CN109816022A - A kind of image-recognizing method based on three decisions and CNN - Google Patents
A kind of image-recognizing method based on three decisions and CNN Download PDFInfo
- Publication number
- CN109816022A CN109816022A CN201910084149.5A CN201910084149A CN109816022A CN 109816022 A CN109816022 A CN 109816022A CN 201910084149 A CN201910084149 A CN 201910084149A CN 109816022 A CN109816022 A CN 109816022A
- Authority
- CN
- China
- Prior art keywords
- decision
- image
- images
- negative sense
- region
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image-recognizing methods based on three decisions and CNN, including input sample data images training CNN classifier;Three decision thoughts are introduced in the image segmentation stage, the forward data collection image classified using CNN classifier and negative sense data images obtain three Decision Classfication devices;Positive region, negative region and Delayed Decision region are divided the image into using three Decision Classfication devices;Classification processing is iterated to the Delayed Decision region that the classifier is partitioned into;Remaining Delayed Decision region is judged, if reach critical value;When Delayed Decision region reaches critical value, illustrate that images to be recognized can not be divided again, CNN image recognition then is carried out to all positive regions;The present invention combines CNN image recognition technology and three decision theories, can make full use of useful information in image, makes to reach higher discrimination.
Description
Technical field
The present invention relates to Intelligent Information Processing and high-performance image identification technology field, and in particular to one kind is based on three
The image-recognizing method of decision and CNN.
Background technique
Image recognition is allowed to handle image using computer, analyzed and understood, to identify different target and right
The technology of elephant.The bottom semanteme of the existing eventful surplus image of image-recognizing method is studied, due to computer self performance
The problem of, can only sequencing the single operational order of execution, cause conventional method low to the discrimination of image.Therefore, someone
It proposes based on three decisions (Three-way Decision, abbreviation 3WD), convolutional neural networks (Convolutional
Neural Network, CNN) the methods of image recognition technology.
Wherein, three decisions have been developed in recent years a kind of method for handling unascertained decision, are a kind of compound
" three points and control " (Trisecting and Acting) model of human cognitive.Three decisions are a kind of to be recognized based on meeting the mankind
The decision-making mode known, it considers that: for people in practice decision process, the things accepted or rejected for tool beyond reasonable doubt can
Quick judgement is made immediately;For the things which cannot make a policy immediately, people often postpone the judgement to event,
That is Delayed Decision.In more practice decision process, there are the complexity of policy setting, the incompleteness of acquisition of information, groups
Each expert opinion inconsistency in decision, the uncertain conditions such as the finiteness and thinking ambiguity of the had knowledge of policymaker, certainly
Plan person is in most cases difficult accurately to provide the specific value of loss function, also results in its accuracy of identification and often reaches not
To necessary requirement.
In addition, convolutional neural networks are to grow up in recent years and a kind of widely used efficient neural network machine
Learning method.The main advantage of CNN can directly input original image in avoiding image pretreated complex process early period.
In general, CNN basic structure includes two layers: feature extraction layer and Feature Mapping layer.Its conventional machines learning method it is main
Difference is that CNN includes a feature extractor being made of a convolutional layer and sub-sampling layer.Sub-sampling is also referred to as pond
(Pooling), usually there are two kinds of forms of mean value sub-sampling (Mean Pooling) and maximum value sub-sampling (Max Pooling).
Sub-sampling is considered as a kind of special convolution process.Convolution sum sub-sampling not only simplifies model complexity, but also reduces
The parameter of model.But the treatment process of convolutional neural networks identification is single, inflexible, the characteristic value needed compares tool
Body causes the feature obtained very fixed;In addition, convolution process is easy to cause the missing of some useful informations in image, so that figure
The accuracy of identification of picture is not satisfactory.
Therefore, how three decision-making techniques to be combined with CNN image-recognizing method, realizes that more preferably image is known
Other technology becomes urgent problem to be solved.
Summary of the invention
In view of the deficiencies of the prior art, the object of the present invention is to provide a kind of image recognitions based on three decisions and CNN
Method, by the way that three decision theories are introduced into CNN image recognition, and the conceptual design classifier based on three decisions, it builds
Formwork erection type carries out image recognition, can be improved image recognition accuracy rate.
In order to achieve the above objectives, The technical solution adopted by the invention is as follows:
A kind of image-recognizing method based on three decisions and CNN, key be the following steps are included:
S1, input images to be recognized;
S2, CNN classifier to be trained is trained using sample image, and has been trained CNN classifier, forward data collection figure
Picture and negative sense data images, the training process of the CNN classifier to be trained are as follows:
S201, CNN classifier to be trained is established, and sets the initial parameter value of filter in CNN classifier to be trained;
S202, input sample image, respectively pre-process each image, and calculate in each image positive area with
Output probability corresponding to negative area;
S203, the overall error for calculating separately output layer;
S204, gradient of the error relative to all weights is calculated using back-propagation algorithm, and updated with gradient descent method
The weight and parameter value of all filters, so that output error minimizes;
S205, setting value of the smallest parameter value of output error as filter is selected, CNN classifier has been trained in acquisition;
S3, three Decision Classfication devices are obtained using the forward data collection image and negative sense data images;
S4, according to the class condition of three Decision Classfication devices, images to be recognized is divided into positive region, negative region, is prolonged
Slow decision region;
S5, using having trained CNN classifier to carry out image recognition processing to the positive region that splits;
S6, output image recognition result.
Further, described to obtain the tool of three Decision Classfication devices using forward data collection image and negative sense data images
Steps are as follows for body:
S301, the input forward data collection image and negative sense data images;
S302, according to the attribute of three decisions, forward data collection image and negative sense data images are respectively trained,
It include: by forward data collection image or negative sense data images respectively according to forward data collection image or negative sense data images
Characteristics of image is split, and the characteristics of image attribute value of forward data collection image or negative sense data images is obtained, according to forward direction
Attribute value λ needed for the characteristics of image attribute value of data images obtains the training of forward data collection imageαp1、λαn1、λβp1、λβn1、
λξp1、λξn1, attribute value needed for obtaining the training of negative sense data images according to the characteristics of image attribute value of negative sense data images
λαp2、λαn2、λβp2、λβn2、λξp2、λξn2;Wherein, λαp1、λαp2It respectively indicates and meets forward data collection image, negative sense data images
Decision condition under take the cost of acceptance decision, λαn1、λαn2It respectively indicates and is unsatisfactory for forward data collection image, negative sense data set
The cost of acceptance decision, λ are taken under the decision condition of imageβp1、λβp2It respectively indicates and meets forward data collection image, negative sense data
Collect the cost that refusal decision is taken under the decision condition of image, λβn1、λβn2It respectively indicates and is unsatisfactory for forward data collection image, negative sense
The cost for taking refusal decision under the decision condition of data images, λξp1、λξp2Respectively indicate meet forward data collection image,
The cost for not promising to undertake decision, λ are taken under the decision condition of negative sense data imagesξn1、λξn2It respectively indicates and is unsatisfactory for forward data
Collect and takes the cost for not promising to undertake decision under the decision condition of image, negative sense data images;
S303, using the decision problem cost matrix of three decisions, calculate separately out the positive region of forward data collection image
Optimal value α1With negative region optimal value β1And negative sense data images positive region optimal value α2With negative region optimal value β2:
S304, the positive region optimal value α according to forward data collection image1With negative region optimal value β1, obtain forward data collection
The Delayed Decision regional value ξ of image1: β1< ξ1< α1;According to negative sense data images positive region optimal value α2It is optimal with negative region
Value β2, obtain the Delayed Decision regional value ξ of forward data collection image2: β2< ξ2< α2;
S305, according to the α1、α2、β1、β2、ξ2And ξ2, respectively obtain three Decision Classfication devices of forward data collection image
With three Decision Classfication devices of negative sense data images.
