CN110097060A - A kind of opener recognition methods towards trunk image - Google Patents

A kind of opener recognition methods towards trunk image Download PDF

Info

Publication number
CN110097060A
CN110097060A CN201910243209.3A CN201910243209A CN110097060A CN 110097060 A CN110097060 A CN 110097060A CN 201910243209 A CN201910243209 A CN 201910243209A CN 110097060 A CN110097060 A CN 110097060A
Authority
CN
China
Prior art keywords
characteristic pattern
similarity
cluster
trunk
trunk image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910243209.3A
Other languages
Chinese (zh)
Other versions
CN110097060B (en
Inventor
陈晋音
林翔
贾澄钰
杨东勇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN201910243209.3A priority Critical patent/CN110097060B/en
Publication of CN110097060A publication Critical patent/CN110097060A/en
Application granted granted Critical
Publication of CN110097060B publication Critical patent/CN110097060B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of opener recognition methods towards trunk image, comprising the following steps: (1) CCN model is designed, after part training sample training CNN model, and construction feature extractor;(2) to the characteristic pattern extracted through feature extractor, the similarity between characteristic pattern is calculated, and characteristic pattern is clustered using DBSCAN algorithm;(3) Loss function is designed according to the cluster result of similarity and characteristic pattern between characteristic pattern, and according to the weight parameter of the parameter of the Loss function optimization feature extractor and similarity function;(4) trunk image to be sorted is input in the feature extractor of optimization, the similarity function of optimization is recycled to calculate the similarity between the characteristic pattern of output, the DBSCAN cluster result to characteristic pattern is finally obtained according to the similarity value for calculating acquisition.The opener recognition methods can realize the opener identification to the trunk image of unknown classification well.

