsive neural networks, describes and mental picture feature with the form of connection weight.
In an embodiment of the present invention, as shown in Figure 5, the connection weight between the sequential relationship provided according to all layer neurons and each layer of STDP rule adjustment impulsive neural networks, concrete steps are as follows:
S301, when a upper strata neuron is provided, the granting moment is denoted as t
f, according to the neuronic connection weight of STDP rule adjustment all lower floors neuron and upper strata.
S302, judges that upper strata neuron is at [t
f-σ, t
f] in whether have granting.
S303, if so, then increases connection weight, if not, then reduces connection weight.
In an embodiment of the present invention, after carrying out above-mentioned unsupervised learning to connection weight, each output neuron can be little by little responsive to certain specific response modes of input neuron group.
Particularly, Fig. 7 is the annexation schematic diagram of input neuron and output neuron in a unit networks.In Fig. 7, left side is input neuron, neuron in each frame is one group (namely having identical receptive field), right side is output neuron, and they are connected with all input neurons separately, and receptive field is spliced by the receptive field of each group of input neuron.When supposing an input image, z
2neuronic film potential is the highest, then within a period of time, only have z in output neuron
2provide (utilizing WTA mechanism), need adjust by STDP rule the connection weight that heavy black line in figure marks, get final product through type (2) and adjust all associated connection weight w
ki(1≤i≤n=7 × 6 × 4):
w
ki＝w
ki+η
ki·Δw
ki(2)
Wherein, Δ w
ki=y
i(t
f) exp (-w
ki)-1, η
kidesirable less constant, as 0.01, also can according to w
kichange through type (3) carry out self-adaptative adjustment, such as
The result of weight adjusting is that second layer neuron is broken up, and each second layer neuron namely in this unit networks can be responsive to the different characteristic of same image-region.
Neuronic for all second layers granting moment is recorded, just can realize the neuronic granting of more top and weight study by similar way.Finally, whole impulsive neural networks describes input picture by the neuronic granting of each layer, and has remembered with the feature of the form of connection weight by each for image yardstick.
The characteristics of image based on impulsive neural networks of the embodiment of the present invention describes and accumulating method, according to the number of plies N of the size determination impulsive neural networks of input picture, and determine a granting in impulsive neural networks in the every preset value of a ground floor neuron according to the image gradient direction after discretize, the neuronic film potential of the second layer is calculated according to ground floor neuronic granting situation, to determine the neuronic granting situation of the second layer, obtain the neuronic granting situation of all layers, connection weight between the sequential relationship provided according to all layer neurons and each layer of STDP rule adjustment impulsive neural networks, describe and mental picture feature with the form of connection weight.Method of the present invention can describe the image of memory plurality of classes simultaneously, and can intactly Recovery image, also has Images Classification function simultaneously.
The characteristics of image based on impulsive neural networks understanding the embodiment of the present invention for convenience describes and accumulating method, and present invention is described and describe in detail can to pass through Fig. 7.
As shown in Figure 7, dot represents neuron, and they have the structure of stratification.In every one deck, neuron is divided into some groups according to the difference of its receptive field, and the neuronic receptive field of each group can cover entire image in zero lap ground, and neuronic arrangement is corresponding with the position of its receptive field.Divide the receptive field of every one deck in figure with dotted line, it should be noted that, top neuronic receptive field covers entire image, so do not divide.
At the bottom of impulsive neural networks, neuronic receptive field is less, and these neurons are comparatively responsive to some simple local features (such as edge feature); Often rise one deck, and neuronic receptive field is spliced by some groups that are positioned at one deck close position below neuronic receptive fields.As shown in Figure 7, have 2 × 2 groups of neurons, often rise one deck, and neuronic receptive field expands four times (in fact, in different levels, the expansion multiple of receptive field can be different).Now, between 2 × 2 groups of neurons of lower floor and one group of neuron on upper strata, set up full connection, and and upper strata other respectively all do not connect between group neurons.