Further, the characteristics of image of the forward data collection image or negative sense data images include image outline, it is bright
One of degree, color, gray scale are a variety of.
Further, according to the class condition of three Decision Classfication devices, the images to be recognized is divided into three areas
Domain includes: that forward data collection image and negative sense data images are quantified as the region that size is 1, is 1 by the size
The positive region optimal value and negative region optimal value of the forward data collection image that region is classified according to three decision-making devices or
The positive region optimal value and negative region optimal value of negative sense data images are respectively divided into three regions, according to risk function, selection
The decision of least risk, obtains evaluation of risk;According to the evaluation of risk, the images to be recognized is taken into three areas
The decision in some region in domain;
Wherein, three regions specifically: [α1,1]、[α2, 1] and it is respectively forward data collection image, negative sense data set figure
The positive region of picture, [0, β1]、[0,β2] be respectively forward data collection image, negative sense data images negative region, (α1,β1)、
(α2,β2) be respectively forward data collection image, negative sense data images Delayed Decision region;Wherein, α1、α2Indicate positive number
According to collection image, the optimal value of negative sense data images, β1、β2Indicate the negative region of forward data collection image, negative sense data images
Optimal value;Corresponding three decisions, use α1、α2It respectively indicates and takes receiving to determine forward data collection image, negative sense data images
Plan, β1、β2It respectively indicates and refusal decision, ξ is taken to forward data collection image, negative sense data images1、ξ2It respectively indicates to just
Delayed Decision is taken to data images, negative sense data images;
The evaluation of risk includes:
Receive risk:
R(α1| y)=λαp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(α2| y)=λαp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
Refuse risk:
R(β1| y)=λβp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(β2| y)=λβp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
Delayed Decision risk:
R(ξ1| y)=λξp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(ξ2| y)=λξp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
According to the evaluation of risk, the images to be recognized takes determining for some region into three regions
Plan includes:
When meeting condition R (α1|y)≤R(r1|y)∧R(α1|y)≤R(n1| when y), images to be recognized is taken to positive number
According to the decision of the receiving of collection image;
When meeting condition R (α2|y)≤R(r2|y)∧R(α2|y)≤R(n2| when y), images to be recognized is taken to negative sense number
According to the decision of the receiving of collection image;
When meeting condition R (r1|y)≤R(α1|y)∧R(r1|y)≤R(n1| when y), images to be recognized is taken to positive number
According to the decision of the refusal of collection image;
When meeting condition R (r2|y)≤R(α2|y)∧R(r2|y)≤R(n2| when y), images to be recognized is taken to negative sense number
According to the decision of the refusal of collection image;
When meeting condition R (n1|y)≤R(α1|y)∧R(n1|y)≤R(r1| when y), images to be recognized is taken to positive number
According to the decision that do not promise to undertake of collection image;
When meeting condition R (n2|y)≤R(α2|y)∧R(n2|y)≤R(r2| when y), images to be recognized is taken to negative sense number
According to the decision that do not promise to undertake of collection image;
Wherein, evaluation function is defined as Pr(Y | [y]), risk function are R (Δ | y), wherein Y indicates to act y decision,
Δ indicates to act the decision of y, and y indicates to determine, R (α1|y)、R(α2| y) indicate to forward data collection image, negative sense data set
Receive the risk function of state, R (β in image decision process1|y)、R(β2| y) indicate to forward data collection image, negative sense number
According to the risk function of disarmed state in collection image decision process, R (ξ1|y)、R(ξ2| it y) indicates to forward data collection image, negative
The risk function of Delayed Decision state into data images decision process;Pβ1(Y|[y])、Pβ2(Y | [y]) indicate to forward direction
The evaluation function of disarmed state in data images, negative sense data images decision process;Risk function R (r1|y)、R(r2|y)
Indicate that forward data collection image, negative sense data images decision act riRisk mathematic expectaion, risk function R (n1|y)、R
(n2| y) indicate forward data collection image, in negative sense data images the risk of decision movement n mathematic expectaion, ∧ indicates simultaneously
Meet.
Further, described according to the evaluation of risk, the images to be recognized is taken into three regions
The decision in some region further include:
For forward data collection image:
Ensure
WhenImages to be recognized takes the decision of the receiving to forward data collection image;
WhenImages to be recognized takes the decision of the refusal to forward data collection image;
WhenImages to be recognized takes the decision that do not promise to undertake to forward data collection image;
Wherein,It indicates acting r to the decision in forward data collection image decision process1Risk
Evaluation function, λαp1Expression meets the cost that acceptance decision is taken under decision condition, λαn1Foot forward data collection figure with thumb down
The cost of acceptance decision, λ are taken under the decision condition of pictureβp1It indicates to meet to take under the decision condition of forward data collection image and refuse
The cost of exhausted decision, λβn1The cost for taking refusal decision under the decision condition of foot forward data collection image with thumb down, λξp1
It indicates to meet under the decision condition of forward data collection image and takes the cost for not promising to undertake decision, λξn1Foot forward data with thumb down
Collect and takes the cost for not promising to undertake decision under the decision condition of image;
For negative sense data images:
Ensure
WhenImages to be recognized takes the decision of the receiving to negative sense data images;
WhenImages to be recognized takes the decision of the refusal to negative sense data images;
WhenImages to be recognized takes the decision that do not promise to undertake to negative sense data images;
Wherein,It indicates acting r to the decision in negative sense data images decision process2Risk
Evaluation function, λαp2Expression meets the cost that acceptance decision is taken under decision condition, λαn2Foot negative sense data set figure with thumb down
The cost of acceptance decision, λ are taken under the decision condition of pictureβp2It indicates to meet to take under the decision conditions of negative sense data images and refuse
The cost of exhausted decision, λβn2The cost for taking refusal decision under the decision condition of foot negative sense data images with thumb down, λξp2
It indicates to meet under the decision conditions of negative sense data images and takes the cost for not promising to undertake decision, λξn2Foot negative sense data with thumb down
Collect and takes the cost for not promising to undertake decision under the decision condition of image.