Description

A kind of opener recognition methods towards trunk image
Technical field
The invention belongs to the deep learning algorithms of artificial intelligence field and opener identification research, and in particular to a kind of face Opener recognition methods to trunk image.
Background technique
For a long time, opener recognition methods is always a research hotspot in deep learning field.This method be intended to from Unknown sample is found in data and therefrom extracts useful feature information, is had in terms of signal identification and Face datection extensive Using.Jain etc. proposes the opener recognition methods based on more classification SVM, they are using SVM to Probabilistic Decision-making score It practises, and unknown sample is refused by setting probability threshold value.Bendale etc., which is proposed, carries out opener knowledge using deep neural network Method for distinguishing, they introduce OpenMax layers, and estimate that input sample belongs to the probability of unknown class with this layer.Ge etc. is utilized GAN generates unknown sample, reuses these data training neural network, is then combined it with OpenMax layers, while also using One threshold value refuses unknown sample.Kuniaki etc. is that known class and unknown sample set up frontiers using dual training method, And training generator makes target sample far from boundary.This method achieves on Office, VisDA and digits data set Relatively good experimental result.
Similarity calculation is a kind of method for measuring similar or different degree between data.Two things are closer, it Similarity measurement it is also bigger, and two things are more become estranged, their similarity measurement is also just smaller.Different similarities Calculation method reflects the degree of closeness of data from different level.Common similarity algorithm include Euclidean distance, Manhattan away from With a distance from, Minkowski, Pearson correlation coefficient, cosine similarity, Jaccard similarity factor etc., these calculation methods are used The similarity degree data is described in the mode of quantification, so that other algorithms classify to these data.
Cluster is that the set of physics or abstract object is divided into the process that multiple classes are made of similar object.Cluster is usual Be grouped according to the similitude between object, thus how the similitude between description object be cluster important prerequisite.K-means Algorithm is a kind of clustering algorithm of classics, and the computation complexity of this method is lower, and on various data sets can obtain compared with Good Clustering Effect.However, the cluster number K in algorithm generally requires artificially to give, which has limited K-means in actual scene In application range.Density clustering method can automatically determine class cluster number, and DBSCAN algorithm is the typical case of such methods It represents, it is defined as core point by introducing the reachable concept of density, the point that will be greater than a certain density, and adjacent core points are mutually straight Connect reachable, all mutually reachable points form a class clusters, effectively compensate for the deficiency of K-means algorithm, and enhance to making an uproar The processing capacity of sound point.
Convolutional neural networks are comprising convolutional calculation and with the feedforward neural network of depth structure, are the generations of deep learning One of table algorithm.Convolutional neural networks copy the vision mechanism of biology to be constructed, and can exercise supervision study and non-supervisory It practises, the sparsity that the convolution kernel parameter sharing in hidden layer is connected with interlayer enables convolutional neural networks with lesser calculating It measures to extract the feature of image.As the proposition of deep learning theory and numerical value calculate the improvement of equipment, convolutional neural networks are obtained Quick development has been arrived, and has been widely used in the fields such as computer vision, natural language processing.
Summary of the invention
The purpose of the present invention is designing a kind of opener recognition methods towards trunk image, pass through training designed, designed CNN model obtains the feature extractor of image, determines loss function in conjunction with similarity calculating method and DBSCAN algorithm, passes through instruction White silk further corrects the parameter of feature extractor, finally carries out feature extraction to the data of all categories on test set And cluster, the opener of trunk image is identified to realize.
The technical solution of the present invention is as follows:
A kind of opener recognition methods towards trunk image, comprising the following steps:
(1) the CCN model towards trunk image is designed, after being trained using part training sample to CNN model, and root According to trained CNN Construction of A Model feature extractor;
(2) to the characteristic pattern extracted through feature extractor, the similarity between characteristic pattern is calculated, and according to similarity value, Characteristic pattern is clustered using DBSCAN algorithm;