By adding WTA constraint for often organizing neuron, when input picture can be made constant, often only have a neuron response comparatively strong in group, the neuron of such upper strata response just can know the position relationship between some neurons that lower floor responds definitely, thus the complex characteristic that the responsive local feature of lower floor's neuron is spliced to form can be shown natural terrain very much.
Further, according to neuronic response, according to STDP rule can be non-supervisory interneuronal connection weight is adjusted, finally make each neuron certain feature-sensitive only to image in its receptive field, and responsive to different characteristics of image with different neuron in group.Neuron is in the memory body of characteristics of image now its connection weight, and when adjusting connection weight, similar feature (feature namely responded by same neuron) is classified as a class automatically.Again because network all remains positional information at every one deck, so this network can carry out complete description to input picture (cat as in Fig. 7) on each abstraction hierarchy.Moreover, utilize the memory of this network, the annexation namely between each layer neuron, new significant image can also be generated.As shown in Figure 8, be the connection weight part facial image that reconstructs out after utilizing the impulsive neural networks of the embodiment of the present invention to learn, illustrate that this network has more complete and memory capability clearly.
It should be noted that, due to the response of output neuron in unit networks and the response of input neuron similar, namely a neuron granting is had at the most in the neuron pool that a period of time enteroception is wild identical, so unit networks is combined cascade to form hierarchical structure is as shown in Figure 7 reasonable and naturally, only need using the input neuron of the output neuron of several unit networks of lower floor as last layer unit networks when cascade.Within the same layer, the receptive field of different units network is not overlapping mutually, and the different piece of their independent processing input pictures, thus can adopt the mode of parallel computation, significantly promotes the processing speed of overall network.Although represent the various features that may occur in a receptive field in impulsive neural networks with a large amount of neuron, to ensure there is stronger descriptive power for other image various types of, cause computation complexity high, but the STDP rule learning to adopt due to weight is a kind of local learning strategy, and parallel computation can be carried out, so in processing speed, do not have larger sacrifice when Description Image.
Describe and can be understood in process flow diagram or in this any process otherwise described or method, represent and comprise one or more for realizing the module of the code of the executable instruction of the step of specific logical function or process, fragment or part, and the scope of the preferred embodiment of the present invention comprises other realization, wherein can not according to order that is shown or that discuss, comprise according to involved function by the mode while of basic or by contrary order, carry out n-back test, this should understand by embodiments of the invention person of ordinary skill in the field.
In the description of this instructions, specific features, structure, material or feature that the description of reference term " embodiment ", " some embodiments ", " example ", " concrete example " or " some examples " etc. means to describe in conjunction with this embodiment or example are contained at least one embodiment of the present invention or example.In this manual, identical embodiment or example are not necessarily referred to the schematic representation of above-mentioned term.And the specific features of description, structure, material or feature can combine in an appropriate manner in any one or more embodiment or example.
Although illustrate and describe embodiments of the invention, those having ordinary skill in the art will appreciate that: can carry out multiple change, amendment, replacement and modification to these embodiments when not departing from principle of the present invention and aim, scope of the present invention is by claim and equivalency thereof.
Claims (8)
1. the characteristics of image based on impulsive neural networks describes and an accumulating method, it is characterized in that, comprises the following steps:
Input M opens normalized image;
According to the number of plies N of the size determination impulsive neural networks of described image, wherein, N be greater than 1 integer;
Pre-service is carried out to described image, obtains the gradient direction at each pixel place in described image, and turn to a preset value value by discrete for described gradient direction;
A granting in described impulsive neural networks in the every described preset value neuron of ground floor is determined according to the described gradient direction after discretize;
Calculate the neuronic film potential of the second layer according to described ground floor neuronic granting situation, to determine the neuronic granting situation of the described second layer, and then obtain the neuronic granting situation of all layers;
Connection weight described in the sequential relationship provided according to described all layer neurons and STDP rule adjustment between each layer of impulsive neural networks, describes with the form of described connection weight and remembers described characteristics of image.
2. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, the size of each layer receptive field of described impulsive neural networks is determined according to the size of described image and the described number of plies, wherein, in described impulsive neural networks, the neuronic receptive field size of ground floor is a pixel.
3. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, describedly carry out pre-service to described image, the method obtaining the gradient direction at each pixel place in described image adopts the one in Prewitt operator, Sobel operator, Kirsch operator and compass operator.
4. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, described preset value is 4.
5. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, the unit networks of N-1 type is comprised in the network of described N layer, wherein, described unit networks is be arranged in the first network formed of the two-layer partial nerve of continuous print by described N layer network.
6. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, described granting neuron is by Poisson process granting.
7. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, described according to the neuronic film potential of the described ground floor neuronic granting situation calculating second layer, to determine the neuronic granting situation of the described second layer, concrete steps are as follows:
Using the random number of independent identically distributed exponential distribution as Time Of Release interval;
Obtain providing the moment according to described Time Of Release interval;
With one millisecond for time step, by described granting moment discretize;
According to the granting moment after described discretize, generate described ground floor neuronic granting pulse y
i(t);
According to described y
it () is by the neuronic described film potential u of the second layer described in following formulae discovery
k(t):
Wherein,
w
kifor described ground floor neuron and the neuronic connection weight of the described second layer, y
ifor described ground floor neuron, 1≤i≤n, n is the neuronic number of described ground floor, and 1≤k≤K, K is the neuronic number of the described second layer, and σ is described y
ito the unit pulse width that the second layer neuron of correspondence sends during granting;
By adding WTA constraint for the second layer neuron of described ground floor neuron granting correspondence, select the neuron that the highest described second layer neuron of film potential is the granting of the described second layer.
8. describe and accumulating method based on the characteristics of image of impulsive neural networks as claimed in claim 1, it is characterized in that, connection weight described in the described sequential relationship provided according to described all layer neurons and STDP rule adjustment between each layer of impulsive neural networks, concrete steps are as follows:
When a upper strata neuron is provided, the granting moment is denoted as t
f, according to the neuronic connection weight of STDP rule adjustment all lower floors neuron and upper strata:
Judge that upper strata neuron is at [t
f-σ, t
f] in whether have granting;
If so, then increase described connection weight, if not, then reduce described connection weight.
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 true CN105404902A (en) | 2016-03-16 |
| CN105404902B 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) |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| 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 |
| CN107256554A (en) * | 2017-05-18 | 2017-10-17 | 西安电子科技大学 | A kind of individual layer impulsive neural networks structure split for image |
| CN108304767A (en) * | 2017-12-12 | 2018-07-20 | 中国科学院自动化研究所 | Human action's intention assessment training method based on more brain area cooperated computings |
| CN108446757A (en) * | 2017-12-29 | 2018-08-24 | 北京理工大学 | A kind of cascade speed identifying system of multistage delay and method based on impulsive neural networks |
| 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 |
| 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 |
| CN109142358A (en) * | 2018-08-22 | 2019-01-04 | 王天骄 | Metal material mechanics performance test methods neural network based |