Further, the image to be identified is divided into three by the class condition according to three Decision Classfication devices
Region further include: introduce a pair of of threshold value (αi,βi), then set U is divided for following three regions:
To forward data collection image:
Positive region: POS (α1,β1)={ u ∈ ∪ | v (u) >=α1};
Negative region: NEG (α1,β1)={ u ∈ ∪ | v (u)≤β1};
Boundary Region:
Three decision rules can be constructed by above three region: positive region is corresponding to be received, the corresponding refusal of negative region, side
Boundary domain corresponds to Delayed Decision, and Boundary Region is also Delayed Decision region, wherein and u ∈ ∪ | v (u) >=α1Indicate to meet v (u) >=α1
And when u ∈ ∪, the value of image u;{u∈∪|v(u)≤β1Indicate to meet v (u)≤β1And when u ∈ ∪, the value of image u;It indicates to meetAnd when u ∈ ∪, the value of image u;Corresponding ordering relation
≤,Show stringent ordering relation;That is:And if only if β1≤v(u)≤α1And β1≠v(u)≠α1;α1Table
Show and acceptance decision, β are taken to forward data collection image1It indicates to take refusal decision, ξ to forward data collection image1Expression pair
Decision is not promised to undertake in taking for forward data collection image;U indicates that forward data image, v (u) indicate to quantify forward data image u
Value afterwards;
To negative sense data images:
Positive region: POS (α2,β2)={ u ∈ ∪ | v (u) >=α2};
Negative region: NEG (α2,β2)={ u ∈ ∪ | v (u)≤β2};
Boundary Region:
Three decision rules can be constructed by above three region: positive region is corresponding to be received, the corresponding refusal of negative region, side
Boundary domain corresponds to Delayed Decision, and Boundary Region is also Delayed Decision region, wherein and u ∈ ∪ | v (u) >=α2Indicate to meet v (u) >=α2
And when u ∈ ∪ image u value;{u∈∪|v(u)≤β2Indicate to meet v (u)≤β2And when u ∈ ∪ image u value;It indicates to meetAnd when u ∈ ∪ image u value;Corresponding ordering relation
≤,Indicate stringent ordering relation;That is:And if only if β2≤v(u)≤α2And β2≠v(u)≠α2;α2
Expression takes acceptance decision, β to negative sense data images2It indicates to take refusal decision, ξ to negative sense data images2It indicates
Decision is not promised to undertake to taking for negative sense data images;U indicates that negative sense data image, v (u) are indicated to after negative sense data u quantization
Value.
Further, the class condition according to three Decision Classfication devices, is divided into the image to be identified
Behind three regions, to Delayed Decision region subseries again, comprising: using the positive region sorted out as new positive sample,
The negative region sorted out is training condition as new negative sample, is trained again to three Decision Classfication devices, directly
It can not divide again to Delayed Decision region part, reach critical value.
Further, the formula for having trained CNN classifier to carry out image recognition processing to the positive region split
Are as follows:
Wherein, xiIndicate input vector, i.e., R, G, B eigenmatrix of every positive region image, i ∈ (1,2,3);WiIt indicates
Input vector xiCorresponding weight, i.e. convolution kernel;B indicates biasing;hW,bIt (x) is output characteristic value.
Further, the establishment process of the CNN classifier to be trained includes the sample number strong point according to images to be recognized
Obtained three classes sample makes classifier.
Remarkable result of the invention is: for image recognition technology in the prior art adaptivity it is poor the problems such as, pass through
The principle of three decision theories and CNN image recognition is analyzed, it is creative that three decision theories have been introduced into CNN figure
In identifying, and the conceptual design based on three decisions establishes CNN classifier, and has carried out adaptive instruction to CNN classifier
Practice, more accurate image feature information can be got, three Decision Classfication devices pair are generated according to training the data obtained collection image later
Images to be recognized is iterated segmentation, is finally known using the CNN classifier after training to resulting positive region image is divided
Other places reason, effectively raises the accurate rate of image recognition.
Detailed description of the invention
Fig. 1 is flow chart of the method for the present invention;
Fig. 2 is the training flow diagram of the CNN classifier to be trained;
Fig. 3 is the flow diagram of three Decision Classfication devices in the present invention.
Specific embodiment
Specific embodiment and working principle of the present invention will be described in further detail with reference to the accompanying drawing.
As shown in Figure 1, a kind of image-recognizing method based on three decisions and CNN, the specific steps are as follows:
S1, input images to be recognized;
S2, CNN classifier to be trained is trained using sample image, and has been trained CNN classifier, forward data collection figure
Picture and negative sense data images;
As shown in Fig. 2, the training process of CNN classifier to be trained described in this example are as follows:
S201, CNN classifier to be trained is established, and sets the initial parameter value of filter in CNN classifier to be trained;
S202, input sample image, respectively pre-process each image, and calculate in each image positive area with
Output probability corresponding to negative area;
S203, the overall error for calculating separately output layer;
S204, gradient of the error relative to all weights is calculated using back-propagation algorithm, and updated with gradient descent method
The weight and parameter value of all filters, so that output error minimizes;
S205, setting value of the smallest parameter value of output error as filter is selected, CNN classifier has been trained in acquisition;
S3, three Decision Classfication devices are obtained using the forward data collection image and negative sense data images;Referring to attached drawing
3, described using forward data collection image and negative sense data images to obtain three Decision Classfication devices specific step is as follows:
S301, the input forward data collection image and negative sense data images;
S302, according to the attribute of three decisions, forward data collection image and negative sense data images are respectively trained,
It include: by forward data collection image or negative sense data images respectively according to forward data collection image or negative sense data images
Characteristics of image is split, and the characteristics of image attribute value of forward data collection image or negative sense data images is obtained, according to forward direction
Attribute value λ needed for the characteristics of image attribute value of data images obtains the training of forward data collection imageαp1、λαn1、λβp1、λβn1、
λξp1、λξn1, attribute value needed for obtaining the training of negative sense data images according to the characteristics of image attribute value of negative sense data images
λαp2、λαn2、λβp2、λβn2、λξp2、λξn2;Wherein, λαp1、λαp2It respectively indicates and meets forward data collection image, negative sense data images
Decision condition under take the cost of acceptance decision, λαn1、λαn2It respectively indicates and is unsatisfactory for forward data collection image, negative sense data set
The cost of acceptance decision, λ are taken under the decision condition of imageβp1、λβp2It respectively indicates and meets forward data collection image, negative sense data
Collect the cost that refusal decision is taken under the decision condition of image, λβn1、λβn2It respectively indicates and is unsatisfactory for forward data collection image, negative sense
The cost for taking refusal decision under the decision condition of data images, λξp1、λξp2Respectively indicate meet forward data collection image,
The cost for not promising to undertake decision, λ are taken under the decision condition of negative sense data imagesξn1、λξn2It respectively indicates and is unsatisfactory for forward data
Collect and takes the cost for not promising to undertake decision under the decision condition of image, negative sense data images;
S303, using the decision problem cost matrix of three decisions, calculate separately out the positive region of forward data collection image
Optimal value α1With negative region optimal value β1And negative sense data images positive region optimal value α2With negative region optimal value β2:
1 two states decision problem cost matrix of table
S304, the positive region optimal value α according to forward data collection image1With negative region optimal value β1, obtain forward data collection
The Delayed Decision regional value ξ of image1: β1< ξ1< α1;According to negative sense data images positive region optimal value α2It is optimal with negative region
Value β2, obtain the Delayed Decision regional value ξ of forward data collection image2: β2< ξ2< α2;
S305, according to the α1、α2、β1、β2、ξ2And ξ2, respectively obtain three Decision Classfication devices of forward data collection image
With three Decision Classfication devices of negative sense data images.