(3) Loss function is designed according to the cluster result of similarity and characteristic pattern between characteristic pattern, and according to the Loss Function and remaining training sample optimize the parameter of feature extractor and the weight parameter of similarity function;
(4) trunk image to be sorted is input in the feature extractor of optimization, is computed and obtains corresponding characteristic pattern, The similarity function of recycling optimization calculates the similarity between characteristic pattern, finally according to the similarity value acquisition pair for calculating acquisition The DBSCAN cluster result of characteristic pattern.
Opener recognition methods provided by the invention to trunk image can accurately extract inhomogeneity using feature extractor Other otherness feature, and unknown class cluster can be captured by DBSCAN algorithm, it realizes and the opener of trunk data set is identified.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technical description to do simply to introduce, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments of invention for those of ordinary skill in the art, can be with root under the premise of not making the creative labor Other accompanying drawings are obtained according to these attached drawings.
Fig. 1 is the flow chart of the opener recognition methods provided by the invention towards trunk image;
Fig. 2 is the structural schematic diagram for the CNN model that the present invention constructs;
Fig. 3 is the real image in experimental example for 4 class trunks of training;
Fig. 4 is the real image in experimental example for the other 3 class trunk of opener identification.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention more comprehensible, with reference to the accompanying drawings and embodiments to this Invention is described in further detail.It should be appreciated that the specific embodiments described herein are only used to explain the present invention, And the scope of protection of the present invention is not limited.
Present invention combination convolutional neural networks and density-based algorithms provide a kind of opening towards trunk image Set identification method is trained feature extractor and similarity function using the trunk image of known class, utilizes training Good feature extractor and similarity function, in conjunction with clustering method can the trunk image to unknown classification identify, It realizes and the opener of trunk image is identified.
As shown in Figure 1, the opener recognition methods provided by the invention towards trunk image comprising the steps of:
S101 designs the CNN model towards trunk image.
It is repeatedly inquired into through testing, in order to promote the comprehensive and accuracy for extracting trunk characteristics of image, design is as shown in Figure 2 CNN model, specifically, CNN model includes sequentially connected input layer, the first convolutional layer, the first maximum pond layer, volume Two Lamination, the second maximum pond layer, third convolutional layer, third maximum pond layer, Volume Four lamination, the first full articulamentum, second are entirely Articulamentum and output layer, wherein the size of input layer is 256 × 256 × 3, that is, requires to input 3 Color Channels expressions, rulers The very little trunk image for being 256 × 256;The template size of first convolution kernel is 11 × 11, port number 64, and convolution step-length is 4, warp The size for crossing image after the first convolution operation becomes 64 × 64 × 64;The pond of the latter linked first maximum pond of first convolutional layer layer Changing step-length is 2, and by the first Chi Huahou, the size of trunk image becomes 32 × 32 × 64;The template size of second convolution kernel is 5 × 5, port number 128, the picture size after the second convolution operation is 32 × 32 × 128;Is connected after second convolutional layer The pond step-length of two maximum pond layers is 2, and by the second Chi Huahou, the size of trunk image becomes 16 × 16 × 128;Third volume The template size of product core is 3 × 3, port number 192, and the picture size after third convolution operation is 16 × 16 × 192;The The pond step-length of three convolutional layers latter linked third maximum pond layer is 2, and by third Chi Huahou, the size of trunk image becomes 8×8×192;The size of Volume Four product core is 3 × 3, port number 256, and the picture size after Volume Four product operation is 8 ×8×256;The first full articulamentum is connected after Volume Four lamination, the image after first overall situation is average becomes 256 nerves 256 neurons are then mapped to 128 neurons by member, the full articulamentum of the second layer, finally by this 128 neurons and output layer It is connected, through being exported as a result, the number of the neuron of output layer is corresponding with the classification number of classification after softmax.
S102 is trained CNN model using part training sample, and construction feature extractor.