| 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 |
| CN109325583A (en) * | 2017-07-31 | 2019-02-12 | 财团法人工业技术研究院 | Deep neural network, method and readable media using deep neural network |
| WO2019136772A1 (en) * | 2018-01-11 | 2019-07-18 | 深圳大学 | Blurred image restoration method, apparatus and device, and storage medium |
| CN110059812A (en) * | 2019-01-26 | 2019-07-26 | 中国科学院计算技术研究所 | Impulsive neural networks operation chip and related operation method |
| CN110111234A (en) * | 2019-04-11 | 2019-08-09 | 上海集成电路研发中心有限公司 | A kind of image processing system framework neural network based |
| CN110210563A (en) * | 2019-06-04 | 2019-09-06 | 北京大学 | The study of pattern pulse data space time information and recognition methods based on Spike cube SNN |
| CN110443363A (en) * | 2018-05-04 | 2019-11-12 | 北京市商汤科技开发有限公司 | Characteristics of image learning method and device |
| CN110781968A (en) * | 2019-10-29 | 2020-02-11 | 中国人民解放军国防科技大学 | Extensible class image identification method based on plastic convolution neural network |
| CN110826602A (en) * | 2019-10-23 | 2020-02-21 | 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 |
| CN111723726A (en) * | 2020-06-17 | 2020-09-29 | 苏州大学 | Identification method of identification model of drying agent packaging bag of pulse neurolemma system |
| CN111811532A (en) * | 2020-07-02 | 2020-10-23 | 浙江大学 | Path planning method and device based on impulse neural network |
| CN112232440A (en) * | 2020-11-10 | 2021-01-15 | 北京理工大学 | Method for realizing information memory and distinction of impulse neural network by using specific neuron groups |
| US20210342579A1 (en) * | 2018-09-05 | 2021-11-04 | Daimler Ag | Method for identifying a hand pose in a vehicle |
| CN113642535A (en) * | 2021-10-13 | 2021-11-12 | 聊城高新生物技术有限公司 | Biological branch detection method and device and electronic equipment |
| CN113673555A (en) * | 2021-07-09 | 2021-11-19 | 浙江大学 | Unsupervised domain adaptive picture classification method based on memory |
| US11308386B2 (en) | 2016-11-28 | 2022-04-19 | Huawei Technologies Co., Ltd. | Signal processing method and apparatus based on spiking neural network |
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 |
|---|
| SUDEB DAS ET AL: "NSCT-based multimodal medical image fusion using pulse-coupled neural network and modified spatial frequency", 《MED BIOL ENG COMPUT》 * |
| WEN JIANG ET AL: "Image segmentation with pulse-coupled neural network and Canny operators", 《COMPUTERS AND ELECTRICAL ENGINEERING》 * |
| 黄敏 等: "基于脉冲神经网络边缘提取的虹膜定位", 《电子测量技术》 * |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11308386B2 (en) | 2016-11-28 | 2022-04-19 | Huawei Technologies Co., Ltd. | Signal processing method and apparatus 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 |
| CN107256554A (en) * | 2017-05-18 | 2017-10-17 | 西安电子科技大学 | A kind of individual layer impulsive neural networks structure split for image |
| 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 |
| CN109325583B (en) * | 2017-07-31 | 2022-03-08 | 财团法人工业技术研究院 | Deep neural network structure, method using deep neural network, and readable medium |
| CN109325583A (en) * | 2017-07-31 | 2019-02-12 | 财团法人工业技术研究院 | Deep neural network, method and readable media using deep neural network |
| 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 |
| CN108304767A (en) * | 2017-12-12 | 2018-07-20 | 中国科学院自动化研究所 | Human action's intention assessment training method based on more brain area cooperated computings |
| CN108446757A (en) * | 2017-12-29 | 2018-08-24 | 北京理工大学 | A kind of cascade speed identifying system of multistage delay and method based on impulsive neural networks |
| 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 |
| WO2019136772A1 (en) * | 2018-01-11 | 2019-07-18 | 深圳大学 | Blurred image restoration method, apparatus and device, and storage medium |
| CN110443363A (en) * | 2018-05-04 | 2019-11-12 | 北京市商汤科技开发有限公司 | Characteristics of image 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 |
| CN109142358A (en) * | 2018-08-22 | 2019-01-04 | 王天骄 | Metal material mechanics performance test methods neural network based |
| US11887396B2 (en) * | 2018-09-05 | 2024-01-30 | Mercedes-Benz Group AG | Method for identifying a hand pose in a vehicle |
| US20210342579A1 (en) * | 2018-09-05 | 2021-11-04 | Daimler Ag | Method for identifying a hand pose in a vehicle |