S4, according to the class condition of three Decision Classfication devices, images to be recognized is divided into positive region, negative region, is prolonged
Slow decision region, and iterate to Delayed Decision region and can not divide again;
Preferably, alternatively, according to the class condition of three Decision Classfication devices, described to be identified
It includes: that forward data collection image and negative sense data images are quantified as the region that size is 1 that image, which is divided into three regions, will
The positive region optimal value for the forward data collection image that the region that the size is 1 is classified according to three decision-making devices and negative
The positive region optimal value and negative region optimal value of region optimal value or negative sense data images are respectively divided into three regions, according to wind
Dangerous function, the smallest decision of risk of selection, obtains evaluation of risk;According to the evaluation of risk, the images to be recognized take into
Enter the decision in some region in three regions;
Wherein, three regions specifically: [α1,1]、[α2, 1] and it is respectively forward data collection image, negative sense data set figure
The positive region of picture, [0, β1]、[0,β2] be respectively forward data collection image, negative sense data images negative region, (α1,β1)、
(α2,β2) be respectively forward data collection image, negative sense data images Delayed Decision region;Wherein, α1、α2Indicate positive number
According to collection image, the optimal value of negative sense data images, β1、β2Indicate the negative region of forward data collection image, negative sense data images
Optimal value;Corresponding three decisions, use α1、α2It respectively indicates and takes receiving to determine forward data collection image, negative sense data images
Plan, β1、β2It respectively indicates and refusal decision, ξ is taken to forward data collection image, negative sense data images1、ξ2It respectively indicates to just
Delayed Decision is taken to data images, negative sense data images;
The evaluation of risk includes:
Receive risk:
R(α1| y)=λαp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(α2| y)=λαp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
Refuse risk:
R(β1| y)=λβp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(β2| y)=λβp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
Delayed Decision risk:
R(ξ1| y)=λξp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(ξ2| y)=λξp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
According to the evaluation of risk, the images to be recognized takes determining for some region into three regions
Plan includes:
When meeting condition R (α1|y)≤R(r1|y)∧R(α1|y)≤R(n1| when y), images to be recognized is taken to positive number
According to the decision of the receiving of collection image;
When meeting condition R (α2|y)≤R(r2|y)∧R(α2|y)≤R(n2| when y), images to be recognized is taken to negative sense number
According to the decision of the receiving of collection image;
When meeting condition R (r1|y)≤R(α1|y)∧R(r1|y)≤R(n1| when y), images to be recognized is taken to positive number
According to the decision of the refusal of collection image;
When meeting condition R (r2|y)≤R(α2|y)∧R(r2|y)≤R(n2| when y), images to be recognized is taken to negative sense number
According to the decision of the refusal of collection image;
When meeting condition R (n1|y)≤R(α1|y)∧R(n1|y)≤R(r1| when y), images to be recognized is taken to positive number
According to the decision that do not promise to undertake of collection image;
When meeting condition R (n2|y)≤R(α2|y)∧R(n2|y)≤R(r2| when y), images to be recognized is taken to negative sense number
According to the decision that do not promise to undertake of collection image;
Wherein, evaluation function is defined as Pr(Y | [y]), risk function are R (Δ | y), wherein Y indicates to act y decision,
Δ indicates to act the decision of y, and y indicates to determine, R (α1|y)、R(α2| y) indicate to forward data collection image, negative sense data set
Receive the risk function of state, R (β in image decision process1|y)、R(β2| y) indicate to forward data collection image, negative sense number
According to the risk function of disarmed state in collection image decision process, R (ξ1|y)、R(ξ2| it y) indicates to forward data collection image, negative
The risk function of Delayed Decision state into data images decision process;Pβ1(Y|[y])、Pβ2(Y | [y]) indicate to forward direction
The evaluation function of disarmed state in data images, negative sense data images decision process;Risk function R (r1|y)、R(r2|y)
Indicate that forward data collection image, negative sense data images decision act riRisk mathematic expectaion, risk function R (n1|y)、R
(n2| y) indicate forward data collection image, in negative sense data images the risk of decision movement n mathematic expectaion, ∧ indicates simultaneously
Meet.
In the specific implementation process, described according to the evaluation of risk, the images to be recognized is taken into described three
The decision in some region in region further include:
For forward data collection image:
Ensure
WhenImages to be recognized takes the decision of the receiving to forward data collection image;
WhenImages to be recognized takes the decision of the refusal to forward data collection image;
WhenImages to be recognized takes the decision that do not promise to undertake to forward data collection image;
Wherein,It indicates acting r to the decision in forward data collection image decision process1Risk
Evaluation function, λαp1Expression meets the cost that acceptance decision is taken under decision condition, λαn1Foot forward data collection figure with thumb down
The cost of acceptance decision, λ are taken under the decision condition of pictureβp1It indicates to meet to take under the decision condition of forward data collection image and refuse
The cost of exhausted decision, λβn1The cost for taking refusal decision under the decision condition of foot forward data collection image with thumb down, λξp1
It indicates to meet under the decision condition of forward data collection image and takes the cost for not promising to undertake decision, λξn1Foot forward data with thumb down
Collect and takes the cost for not promising to undertake decision under the decision condition of image;
For negative sense data images:
Ensure
WhenImages to be recognized takes the decision of the receiving to negative sense data images;
WhenImages to be recognized takes the decision of the refusal to negative sense data images;
WhenImages to be recognized takes the decision that do not promise to undertake to negative sense data images;
Wherein,It indicates acting r to the decision in negative sense data images decision process2Risk
Evaluation function, λαp2Expression meets the cost that acceptance decision is taken under decision condition, λαn2Foot negative sense data set figure with thumb down
The cost of acceptance decision, λ are taken under the decision condition of pictureβp2It indicates to meet to take under the decision conditions of negative sense data images and refuse
The cost of exhausted decision, λβn2The cost for taking refusal decision under the decision condition of foot negative sense data images with thumb down, λξp2
It indicates to meet under the decision conditions of negative sense data images and takes the cost for not promising to undertake decision, λξn2Foot negative sense data with thumb down
Collect and takes the cost for not promising to undertake decision under the decision condition of image.
Further, the image to be identified is divided into three by the class condition according to three Decision Classfication devices
Region further include: introduce a pair of of threshold value (αi,βi), then set U is divided for following three regions:
To forward data collection image:
Positive region: POS (α1,β1)={ u ∈ ∪ | v (u) >=α1};
Negative region: NEG (α1,β1)={ u ∈ ∪ | v (u)≤β1};
Boundary Region:
Three decision rules can be constructed by above three region: positive region is corresponding to be received, the corresponding refusal of negative region, side
Boundary domain corresponds to Delayed Decision, and Boundary Region is also Delayed Decision region, in which:
{u∈∪|v(u)≥α1Indicate to meet v (u) >=α1And when u ∈ ∪, the value of image u;
{u∈∪|v(u)≤β1Indicate to meet v (u)≤β1And when u ∈ ∪, the value of image u;
It indicates to meetAnd when u ∈ ∪, the value of image u;
Corresponding ordering relation≤,Indicate stringent ordering relation;That is:And if only if β1≤v(u)
≤α1And β1≠v(u)≠α1;α1Expression takes acceptance decision, β to forward data collection image1It indicates to forward data collection image
Take refusal decision, ξ1Expression does not promise to undertake decision to taking for forward data collection image;U indicates forward data image, v (u)
It indicates to the value after forward data image u quantization;
To negative sense data images:
Positive region: POS (α2,β2)={ u ∈ ∪ | v (u) >=α2};
Negative region: NEG (α2,β2)={ u ∈ ∪ | v (u)≤β2};
Boundary Region:
Three decision rules can be constructed by above three region: positive region is corresponding to be received, the corresponding refusal of negative region, side
Boundary domain corresponds to Delayed Decision, and Boundary Region is also Delayed Decision region, wherein and u ∈ ∪ | v (u) >=α2Indicate to meet v (u) >=α2
And when u ∈ ∪ image u value;{u∈∪|v(u)≤β2Indicate to meet v (u)≤β2And when u ∈ ∪ image u value;It indicates to meetAnd when u ∈ ∪ image u value;Corresponding ordering relation
≤,Indicate stringent ordering relation;That is:And if only if β2≤v(u)≤α2And β2≠v(u)≠α2;α2
Expression takes acceptance decision, β to negative sense data images2It indicates to take refusal decision, ξ to negative sense data images2It indicates
Decision is not promised to undertake to taking for negative sense data images;U indicates that negative sense data image, v (u) are indicated to after negative sense data u quantization
Value.