After designing CNN model, i.e., CNN model is trained, specifically, using the tree comprising part trunk classification The training sample of dry image and the practical category composition of corresponding trunk image is trained CNN model, loss when training Function is the cross entropy for predicting category and practical category:
Wherein, p (i) indicates the practical category of i-th of trunk image, and q (i) indicates i-th of trunk image through CNN model meter The prediction category of output is calculated, cross entropy L is smaller, then it represents that trunk image more connects with the characteristic pattern distribution to trunk image zooming-out Closely, i.e. the prediction result of CNN model is more accurate.
After obtaining trained CNN model, remove the second full articulamentum in CNN model, remaining network structure with And corresponding network architecture parameters constitute the feature extractor of trunk image.
S103 calculates the similarity between characteristic pattern to the characteristic pattern extracted through feature extractor, and according to similarity Value, clusters characteristic pattern using DBSCAN algorithm.
After feature extractor extracts and obtains characteristic pattern, find to be to be between characteristic pattern there are certain similarity difference From the similarity difference measured between different data with multi-angle, the present invention is linearly folded using cosine similarity and space similarity The mode added obtains the similarity between two characteristic patterns, specifically, the calculation formula of similarity are as follows:
D(Ai,Aj)=- α [d (Ai,Aj)c+1]/2+βd(Ai,Aj)n (2)
Wherein, d (Ai,Aj)cIndicate ith feature figure AiWith j-th of characteristic pattern AjBetween cosine similarity, it use two Angle between a vector describes the otherness between two characteristic patterns, specifically, d (Ai,Aj)cAre as follows:
It can be obtained from formula (3), d (Ai,Aj)cValue be [- 1,1], d (Ai,Aj)cValue it is smaller, then characteristic pattern AiAnd jth A characteristic pattern AjSimilitude it is smaller.
For the ease of the use of DBSCAN algorithm later, to certain processing is carried out in formula (2), i.e., using-[d (Ai,Aj)c+ 1] function D (A/2 is replacedi,Aj) in d (Ai,Aj)cIt is calculated, at this time-[d (Ai,Aj)c+ 1] value range/2 For [0,1], and-[d (Ai,Aj)c+ 1] value/2 is smaller, then characteristic pattern AiWith j-th of characteristic pattern AjSimilitude it is bigger, except this it Outside, processing can guarantee D (A in this wayi,Aj) be positive always.
d(Ai,Aj)nIndicate ith feature figure AiWith j-th of characteristic pattern AjBetween Euclidean distance, as space phase knowledge and magnanimity Value, for measuring absolute distance of two characteristic patterns in hyperspace, its calculation formula is:
Wherein, r indicates characteristic pattern AiWith characteristic pattern AjData dimension,WithRespectively indicate characteristic pattern AiAnd characteristic pattern AjValue in kth dimension,Indicate characteristic pattern AiWith characteristic pattern AjIn kth dimension Distance after normalization.It is easy to get from formula (4), d (Ai,Aj)nValue be [0,1], d (Ai,Aj)nValue it is smaller, then it is special Sign figure AiWith characteristic pattern AjSimilitude it is bigger.
In formula (2), α and β are two weight coefficients, for adjusting Euclidean distance and cosine similarity in similarity meter Proportion in calculation, and have α >=0, β >=0, alpha+beta=1 under initial situation, enable α=β=0.5, weigh during successive iterations Weight parameter alpha and β will be obtained optimization.
After obtaining the similarity between characteristic pattern, according to similarity value, characteristic pattern is gathered using DBSCAN algorithm Class, specifically, DBSCAN algorithm includes:
Define 1 (Eps neighborhood): for a characteristic pattern Ai, characteristic pattern AiEps neighborhood definition be with characteristic pattern AiFor core The heart, using Eps as the hypersphere body region of radius, i.e.,
NEps(Ai)={ Aj∈M|D(Ai,Aj)≤Eps} (5)
Wherein, M indicates all feature set of graphs;
Define 2 (core point and boundary points): for a characteristic pattern Ai, an integer minPts is given, if characteristic pattern Ai's Object in Eps neighborhood meets
|NEps(Ai)|≥min Pts (6)
Then claim characteristic pattern AiFor the core point under the conditions of (Eps, minPts);
Define 3 (density is through): if characteristic pattern AiPositioned at characteristic pattern AjEps neighborhood in, and AjIt is kernel object, then claims Characteristic pattern AiBy characteristic pattern AjDensity is through;
Define 4 (density is reachable): for characteristic pattern AiWith characteristic pattern Aj, sample sequence p if it exists1,p2,...,pt, pt+1,...,pT, meet p1=Ai, pT=Aj, and pt+1By ptDensity is through, then claims AjBy AiDensity is reachable;It is tool that i.e. density is reachable There is the density of transitivity through;
Work as Eps, after minPts is determined, for the characteristic pattern extracted through feature extractor, is made by any one characteristic pattern It is starting for core point, a class cluster will be constituted with the reachable all characteristic patterns of the core dot density, when all core points quilt When being divided to certain a kind of cluster, cluster is finished, and class cluster number will be determined automatically at this time.