| CN110059812A (en) * | 2019-01-26 | 2019-07-26 | 中国科学院计算技术研究所 | Impulsive neural networks operation chip and related operation method |
| 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 |
| CN110111234A (en) * | 2019-04-11 | 2019-08-09 | 上海集成电路研发中心有限公司 | A kind of image processing system framework neural network based |
| 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 |
| CN110210563A (en) * | 2019-06-04 | 2019-09-06 | 北京大学 | The study of pattern pulse data space time information and recognition methods 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 |
| CN110826602A (en) * | 2019-10-23 | 2020-02-21 | 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 |
| CN110781968A (en) * | 2019-10-29 | 2020-02-11 | 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 |
| CN111723726A (en) * | 2020-06-17 | 2020-09-29 | 苏州大学 | Identification method of identification model of drying agent packaging bag of pulse neurolemma system |
| CN111811532A (en) * | 2020-07-02 | 2020-10-23 | 浙江大学 | Path planning method and device based on impulse neural network |
| 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 |
| CN112232440A (en) * | 2020-11-10 | 2021-01-15 | 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 |
| CN113673555A (en) * | 2021-07-09 | 2021-11-19 | 浙江大学 | Unsupervised domain adaptive picture classification method based on memory |
| 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 |
| CN113642535A (en) * | 2021-10-13 | 2021-11-12 | 聊城高新生物技术有限公司 | Biological branch detection method and device and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105404902B (en) | 2019-02-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105404902A (en) | Impulsive neural network-based image feature describing and memorizing method | |
| Majdi et al. | Drive-net: Convolutional network for driver distraction detection | |
| CN110235146A (en) | Method and apparatus for the object of interest in detection image | |
| Mendoza et al. | Interval type-2 fuzzy logic and modular neural networks for face recognition applications | |
| CN107967451A (en) | A kind of method for carrying out crowd's counting to static image using multiple dimensioned multitask convolutional neural networks | |
| CN103258214A (en) | Remote sensing image classification method based on image block active learning | |
| CN106372666B (en) | A kind of target identification method and device | |
| CN106845541A (en) | A kind of image-recognizing method based on biological vision and precision pulse driving neutral net | |
| CN107403154A (en) | A kind of gait recognition method based on dynamic visual sensor | |
| CN106897738A (en) | A kind of pedestrian detection method based on semi-supervised learning | |
| CN104992142A (en) | Pedestrian recognition method based on combination of depth learning and property learning | |
| CN106909902A (en) | A kind of remote sensing target detection method based on the notable model of improved stratification | |
| CN106778796A (en) | Human motion recognition method and system based on hybrid cooperative model training | |
| CN107909151A (en) | For realizing the method and system of notice mechanism in artificial neural network | |
| US20200074190A1 (en) | Lane and object detection systems and methods | |
| CN111507227B (en) | Multi-student individual segmentation and state autonomous identification method based on deep learning | |
| CN110281949B (en) | Unified hierarchical decision-making method for automatic driving | |
| CN105046269A (en) | Multi-instance multi-label scene classification method based on multinuclear fusion | |
| Zinchenko et al. | Computer vision in control and optimization of road traffic | |
| CN110395257A (en) | A kind of lane line example detection method and apparatus, automatic driving vehicle | |
| KR20200052429A (en) | Method and system for multi-pedestrian tracking using teacher-student random ferns | |
| del Egio et al. | Self-driving a car in simulation through a CNN | |
| CN103544503B (en) | A kind of Activity recognition method based on many examples Markov model | |
| CN104778452A (en) | Feasible region detecting method based on machine learning | |
| CN113421243A (en) | Method and device for detecting type of fundus image based on knowledge distillation network |
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 |