Finally, the image to be identified, is divided into three by the class condition according to three Decision Classfication devices
Behind region, to Delayed Decision region subseries again, comprising: using the positive region sorted out as new positive sample, divided
The negative region that class goes out is training condition as new negative sample, is trained again to three Decision Classfication devices, Zhi Daosuo
Stating Delayed Decision region part can not divide again, reach critical value.
S5, using CNN classifier has been trained, to the positive region progress image recognition processing split, detailed process is such as
Under:
In the convolutional layer of convolutional neural networks, a neuron is only connect with part adjacent bed neuron.Usually CNN's
It can include several characteristic planes (Feature Map), wherein being made of the neuron of some rectangular arrangeds in one convolutional layer
Each characteristic plane, and the neuron of same characteristic plane shares weight, shared weight is exactly convolution kernel here.General convolution
Core is initialized by the form of random decimal matrix, and then convolution kernel obtains study reasonably in the training process of network
Weight.Reducing the most direct mode of connection between each layer of network is shared weight (convolution kernel), but reduces over-fitting simultaneously
Risk.Sub-sampling is also referred to as pond (Pooling), usually there is mean value sub-sampling (Mean Pooling) and maximum value sub-sampling
(Max Pooling) two kinds of forms.Sub-sampling is considered as a kind of special convolution process.Convolution sum sub-sampling not only simplifies
Model complexity, and also reduce the parameter of model.
According to the theory of neural network, it is as follows to correspond to formula:
Wherein, xiIndicate input vector, i.e., R, G, B eigenmatrix of every positive region image, i ∈ (1,2,3);WiIt indicates
Input vector xiCorresponding weight, i.e. convolution kernel;B indicates biasing;hW,bIt (x) is output characteristic value.
The CNN image recognition processes include:
Neural network model is multiple units to combine, and contain layered structure, wherein corresponding formula is as follows:
Wherein, x1, x2, x3Indicate input vector;W indicates weight;B indicates biasing;hW,bIt (x) is output;
Indicate activation primitive, activate Neuron characteristics, feature is retained by function and is mapped out come.Numerous neurons receive a large amount of non-
Linear shape inputs information, this information is known as input vector;Hidden layer is numerous neurons and link-group between input layer, output layer
At every aspect;Output layer is expressed as information and transmits in neuron link, analyzes, weighs obtained output result.
Corresponding three decisions and CNN image recognition are nominal variables since CNN is suitable for dependent variable, and by three or three
The case where a above classification composition, according to the process of three decisions, we take the state set to beIt respectively indicates
The object belongs to C and is not belonging to C.Action collection A={ aP,aB,aN, wherein aP, aB, aNRespectively indicate can take when x classification three
Kind is taken action, and determines x ∈ POS (C) respectively, x ∈ BND (C), generated risk under three kinds of states of x ∈ NEG (C), thus structure
A loss matrix is made, as shown in table 2.
The loss matrix of 2 two states decision problem of table
In table, λPP, λBP, λNPIt respectively indicates as x ∈ C, take action aP, aB, aNWhen loss;λPN, λBN, λNNPoint
Biao Shi it not work asWhen, take action aP, aB, aNWhen loss.If Pr(C | [x]) is the conditional probability of object x ∈ C,
Object x will be described by its equivalence class [x] when calculating.For an object, expected loss R when three kinds of different action is taken
(ai| [x]) it may be expressed as:
R(aP| [x])=λPPPr(C|[x])+λPNPr(C/[x])
R(aB| [x])=λBPPr(C|[x])+λBNPr(C|[x])
R(aN| [x])=λNPPr(C|[x])+λNNPr(C|[x])
S6, output image recognition result.
The characteristics of image of the forward data collection image or negative sense data images includes image outline, brightness, color, ash
One of degree is a variety of.
Training block includes these sample data set images in a random order, but some trained blocks may be wrapped than other classification
Containing more images.Training block clock is every a kind of comprising 5000 images.Raw data set is all monocular logo image, in order to verify three
4 images in original sample data set image are synthesized one by the accuracy rate of branch Decision Classfication device, the present invention, i.e., each survey
Examination item has 4 targets to be identified.
Further, alternatively, 4000 picture conducts provided in sample data set image are chosen
Image to be identified, wherein every class random selection picture 400 is opened, and experiment is 100 pictures every time, and final result is averaged.
Statistical data can obtain table 3, and as shown in table 2, the image recognition based on three decisions can be achieved on the identification of multi-Target Image.
According to the data of table 3 it is found that method of the invention accuracy rate average in data set cifar-10 is
90.85%, average F value is 91.25%, it can thus be concluded that the image recognition based on three Decision Classfication devices can obtain higher identification
Rate and accuracy rate, cifar-10 are that Canadian government is taken the lead one of investment advanced science project research collected one and is used for
The data set of pervasive object identification.
Table 3 is based on three decision cifar-10 multi-targets recognition rates
Further, as another optional way, using Car detection on UIUC data set as to be identified
Image: the multi-targets recognition of Car detection on UIUC data set is compared, in order to verify three Decision Classfication devices pair
In the accuracy rate of multi-targets recognition, present invention uses data set Car detection on UIUC as a comparison.Wherein, Car
Detection on UIUC data set is University of Illinois's automotive check image data base, wherein including 1050 training figures
This 1050 training images are split according to characteristics of image, obtain positive sample 550, negative sample 500 by picture;There are two
Test data set, one has 200 targets to be identified comprising 170 pictures, another include 108 pictures have 139 wait know
Other target.
As shown in table 4, for UIUC data set, the identification target of characteristics of image compares zero in sample data set image
It dissipating, is easier to ignore useful information, wherein CNN-3WD is method proposed by the invention, and FD is quick detection framework,
Precision is accuracy rate, and Recall is recall rate, and F-Measure is F value, and all data are all figures in data set in table
As data are averaged, analysis can obtain the image recognition accurate rate of the invention based on three decisions and CNN and be higher than quickly detection
Frame FD.The method of the present invention discrimination reaches 99.41%, F value and is slightly above FD method for 99.26%.
Table 4 is compared based on UIUC discrimination
Further, as a kind of arbitrary way, using KITTI data set to the image of the invention based on three decisions
Recognition methods is verified.Wherein, KITTI data set is by the Karlsruhe, Germany Institute of Technology and American Institute of Technology, Toyota
Joint is established, and is the computer vision algorithms make evaluation and test data set under current automatic Pilot scene maximum in the world.The evaluation and test number
Real image data according to collection comprising the scenes acquisition such as urban district, rural area and highway, in every image for up to 15 vehicles and
30 pedestrians are blocked and are truncated there are also various degrees of.
Wherein, CS-AdaBoost is that Masnadishirazi H is proposed in paper Cost-sensitive boosting
Method, AdaBoost+LDA be Wu J, Brubaker S C, Mullin M D, et al is in paper Fast Asymmetric
The method proposed in Learning for Cascade Face Detection, AdaBoost are Viola P, and Jones M is being discussed
The method proposed in literary Robust real-time face detection, Pruning are that Paisitkriangkrai S is being discussed
The method proposed in literary Asymmetric Pruning for Learning Cascade Detectors, FD are Hu Q,
Paisitkriangkrai S, Shen C, et al are in paper Fast Detection of Multiple Objects in
The method proposed in Traffic Scenes With a Common Detection Framework;
KITTI data set is divided into three classifications, is vehicle, pedestrian and bicycle respectively, but due to pedestrian and bicycle sample
There is the case where interfering with each other in this image itself, and its criterion is not known yet, so the present invention uses the image of vehicle
Tested and compared, the method made comparisons be also use image in information of vehicles as object of experiment.