S104 designs Loss function, and according to Loss function optimization feature extractor parameter.
When designing Loss function, according to the cluster result of S103, the core point for choosing all class clusters carries out Loss function Calculating, and loss function is optimized using the optimization method based on gradient, specifically, Loss function are as follows:
Wherein, H indicates the set of all core points, g (Ai,Aj) it is sign function, as characteristic pattern AiWith characteristic pattern AjBelong to When same class cluster, then g (Ai,Aj)=1;As characteristic pattern AiWith characteristic pattern AjWhen being not belonging to same class cluster, then g (Ai,Aj)=- 1.It follows that the purpose of L ' is to make the distance between same cluster core point small as far as possible, the distance between different cluster core points are to the greatest extent Amount is big, so that training characteristics extractor is to improve Clustering Effect.
Using remaining training sample and the Loss function of determination, feature extractor is carried out using gradient descent algorithm Parameter optimization, the feature extractor after optimization can reinforce the otherness feature extraction to each trunk classification.
S105 optimizes the weight parameter of similarity function.
When optimizing to feature extractor, when L ' tends towards stability, stop the parameter optimization to feature extractor, i.e., Obtain the feature extractor of optimization.In optimization feature extractor and then carry out similarity function D (Ai,Aj) in weight parameter α With the optimization of β, due to α >=0, β >=0, alpha+beta=1, then similarity function D (Ai,Aj) write as:
D(Ai,Aj)=- α [d (Ai,Aj)c+1]/2+(1-α)d(Ai,Aj)n (8)
Firstly, Δ α=0.1 is now enabled, from initial value α0=0.5 starts, and calculates separately in α0Locate, in α0At+Δ α and α0-Δ Cluster accuracy rate at α, corresponding α is as α when choosing cluster accuracy rate highest1
Then, according to α1With α0Δ α is updated, and is calculated in α1Locate, in α1At+Δ α and α1Cluster accuracy rate at Δ α, choosing Corresponding α is as α when taking cluster accuracy rate highest2, iteration process following procedure: i.e. according to αmWith αm-1Δ α is updated, and is counted It calculates in αmLocate, in αmAt+Δ α and αmCluster accuracy rate at Δ α, corresponding α is as α when choosing cluster accuracy rate highestm+1, Until finding optimal α, the similarity function D (A of optimal α, β=1- α and optimization is thus obtainedi,Aj)。
When calculating cluster preparation rate, cluster is calculated according to the corresponding practical category of characteristic pattern and acquisition cluster result Accuracy rate.
S106 is computed and is corresponded in application, trunk image to be sorted is input in the feature extractor of optimization Characteristic pattern, recycle optimization similarity function calculate characteristic pattern between similarity, finally according to calculate obtain it is similar Angle value obtains the DBSCAN cluster result to characteristic pattern.
S107 exports cluster result.
Experimental example
In order to verify the classifying quality of the above-mentioned opener recognition methods towards trunk image, above-mentioned opener recognition methods is existed The trunk of actual acquisition on collection according to being tested.
Trunk data set is the data set of team itself statistical collection, which includes 13 kinds of trunks, and every kind of trunk contains There are about 200~500 trunk RPG images.In experiment, chooses preceding 7 class trunk and tested as test object.In order to make the number Reach trainable requirement according to collection, data extending is carried out to such trunk image by intercepting part trunk at random, is then carried out All trunk image processings are finally 256 × 256 sizes by gray processing processing.Finally obtain 500 picture of every kind of trunk, one The data set of totally 3500 trunks.In experiment, before the data of preceding 4 kinds of trunks as shown in Figure 3 80% (totally 1600 figures are used Piece) CNN model is trained as training set, and (i.e. as shown in Figure 4 by after 4 class trunks 20% data and other three classes Classification trunk image) 20% data (totally 700 picture) detected as test set opener identification effect.Through counting, Final cluster accuracy rate has reached 64.3%, and average cluster purity is 61.5%.The cluster accuracy rate of three classes opener trunk point It Wei 58%, 54% and 49%.
The experimental example sufficiently shows the opener recognition methods provided by the invention towards trunk image can be real well Now the opener of the trunk image of unknown classification is identified.
Technical solution of the present invention and beneficial effect is described in detail in above-described specific embodiment, Ying Li Solution is not intended to restrict the invention the foregoing is merely presently most preferred embodiment of the invention, all in principle model of the invention Interior done any modification, supplementary, and equivalent replacement etc. are enclosed, should all be included in the protection scope of the present invention.