As shown in Table 5, information of vehicles in data set KITTI is identified, the discrimination of 3WD method is 89.27%, F value
It is 88.86%, is higher than FD due to the particularity of the data set and only has trained the image of vehicle, acquired results and its other party herein
The information of vehicles of method compares, and as shown in table 5, the method applied in the present invention has higher accuracy rate and F compared to other methods
Value.
Table 5 is compared based on KITTI data set identify rate
Wherein, mBoW is Behley J in paper Laser-based segment classification using a
The method proposed in mixture of bag-of-words, MDPM are Forsyth D in paper Object Detection
The method proposed in Discriminatively Trained Part-Based Models, DPM-C8B1 are Yebes J
J is in paper Supervised learning and evaluation of KITTI's cars detector with DPM
The method of proposition, OC-DPM are Pepikj B in paper Occlusion Patterns for Object Class
The method proposed in Detection, AOG are Li B, and Wu T, Zhu S C is in paper Integrating Context and
The method proposed in Occlusion for Car Detection by Hierarchical And-Or Model, SubCat are
Ohn-Bar E is in paper Fast and Robust Object Detection Using Visual Subcategories
The method of proposition, Regionlets are Wang X in paper Regionlets for Generic Object Detection
The method of proposition.
Technical solution provided by the present invention is described in detail above.Specific case used herein is to this hair
Bright principle and embodiment is expounded, method of the invention that the above embodiments are only used to help understand and its
Core concept.It should be pointed out that for those skilled in the art, in the premise for not departing from the principle of the invention
Under, it can be with several improvements and modifications are made to the present invention, these improvement and modification also fall into the protection of the claims in the present invention
In range.
Claims (9)
1. a kind of image-recognizing method based on three decisions and CNN, which comprises the following steps:
S1, input images to be recognized;
S2, using sample image training CNN classifier train, and trained CNN classifier, forward data collection image and
Negative sense data images, the training process of the CNN classifier to be trained are as follows:
S201, CNN classifier to be trained is established, and sets the initial parameter value of filter in CNN classifier to be trained;
S202, input sample image, respectively pre-process each image, and calculate positive area and negative sense in each image
Output probability corresponding to region;
S203, the overall error for calculating separately output layer;
S204, gradient of the error relative to all weights is calculated using back-propagation algorithm, and updated and owned with gradient descent method
Filter weight and parameter value so that output error minimize;
S205, setting value of the smallest parameter value of output error as filter is selected, CNN classifier has been trained in acquisition;
S3, three Decision Classfication devices are obtained using the forward data collection image and negative sense data images;
S4, according to the class condition of three Decision Classfication devices, images to be recognized is divided into positive region, negative region, delay are determined
Plan region;
S5, using having trained CNN classifier to carry out image recognition processing to the positive region that splits;
S6, output image recognition result.
2. the image-recognizing method according to claim 1 based on three decisions and CNN, it is characterised in that: the utilization
Forward data collection image and negative sense data images obtain three Decision Classfication devices, and specific step is as follows:
S301, the input forward data collection image and negative sense data images;
S302, according to the attribute of three decisions, forward data collection image and negative sense data images are respectively trained, wrap
It includes: by forward data collection image or negative sense data images respectively according to forward data collection image or the figure of negative sense data images
As feature is split, the characteristics of image attribute value of forward data collection image or negative sense data images is obtained, according to positive number
Attribute value λ needed for obtaining the training of forward data collection image according to the characteristics of image attribute value of collection imageαp1、λαn1、λβp1、λβn1、λξp1、
λξn1, attribute value λ needed for obtaining the training of negative sense data images according to the characteristics of image attribute value of negative sense data imagesαp2、
λαn2、λβp2、λβn2、λξp2、λξn2;Wherein, λαp1、λαp2Respectively indicate meet forward data collection image, negative sense data images certainly
The cost of acceptance decision, λ are taken under the conditions of planαn1、λαn2It respectively indicates and is unsatisfactory for forward data collection image, negative sense data images
Decision condition under take the cost of acceptance decision, λβp1、λβp2It respectively indicates and meets forward data collection image, negative sense data set figure
The cost of refusal decision, λ are taken under the decision condition of pictureβn1、λβn2It respectively indicates and is unsatisfactory for forward data collection image, negative sense data
Collect the cost for taking refusal decision under the decision condition of image, λξp1、λξp2It respectively indicates and meets forward data collection image, negative sense
The cost for not promising to undertake decision, λ are taken under the decision condition of data imagesξn1、λξn2It respectively indicates and is unsatisfactory for forward data collection figure
As, negative sense data images decision condition under take the cost for not promising to undertake decision;
S303, using the decision problem cost matrix of three decisions, the positive region for calculating separately out forward data collection image is optimal
Value α1With negative region optimal value β1And negative sense data images positive region optimal value α2With negative region optimal value β2:
S304, the positive region optimal value α according to forward data collection image1With negative region optimal value β1, obtain forward data collection image
Delayed Decision regional value ξ1: β1< ξ1< α1;According to negative sense data images positive region optimal value α2With negative region optimal value β2,
Obtain the Delayed Decision regional value ξ of forward data collection image2: β2< ξ2< α2;
S305, according to the α1、α2、β1、β2、ξ2And ξ2, respectively obtain three Decision Classfication devices of forward data collection image and bear
To three Decision Classfication devices of data images.
3. the image-recognizing method according to claim 1 based on three decisions and CNN, it is characterised in that: the forward direction
The characteristics of image of data images or negative sense data images includes one of image outline, brightness, color, gray scale or more
Kind.