Claims (10)

1. a kind of opener recognition methods towards trunk image, comprising the following steps:
(1) the CCN model towards trunk image is designed, after being trained using part training sample to CNN model, and according to instruction Experienced CNN Construction of A Model feature extractor;
(2) to the characteristic pattern extracted through feature extractor, the similarity between characteristic pattern is calculated, and according to similarity value, is used DBSCAN algorithm clusters characteristic pattern;
(3) Loss function is designed according to the cluster result of similarity and characteristic pattern between characteristic pattern, and according to the Loss function And remaining training sample optimizes the parameter of feature extractor and the weight parameter of similarity function;
(4) trunk image to be sorted is input in the feature extractor of optimization, is computed and obtains corresponding characteristic pattern, then benefit The similarity between characteristic pattern is calculated with the similarity function of optimization, is finally obtained according to the similarity value for calculating acquisition to feature The DBSCAN cluster result of figure.
2. as described in claim 1 towards the opener recognition methods of trunk image, which is characterized in that CNN model includes successively The input layer of connection, the first convolutional layer, the first maximum pond layer, the second convolutional layer, the second maximum pond layer, third convolutional layer, Third maximum pond layer, Volume Four lamination, the first full articulamentum, the second full articulamentum and output layer, wherein the ruler of input layer Very little is 256 × 256 × 3, that is, require to input 3 Color Channels indicate, having a size of 256 × 256 trunk image;First convolution kernel Template size be 11 × 11, port number 64, convolution step-length be 4, the size of image becomes 64 after the first convolution operation ×64×64;The pond step-length of the latter linked first maximum pond of first convolutional layer layer is 2, by the first Chi Huahou, trunk figure The size of picture becomes 32 × 32 × 64;The template size of second convolution kernel is 5 × 5, and port number 128 is grasped by the second convolution Picture size after work is 32 × 32 × 128;The pond step-length of the maximum pond layer of connection second is 2 after second convolutional layer, is passed through The size of second Chi Huahou, trunk image become 16 × 16 × 128;The template size of third convolution kernel is 3 × 3, and port number is 192, the picture size after third convolution operation is 16 × 16 × 192;Third convolutional layer latter linked third maximum pond The pond step-length of layer is 2, and by third Chi Huahou, the size of trunk image becomes 8 × 8 × 192;Volume Four product core size be 3 × 3, port number 256, the picture size after Volume Four product operation is 8 × 8 × 256;First is connected after Volume Four lamination Full articulamentum, the image after first overall situation is average become 256 neurons, and the full articulamentum of the second layer is then by 256 nerves Member maps to 128 neurons, and finally this 128 neurons are connected with output layer, through obtaining output knot after softmax Fruit, the number of the neuron of output layer and the classification number of classification are corresponding.
3. as claimed in claim 2 towards the opener recognition methods of trunk image, which is characterized in that in step (1), using packet The training sample of the trunk image of the classification of trunk containing part and the practical category composition of corresponding trunk image carries out CNN model Training, loss function when training are the cross entropy for predicting category and practical category:
Wherein, p (i) indicates the practical category of i-th of trunk image, and it is defeated that q (i) indicates that i-th of trunk image is calculated through CNN model Prediction category out;
After obtaining trained CNN model, the second full articulamentum in CNN model is removed, remaining network structure and right The network architecture parameters answered constitute the feature extractor of trunk image.
4. as described in claim 1 towards the opener recognition methods of trunk image, which is characterized in that using cosine similarity and The mode of space similarity linear superposition obtains the similarity between two characteristic patterns, specifically, the calculation formula of similarity Are as follows:
D(Ai,Aj)=- α [d (Ai,Aj)c+1]/2+βd(Ai,Aj)n (2)
Wherein, d (Ai,Aj)cIndicate ith feature figure AiWith j-th of characteristic pattern AjBetween cosine similarity, d (Ai,Aj)nIt indicates Ith feature figure AiWith j-th of characteristic pattern AjBetween Euclidean distance, α and β are two weight coefficients, for adjusting Euclidean distance With proportion of the cosine similarity in similarity calculation, and there is α >=0, β >=0, alpha+beta=1.
5. as claimed in claim 4 towards the opener recognition methods of trunk image, which is characterized in that described to be calculated using DBSCAN Method carries out cluster to characteristic pattern
Define 1 (Eps neighborhood): for a characteristic pattern Ai, characteristic pattern AiEps neighborhood definition be with characteristic pattern AiFor core, with Eps is the hypersphere body region of radius, i.e.,
NEps(Ai)={ Aj∈M|D(Ai,Aj)≤Eps} (3)
Wherein, M indicates all feature set of graphs;
Define 2 (core point and boundary points): for a characteristic pattern Ai, an integer minPts is given, if characteristic pattern AiEps Object in neighborhood meets
|NEps(Ai)|≥minPts (4)
Then claim characteristic pattern AiFor the core point under the conditions of (Eps, minPts);
Define 3 (density is through): if characteristic pattern AiPositioned at characteristic pattern AjEps neighborhood in, and AjIt is kernel object, then claims feature Scheme AiBy characteristic pattern AjDensity is through;
Define 4 (density is reachable): for characteristic pattern AiWith characteristic pattern Aj, sample sequence p if it exists1,p2,...,pt,pt+1,..., pT, meet p1=Ai, pT=Aj, and pt+1By ptDensity is through, then claims AjBy AiDensity is reachable;It is with transitivity that i.e. density is reachable Density it is through;
Work as Eps, after minPts is determined, for the characteristic pattern extracted through feature extractor, by any one characteristic pattern as core Heart point is starting, a class cluster will be constituted with the reachable all characteristic patterns of the core dot density, when all core points are divided When to certain a kind of cluster, cluster is finished, and class cluster number will be determined automatically at this time.
6. as claimed in claim 5 towards the opener recognition methods of trunk image, which is characterized in that use-[d (Ai,Aj)c+ 1] function D (A/2 is replacedi,Aj) in d (Ai,Aj)cIt is calculated, to guarantee D (Ai,Aj) be positive always.
7. as claimed in claim 4 towards the opener recognition methods of trunk image, which is characterized in that Loss function are as follows:
Wherein, H indicates the set of all core points, g (Ai,Aj) it is sign function, as characteristic pattern AiWith characteristic pattern AjBelong to same When a class cluster, then g (Ai,Aj)=1;As characteristic pattern AiWith characteristic pattern AjWhen being not belonging to same class cluster, then g (Ai,Aj)=- 1.
8. as claimed in claim 7 towards the opener recognition methods of trunk image, which is characterized in that utilize remaining trained sample Sheet and the Loss function determined carry out parameter optimization to feature extractor using gradient descent algorithm.
9. as claimed in claim 8 towards the opener recognition methods of trunk image, which is characterized in that in optimization feature extractor And then carry out similarity function D (Ai,Aj) in weight parameter α and β optimization, due to α >=0, β >=0, alpha+beta=1 is then similar Spend function D (Ai,Aj) write as:
D(Ai,Aj)=- α [d (Ai,Aj)c+1]/2+(1-α)d(Ai,Aj)n (6)
Firstly, Δ α=0.1 is now enabled, from initial value α0=0.5 starts, and calculates separately in α0Locate, in α0At+Δ α and α0At Δ α Accuracy rate is clustered, corresponding α is as α when choosing cluster accuracy rate highest1
Then, according to α1With α0Δ α is updated, and is calculated in α1Locate, in α1At+Δ α and α1Cluster accuracy rate at Δ α is chosen poly- Corresponding α is as α when class accuracy rate highest2, iteration process following procedure: i.e. according to αmWith αm-1Δ α is updated, and is calculated αmLocate, in αmAt+Δ α and αmCluster accuracy rate at Δ α, corresponding α is as α when choosing cluster accuracy rate highestm+1, until Until finding optimal α, the similarity function D (A of optimal α, β=1- α and optimization is thus obtainedi,Aj)。
10. as claimed in claim 9 towards the opener recognition methods of trunk image, which is characterized in that prepare calculating cluster When rate, cluster accuracy rate is calculated according to the corresponding practical category of characteristic pattern and acquisition cluster result.
CN201910243209.3A 2019-03-28 2019-03-28 Open set identification method for trunk image Active CN110097060B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910243209.3A CN110097060B (en) 2019-03-28 2019-03-28 Open set identification method for trunk image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910243209.3A CN110097060B (en) 2019-03-28 2019-03-28 Open set identification method for trunk image

Publications (2)

Publication Number Publication Date
CN110097060A true CN110097060A (en) 2019-08-06
CN110097060B CN110097060B (en) 2021-04-06

Family

ID=67444070

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910243209.3A Active CN110097060B (en) 2019-03-28 2019-03-28 Open set identification method for trunk image

Country Status (1)

Country Link
CN (1) CN110097060B (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110457511A (en) * 2019-08-16 2019-11-15 成都数之联科技有限公司 Image classification method and system based on attention mechanism and generation confrontation network
CN111179335A (en) * 2019-12-28 2020-05-19 东北林业大学 Standing tree measuring method based on binocular vision
CN111428760A (en) * 2020-03-10 2020-07-17 上海集成电路研发中心有限公司 Feature similarity based self-adaptive clustering method and application
CN111914919A (en) * 2020-07-24 2020-11-10 天津大学 Open set radiation source individual identification method based on deep learning
CN112508062A (en) * 2020-11-20 2021-03-16 普联国际有限公司 Open set data classification method, device, equipment and storage medium
CN113469276A (en) * 2021-07-21 2021-10-01 苏州大学 Fruit tree detection method and device
CN113822390A (en) * 2021-11-24 2021-12-21 杭州贝嘟科技有限公司 User portrait construction method and device, electronic equipment and storage medium
CN113947578A (en) * 2021-10-18 2022-01-18 山东大学 Prediction method of nucleation rate of vermicular cast iron based on DBSCAN clustering algorithm

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107729799A (en) * 2017-06-13 2018-02-23 银江股份有限公司 Crowd's abnormal behaviour vision-based detection and analyzing and alarming system based on depth convolutional neural networks
US9964607B1 (en) * 2017-08-28 2018-05-08 Seagate Technology Llc Recognizing and identifying defect patterns on magnetic media
US10025950B1 (en) * 2017-09-17 2018-07-17 Everalbum, Inc Systems and methods for image recognition
CN109214422A (en) * 2018-08-02 2019-01-15 深圳先进技术研究院 Parking data method for repairing and mending, device, equipment and storage medium based on DCGAN
US10235602B1 (en) * 2018-03-08 2019-03-19 Capital One Services, Llc Machine learning artificial intelligence system for identifying vehicles

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107729799A (en) * 2017-06-13 2018-02-23 银江股份有限公司 Crowd's abnormal behaviour vision-based detection and analyzing and alarming system based on depth convolutional neural networks
US9964607B1 (en) * 2017-08-28 2018-05-08 Seagate Technology Llc Recognizing and identifying defect patterns on magnetic media
US10025950B1 (en) * 2017-09-17 2018-07-17 Everalbum, Inc Systems and methods for image recognition
US10235602B1 (en) * 2018-03-08 2019-03-19 Capital One Services, Llc Machine learning artificial intelligence system for identifying vehicles
CN109214422A (en) * 2018-08-02 2019-01-15 深圳先进技术研究院 Parking data method for repairing and mending, device, equipment and storage medium based on DCGAN