4. the image-recognizing method according to claim 1 based on three decisions and CNN, it is characterised in that: according to described
The class condition of three Decision Classfication devices, the images to be recognized be divided into three regions include: by forward data collection image with
Negative sense data images are quantified as the region that size is 1, and the region that the size is 1 is classified according to three decision-making devices
The positive region optimal value and negative region optimal value of obtained forward data collection image or the positive region of negative sense data images are optimal
Value and negative region optimal value are respectively divided into three regions, and according to risk function, the smallest decision of risk of selection obtains risk and estimates
Meter;According to the evaluation of risk, the images to be recognized takes the decision in some region into three regions;
Wherein, three regions specifically: [α1,1]、[α2, 1] and it is respectively forward data collection image, negative sense data images
Positive region, [0, β1]、[0,β2] be respectively forward data collection image, negative sense data images negative region, (α1,β1)、(α2,β2)
The respectively Delayed Decision region of forward data collection image, negative sense data images;Wherein, α1、α2Indicate forward data collection figure
Picture, the optimal value of negative sense data images, β1、β2Indicate the optimal of the negative region of forward data collection image, negative sense data images
Value;Corresponding three decisions, use α1、α2It respectively indicates and acceptance decision, β is taken to forward data collection image, negative sense data images1、
β2It respectively indicates and refusal decision, ξ is taken to forward data collection image, negative sense data images1、ξ2It respectively indicates to forward data
Collection image, negative sense data images take Delayed Decision;
The evaluation of risk includes:
Receive risk:
R(α1| y)=λαp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(α2| y)=λαp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
Refuse risk:
R(β1| y)=λβp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(β2| y)=λβp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
Delayed Decision risk:
R(ξ1| y)=λξp1·Pβ1(Y|[y])+(1-Pβ1(Y|[y]));
R(ξ2| y)=λξp2·Pβ2(Y|[y])+(1-Pβ2(Y|[y]));
According to the evaluation of risk, the images to be recognized takes the decision package in some region into three regions
It includes:
When meeting condition R (α1|y)≤R(r1|y)∧R(α1|y)≤R(n1| when y), images to be recognized is taken to forward data collection figure
The decision of the receiving of picture;
When meeting condition R (α2|y)≤R(r2|y)∧R(α2|y)≤R(n2| when y), images to be recognized is taken to negative sense data set figure
The decision of the receiving of picture;
When meeting condition R (r1|y)≤R(α1|y)∧R(r1|y)≤R(n1| when y), images to be recognized is taken to forward data collection figure
The decision of the refusal of picture;
When meeting condition R (r2|y)≤R(α2|y)∧R(r2|y)≤R(n2| when y), images to be recognized is taken to negative sense data set figure
The decision of the refusal of picture;
When meeting condition R (n1|y)≤R(α1|y)∧R(n1|y)≤R(r1| when y), images to be recognized is taken to forward data collection figure
The decision that do not promise to undertake of picture;
When meeting condition R (n2|y)≤R(α2|y)∧R(n2|y)≤R(r2| when y), images to be recognized is taken to negative sense data set figure
The decision that do not promise to undertake of picture;
Wherein, evaluation function is defined as Pr(Y | [y]), risk function are R (Δ | y), wherein Y indicates to act y decision, Δ table
Show the decision movement to y, y indicates to determine, R (α1|y)、R(α2| y) indicate to forward data collection image, negative sense data images
Receive the risk function of state, R (β in decision process1|y)、R(β2| y) indicate to forward data collection image, negative sense data set
The risk function of disarmed state in image decision process, R (ξ1|y)、R(ξ2| y) indicate to forward data collection image, negative sense number
According to the risk function of Delayed Decision state in collection image decision process;Pβ1(Y|[y])、Pβ2(Y | [y]) indicate to forward data
The evaluation function of disarmed state in collection image, negative sense data images decision process;Risk function R (r1|y)、R(r2| y) indicate
Forward data collection image, negative sense data images decision act riRisk mathematic expectaion, risk function R (n1|y)、R(n2|
Y) indicate forward data collection image, in negative sense data images the risk of decision movement n mathematic expectaion, ∧ indicates to expire simultaneously
Foot.
5. the image-recognizing method according to claim 4 based on three decisions and CNN, it is characterised in that: the basis
The evaluation of risk, the images to be recognized take the decision in some region into three regions further include:
For forward data collection image:
Ensure
WhenImages to be recognized takes the decision of the receiving to forward data collection image;
WhenImages to be recognized takes the decision of the refusal to forward data collection image;
WhenImages to be recognized takes the decision that do not promise to undertake to forward data collection image;
Wherein,It indicates acting r to the decision in forward data collection image decision process1Risk comment
Valence function, λαp1Expression meets the cost that acceptance decision is taken under decision condition, λαn1Foot forward data collection image with thumb down
The cost of acceptance decision, λ are taken under decision conditionβp1Indicating to meet under the decision condition of forward data collection image takes refusal to determine
The cost of plan, λβn1The cost for taking refusal decision under the decision condition of foot forward data collection image with thumb down, λξp1It indicates
Meet the cost taken under the decision condition of forward data collection image and do not promise to undertake decision, λξn1Foot forward data collection figure with thumb down
The cost for not promising to undertake decision is taken under the decision condition of picture;
For negative sense data images:
Ensure
WhenImages to be recognized takes the decision of the receiving to negative sense data images;
WhenImages to be recognized takes the decision of the refusal to negative sense data images;
WhenImages to be recognized takes the decision that do not promise to undertake to negative sense data images;
Wherein,It indicates acting r to the decision in negative sense data images decision process2Risk comment
Valence function, λαp2Expression meets the cost that acceptance decision is taken under decision condition, λαn2Foot negative sense data images with thumb down
The cost of acceptance decision, λ are taken under decision conditionβp2Indicating to meet under the decision conditions of negative sense data images takes refusal to determine
The cost of plan, λβn2The cost for taking refusal decision under the decision condition of foot negative sense data images with thumb down, λξp2It indicates
Meet the cost taken under the decision condition of negative sense data images and do not promise to undertake decision, λξn2Foot negative sense data set figure with thumb down
The cost for not promising to undertake decision is taken under the decision condition of picture.
6. the image-recognizing method according to claim 1 based on three decisions and CNN, it is characterised in that: the basis
The image to be identified is divided into three regions by the class condition of three Decision Classfication devices further include: introduces a pair of of threshold value
(αi,βi), then set U is divided for following three regions:
To forward data collection image:
Positive region: POS (α1,β1)={ u ∈ ∪ | v (u) >=α1};
Negative region: NEG (α1,β1)={ u ∈ ∪ | v (u)≤β1};
Boundary Region:
Three decision rules can be constructed by above three region: positive region is corresponding to be received, the corresponding refusal of negative region, Boundary Region
Corresponding Delayed Decision, Boundary Region is also Delayed Decision region, wherein and u ∈ ∪ | v (u) >=α1Indicate to meet v (u) >=α1And u ∈
When ∪, the value of image u;{u∈∪|v(u)≤β1Indicate to meet v (u)≤β1And when u ∈ ∪, the value of image u;It indicates to meetAnd when u ∈ ∪, the value of image u;Corresponding ordering relation
≤,Indicate stringent ordering relation;That is:And if only if β1≤v(u)≤α1And β1≠v(u)≠α1;α1
Expression takes acceptance decision, β to forward data collection image1It indicates to take refusal decision, ξ to forward data collection image1It indicates
Decision is not promised to undertake to taking for forward data collection image;U indicates that forward data image, v (u) are indicated to forward data image u amount
Value after change;
To negative sense data images:
Positive region: POS (α2,β2)={ u ∈ ∪ | v (u) >=α2};
Negative region: NEG (α2,β2)={ u ∈ ∪ | v (u)≤β2};
Boundary Region:
Three decision rules can be constructed by above three region: positive region is corresponding to be received, the corresponding refusal of negative region, and Boundary Region is corresponding
Delayed Decision, Boundary Region are also Delayed Decision region, wherein and u ∈ ∪ | v (u) >=α2Indicate to meet v (u) >=α2And image when u ∈ ∪
The value of u;{u∈∪|v(u)≤β2Indicate to meet v (u)≤β2And when u ∈ ∪ image u value;
It indicates to meetAnd when u ∈ ∪ image u value;Corresponding ordering relation≤,Indicate that stringent total order is closed
System;That is:And if only if β2≤v(u)≤α2And β2≠v(u)≠α2;α2It indicates to negative sense data images
Take acceptance decision, β2It indicates to take refusal decision, ξ to negative sense data images2Negative sense data images are taken in expression
Do not promise to undertake decision;U indicates that negative sense data image, v (u) are indicated to the value after negative sense data u quantization.
7. the image-recognizing method based on three decisions and CNN described in -6 according to claim 1, it is characterised in that: described
According to the class condition of three Decision Classfication devices, after the image to be identified is divided into three regions, determine to the delay
Plan region subseries again, comprising: using the positive region sorted out as new positive sample, the negative region sorted out as new
Negative sample is training condition, is trained again to three Decision Classfication devices, until Delayed Decision region part not
It can divide again, reach critical value.
8. the image-recognizing method according to claim 1 based on three decisions and CNN, it is characterised in that: described to have instructed
Practice the formula that CNN classifier carries out image recognition processing to the positive region split are as follows:
Wherein, xiIndicate input vector, i.e., R, G, B eigenmatrix of every positive region image, i ∈ (1,2,3);WiIndicate input
Vector xiCorresponding weight, i.e. convolution kernel;B indicates biasing;hW,bIt (x) is output characteristic value.
9. the image-recognizing method according to claim 1 based on three decisions and CNN, it is characterised in that: described wait instruct
The establishment process for practicing CNN classifier includes the three classes sample obtained according to the sample number strong point of images to be recognized, makes classifier.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910084149.5A CN109816022A (en) | 2019-01-29 | 2019-01-29 | A kind of image-recognizing method based on three decisions and CNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910084149.5A CN109816022A (en) | 2019-01-29 | 2019-01-29 | A kind of image-recognizing method based on three decisions and CNN |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109816022A true CN109816022A (en) | 2019-05-28 |
Family
ID=66605564
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910084149.5A Pending CN109816022A (en) | 2019-01-29 | 2019-01-29 | A kind of image-recognizing method based on three decisions and CNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109816022A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110473251A (en) * | 2019-08-14 | 2019-11-19 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Custom field spatial data area statistics method based on grid spatial index |
CN111046926A (en) * | 2019-11-26 | 2020-04-21 | 山东浪潮人工智能研究院有限公司 | Computer vision image classification integrated learning method |
CN112580785A (en) * | 2020-12-18 | 2021-03-30 | 河北工业大学 | Neural network topological structure optimization method based on three-branch decision |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH07152543A (en) * | 1993-11-30 | 1995-06-16 | Kobe Nippon Denki Software Kk | Decision table input device |
US5956428A (en) * | 1997-05-19 | 1999-09-21 | Ausbeck, Jr.; Paul James | Stroke code for image partitions |
CN106796647A (en) * | 2014-09-05 | 2017-05-31 | 北京市商汤科技开发有限公司 | Scene text detecting system and method |
CN108154172A (en) * | 2017-12-21 | 2018-06-12 | 重庆邮电大学 | Image-recognizing method based on three decisions |
CN108226889A (en) * | 2018-01-19 | 2018-06-29 | 中国人民解放军陆军装甲兵学院 | A kind of sorter model training method of radar target recognition |
CN109101108A (en) * | 2018-07-25 | 2018-12-28 | 重庆邮电大学 | Method and system based on three decision optimization intelligence cockpit human-computer interaction interfaces |
-
2019
- 2019-01-29 CN CN201910084149.5A patent/CN109816022A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH07152543A (en) * | 1993-11-30 | 1995-06-16 | Kobe Nippon Denki Software Kk | Decision table input device |
US5956428A (en) * | 1997-05-19 | 1999-09-21 | Ausbeck, Jr.; Paul James | Stroke code for image partitions |
CN106796647A (en) * | 2014-09-05 | 2017-05-31 | 北京市商汤科技开发有限公司 | Scene text detecting system and method |
CN108154172A (en) * | 2017-12-21 | 2018-06-12 | 重庆邮电大学 | Image-recognizing method based on three decisions |
CN108226889A (en) * | 2018-01-19 | 2018-06-29 | 中国人民解放军陆军装甲兵学院 | A kind of sorter model training method of radar target recognition |
CN109101108A (en) * | 2018-07-25 | 2018-12-28 | 重庆邮电大学 | Method and system based on three decision optimization intelligence cockpit human-computer interaction interfaces |
Non-Patent Citations (1)
Title |
---|
JIN DAI等: "An Novel Image Recognition method based on Three-way Decision", 《第六届国际(洛阳)工业设计高峰论坛(IFID2018)》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110473251A (en) * | 2019-08-14 | 2019-11-19 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Custom field spatial data area statistics method based on grid spatial index |
CN110473251B (en) * | 2019-08-14 | 2022-03-25 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Self-defined range spatial data area statistical method based on grid spatial index |
CN111046926A (en) * | 2019-11-26 | 2020-04-21 | 山东浪潮人工智能研究院有限公司 | Computer vision image classification integrated learning method |
CN111046926B (en) * | 2019-11-26 | 2023-09-19 | 山东浪潮科学研究院有限公司 | Computer vision image classification integrated learning method |
CN112580785A (en) * | 2020-12-18 | 2021-03-30 | 河北工业大学 | Neural network topological structure optimization method based on three-branch decision |
CN112580785B (en) * | 2020-12-18 | 2022-04-05 | 河北工业大学 | Neural network topological structure optimization method based on three-branch decision |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110119676B (en) | Driver fatigue detection method based on neural network | |
Lin et al. | Intelligent traffic-monitoring system based on YOLO and convolutional fuzzy neural networks | |
CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
CN109816022A (en) | A kind of image-recognizing method based on three decisions and CNN | |
CN107689158A (en) | A kind of intellectual traffic control method based on image procossing | |
CN109993061A (en) | A kind of human face detection and tracing method, system and terminal device | |
CN108154172B (en) | Image-recognizing method based on three decisions | |
CN109034020A (en) | A kind of community's Risk Monitoring and prevention method based on Internet of Things and deep learning | |
Yusof et al. | Automated asphalt pavement crack detection and classification using deep convolution neural network | |
CN111223087A (en) | Automatic bridge crack detection method based on generation countermeasure network | |
Luo et al. | An application of object detection based on YOLOv3 in traffic | |
CN115984537A (en) | Image processing method and device and related equipment | |
Hasanvand et al. | Machine learning methodology for identifying vehicles using image processing | |
Sikora et al. | Classification of railway level crossing barrier and light signalling system using YOLOv3 | |
Karthigadevi et al. | Speed Limit Signboard Recognition and Audio Alert for Driver Using CNN | |
de S. Matos et al. | Hierarchical classification of vehicle images using NN with conditional adaptive distance | |
CN112084897A (en) | Rapid traffic large-scene vehicle target detection method of GS-SSD | |
Mahenge et al. | RCNN-GAN: an enhanced deep learning approach towards detection of road cracks | |
Nhat-Duc et al. | Computer Vision-Based Severity Classification of Asphalt Pavement Raveling Using Advanced Gradient Boosting Machines and Lightweight Texture Descriptors | |
CN115393802A (en) | Railway scene unusual invasion target identification method based on small sample learning | |
CN115187886A (en) | Vehicle violation detection method and device and electronic equipment | |
Ranganayaki et al. | Machine Learning Approaches on Pedestrian Detection in an autonomous vehicle | |
Dorrani | Traffic Scene Analysis and Classification using Deep Learning | |
Kinattukara et al. | Clustering based neural network approach for classification of road images | |
Lin et al. | Application of the efficientdet algorithm in traffic flow statistics |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
CB02 | Change of applicant information |
Address after: 400020 Jiangbei District, Chongqing electric measuring Village No. 231 Applicant after: Chongqing geographic information and Remote Sensing Application Center (Chongqing surveying and mapping product quality inspection and testing center) Address before: 400000 electric survey village, Jiangbei District, Chongqing City, No. 231 Applicant before: Chongqing Geographical Information Center |
|
CB02 | Change of applicant information | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190528 |
|
RJ01 | Rejection of invention patent application after publication |