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
CHEN JINYIN等: ""A novel cluster center fast determination clustering algorithm"", 《APPLIED SOFT COMPUTING》 *
CHENBO FU等: ""Link weight prediction using supervised learning methods and its application to yelp layered network"", 《IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING》 *
JINYIN CHEN等: ""A novel image segmentation method based on fast density clustering algorithm"", 《ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE》 *
JINYIN CHEN等: ""FGCH: a fast and grid based clustering algorithm for hybrid data stream"", 《APPLIED INTELLIGENCE》 *
JUN HAN等: ""FlowNet: A deep learning framework for clustering and selection of streamlines and stream surfaces"", 《IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS》 *
刘超等: ""超低照度下微光图像增强神经网络损失函数设计分析"", 《国防科技大学学报》 *
李真等: ""树干与地面点云分类K-means方法的改进"", 《东北林业大学学报》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110457511A (en) * 2019-08-16 2019-11-15 成都数之联科技有限公司 Image classification method and system based on attention mechanism and generation confrontation network
CN110457511B (en) * 2019-08-16 2022-12-06 成都数之联科技股份有限公司 Image classification method and system based on attention mechanism and generation countermeasure network
CN111179335A (en) * 2019-12-28 2020-05-19 东北林业大学 Standing tree measuring method based on binocular vision
CN111428760A (en) * 2020-03-10 2020-07-17 上海集成电路研发中心有限公司 Feature similarity based self-adaptive clustering method and application
CN111428760B (en) * 2020-03-10 2023-09-08 上海集成电路研发中心有限公司 Self-adaptive clustering method based on feature similarity and application
CN111914919A (en) * 2020-07-24 2020-11-10 天津大学 Open set radiation source individual identification method based on deep learning
CN112508062A (en) * 2020-11-20 2021-03-16 普联国际有限公司 Open set data classification method, device, equipment and storage medium
CN113469276A (en) * 2021-07-21 2021-10-01 苏州大学 Fruit tree detection method and device
CN113947578A (en) * 2021-10-18 2022-01-18 山东大学 Prediction method of nucleation rate of vermicular cast iron based on DBSCAN clustering algorithm
CN113947578B (en) * 2021-10-18 2024-08-09 山东大学 Vermicular graphite cast iron nucleation rate prediction method based on DBSCAN clustering algorithm
CN113822390A (en) * 2021-11-24 2021-12-21 杭州贝嘟科技有限公司 User portrait construction method and device, electronic equipment and storage medium
CN113822390B (en) * 2021-11-24 2022-03-15 杭州贝嘟科技有限公司 User portrait construction method and device, electronic equipment and storage medium

Also Published As

Publication number Publication date
CN110097060B (en) 2021-04-06

Similar Documents

Publication Publication Date Title
CN110097060A (en) A kind of opener recognition methods towards trunk image
CN112308158B (en) Multi-source field self-adaptive model and method based on partial feature alignment
CN106779087B (en) A kind of general-purpose machinery learning data analysis platform
CN110348399B (en) Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network
CN109034205A (en) Image classification method based on the semi-supervised deep learning of direct-push
CN114841257B (en) Small sample target detection method based on self-supervision comparison constraint
CN106960214A (en) Object identification method based on image
CN109543723B (en) Robust image clustering method
CN110807760B (en) Tobacco leaf grading method and system
CN110135459A (en) A kind of zero sample classification method based on double triple depth measure learning networks
CN110866134B (en) Image retrieval-oriented distribution consistency keeping metric learning method
CN102324038A (en) A kind of floristics recognition methods based on digital picture
CN106250925B (en) A kind of zero Sample video classification method based on improved canonical correlation analysis
CN110297888A (en) A kind of domain classification method based on prefix trees and Recognition with Recurrent Neural Network
CN110321451A (en) Image retrieval algorithm based on Distribution Entropy gain loss function
Huang et al. MultiSpectralNet: Spectral clustering using deep neural network for multi-view data
CN109993208A (en) A kind of clustering processing method having noise image
CN113628197A (en) Weakly supervised full-section histopathology image classification method based on contrast learning
CN110991554B (en) Improved PCA (principal component analysis) -based deep network image classification method
CN115393631A (en) Hyperspectral image classification method based on Bayesian layer graph convolution neural network
CN109948662A (en) A kind of facial image depth clustering method based on K-means and MMD
CN112149556B (en) Face attribute identification method based on deep mutual learning and knowledge transfer
CN109947960A (en) The more attribute Combined estimator model building methods of face based on depth convolution
CN113707213A (en) Protein-ligand binding site prediction method based on deep learning
CN116229330A (en) Method, system, electronic equipment and storage medium for determining video effective frames

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant