CN110222778B - Online multi-view classification method, system and device based on deep forest - Google Patents
Online multi-view classification method, system and device based on deep forest Download PDFInfo
- Publication number
- CN110222778B CN110222778B CN201910502702.2A CN201910502702A CN110222778B CN 110222778 B CN110222778 B CN 110222778B CN 201910502702 A CN201910502702 A CN 201910502702A CN 110222778 B CN110222778 B CN 110222778B
- Authority
- CN
- China
- Prior art keywords
- view
- node
- classification
- forest
- updating
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 79
- 238000004040 coloring Methods 0.000 claims abstract description 9
- 238000007637 random forest analysis Methods 0.000 claims description 31
- 238000005070 sampling Methods 0.000 claims description 25
- 238000012549 training Methods 0.000 claims description 14
- 230000000007 visual effect Effects 0.000 claims description 14
- 239000000126 substance Substances 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000013136 deep learning model Methods 0.000 abstract description 3
- 238000010801 machine learning Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 10
- 230000008569 process Effects 0.000 description 10
- SZKKRCSOSQAJDE-UHFFFAOYSA-N Schradan Chemical compound CN(C)P(=O)(N(C)C)OP(=O)(N(C)C)N(C)C SZKKRCSOSQAJDE-UHFFFAOYSA-N 0.000 description 6
- 230000011218 segmentation Effects 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 230000010287 polarization Effects 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000002790 cross-validation Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241001270131 Agaricus moelleri Species 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 244000025254 Cannabis sativa Species 0.000 description 1
- PEDCQBHIVMGVHV-UHFFFAOYSA-N Glycerine Chemical compound OCC(O)CO PEDCQBHIVMGVHV-UHFFFAOYSA-N 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000009435 building construction Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of machine learning, and particularly relates to an online multi-view classification method, system and device based on deep forests, aiming at solving the problems that a deep learning model cannot be updated online and a shallow online learning model is low in classification accuracy. The method comprises the following steps: acquiring a classification result graph of multi-view data through a multi-view depth forest network; arranging the prediction labels of the classification result picture into the same size with the multi-view data picture according to columns; and coloring the arranged classification result graph into a gray classification result graph according to the size of the label and outputting the gray classification result graph. The invention learns from the data stream in an online mode, updates the model structure and the weight, ensures that the model has strong adaptability and expansibility on different data sets, simultaneously fully utilizes the multi-view data and the information among the multi-view data, and effectively develops deep characteristic information, thereby obtaining higher online classification accuracy, and the online mode does not need to store all data, thereby effectively saving resources.
Description
Technical Field
The invention belongs to the technical field of machine learning, and particularly relates to an online multi-view classification method, system and device based on deep forests.
Background
At present, the deep neural network method is widely applied in many fields and has achieved great success, especially in the fields of images and voice. Although the deep neural network is very powerful, there are many drawbacks, such as the training process requires a lot of training data and equipment with powerful computing power, the model is very complex, the hyper-parameters are too many and different tasks are sensitive to different parameters. Therefore, a deep forest network [1] with a cascade structure is designed by Zhou et al, so that the hyper-parameters are less, and the parameters are more robust to different tasks; the training costs are controllable and are applicable to small-scale data sets. However, the deep forest algorithm is an off-line learning method, which requires all training data to be available at the beginning of learning, and is not suitable for the situation that data is continuously acquired in a data stream form in an actual situation.
The online learning can effectively process the real-time data flow problem, and is a research hotspot in the field of machine learning. Online learning can incrementally learn classification models from data streams and does not reuse previous samples, applicable to dynamically growing data sets. The existing methods comprise a perceptron algorithm, an online passive attack algorithm, a support vector machine method based on convex hull vertex selection, an online random forest method [2] and the like. Data acquired from different information sources, spaces and modalities are becoming more and more abundant, and the data with different attributes form a multi-view data set. Compared with single-view learning, multi-view learning can explore useful features of each view to improve learning ability. The real-time updating method of the polarized SAR data classifier based on multi-view learning utilizes the consistency and complementarity between views to effectively improve the classification precision, however, the model is a shallow linear model, and the classification performance is not good enough [3 ]. The method takes the online classification of the polarized SAR data as the background, the data is firstly subjected to superpixel segmentation before classification, and the self-adaptive local iterative clustering method [4] is an effective polarized SAR data superpixel segmentation method and can improve the efficiency and robustness of subsequent classification.
In general, the existing deep learning model cannot be updated on line, and the classification accuracy of the shallow online learning model is not high enough.
The following documents are background information related to the present invention:
[1]Zhou Z H,Feng J.Deep forest:towards an alternative to deep neural networks.Proceedings of the 26th International Joint Conference on Artificial Intelligence.AAAI Press,2017:3553-3559.
[2]B.Lakshminarayanan,D.M.Roy,and Y.W.The.Mondrian forests:Efficient online random forests.Advances in neural information processing systems,2014.
[3] ney auspicious, yellow summer moon, Dingshuang, Qiaohong, Zhangbo: a polarized SAR data classifier real-time updating method based on multi-view learning, 2017-12-29.
[4]Xiang D,Ban Y,Wang W,et al.Adaptive superpixel generation for polarimetric SAR images with local iterative clustering and SIRV model[J].IEEE Transactions on Geoscience and Remote Sensing,2017,55(6):3115-3131.
Disclosure of Invention
In order to solve the above problems in the prior art, namely, the problem that a deep learning model cannot be updated online and the classification accuracy of a shallow online learning model is low, the invention provides an online multi-view classification method based on a deep forest, which comprises the following steps:
step S10, acquiring the characteristics of the same target with different attributes or the data of different sensors as input multi-view data;
step S20, based on the input multi-view data, obtaining a classification result graph of the input multi-view data through a multi-view depth forest network; the multi-view depth forest network comprises a first preset number of layers, and each layer comprises a second preset number of random forests;
step S30, arranging the prediction labels of the classification result picture in a column with the same size as the input multi-view data picture;
in step S40, the arranged classification result maps are colored into gray classification result maps according to the size of the label.
In some preferred embodiments, the training method of the multi-view deep forest network is as follows:
step B10, initializing weights of different visual angles of the multi-visual angle depth forest network, and obtaining the initialized multi-visual angle depth forest network;
b20, updating a multi-view depth forest network structure at the t moment based on the multi-view depth forest network at the t-1 moment, the multi-view data at the t moment and a corresponding real label, acquiring a prediction label of the multi-view data at the t moment, calculating a loss function of the comparison between the prediction label and the real label, and updating weights of different views;
and step B30, repeating step B20 until a preset training time is reached or the loss function value of the prediction tag is lower than a set threshold value, wherein t is t + 1.
In some preferred embodiments, step B10 "initialize the weights between different perspectives of the multi-perspective depth forest network" by:
in some preferred embodiments, in step B20, "updating the structure of the multi-view depth forest network" includes:
step B211, the structure of the current tree in the current layer is:
T=(T,δ,ξ,τ)
wherein T represents the current tree, delta represents the dimension of occurrence of node splitting of the tree, xi represents the position of occurrence of node splitting of the tree, and tau represents the time of occurrence of node splitting of the tree;
step B212, for the input sample D of the ith view angle(i)=(x(i)Y), at the jth node of the tree, set:
wherein the content of the first and second substances,are input samples x respectively(i)The lower and upper bounds of the element-by-element calculation of the space in which the device is located; e.g. of the typel、euAre input samples x respectively(i)The difference between the current node and the lower and upper bounds when the current node falls outside the given interval, wherein j is the jth node of the tree;
step B213, from the parametersThe method comprises the following steps of sampling in exponential distribution, obtaining splitting time E which needs to be increased when nodes of the tree are split, and updating as follows:
if τparen(t)j+E<τjThen the sampling yields the probability and of the splitting dimension δ, δ ═ dIn direct proportion, a set sampling interval is selected for sampling to obtain a splitting position, a new father node is inserted above the current node, and a new leaf node is generated below the father node; if τparent(j)+E≥τjUpdating the upper and lower bounds of the input space, judging whether the current node j is a leaf node, if so, stopping updating the random forest of the current layer, and if not, continuing to iterate downwards;
wherein, taujIs the split time of internal node j, parent (j) is the parent of node j;
and step B214, after the random forest of the current layer is updated, updating the random forest of the next layer according to the mode from the step B201 to the step B204 until the structure updating of the whole depth forest is completed.
In some preferred embodiments, the step B213 "select the set sampling interval" includes:
Wherein the content of the first and second substances,input sample x for ith view(i)The value at the dimension delta is taken from,as a lower boundary of the sample spaceThe value at dimension δ.
In some preferred embodiments, step B20 "update the weights of different views" is performed by:
step B221, after the updating of the random forest structure of a certain layer is completed, the prediction labels corresponding to the random forests of all the viewing angles are respectively calculated
Wherein f isiA prediction vector for view i;
step B222, respectively corresponding the predicted labels of the I visual anglesComparing with the data real label y, and updating the output weight of each visual angle:
step B223, normalizing each view angle weight:
wherein I is the number of viewing angles.
In some preferred embodiments, the loss function is:
On the other hand, the invention provides an online multi-view classification system based on a deep forest, which comprises an input module, a multi-view classification module, an arrangement module, a coloring module and an output module;
the input module is configured to acquire and input multi-view data which are characteristics of the same target with different attributes or data of different sensors;
the multi-view classification module is configured to obtain a classification result graph of the input multi-view data through a multi-view depth forest network based on the input multi-view data;
the arrangement module is configured to arrange the prediction labels of the classification result graph into the same size as the input multi-view data picture in a column;
the coloring module is configured to color the arranged classification result graph into a gray classification result graph according to the size of the label;
the output module is configured to output the acquired gray level classification result image.
In a third aspect of the present invention, a storage device is proposed, in which a plurality of programs are stored, the programs being adapted to be loaded and executed by a processor to implement the above-mentioned method for online multi-view classification based on deep forests.
In a fourth aspect of the present invention, a processing apparatus is provided, which includes a processor, a storage device; the processor is suitable for executing various programs; the storage device is suitable for storing a plurality of programs; the program is adapted to be loaded and executed by a processor to implement the above-described depth forest based online multi-view classification method.
The invention has the beneficial effects that:
(1) the deep forest based on the online multi-view classification method learns a deep forest model from a data stream in an online mode, and updates the structure of a random forest and the weights of different views according to new data, so that the model has strong adaptability and expansibility on dynamic change data sets of different scales.
(2) The online multi-view classification method based on the deep forest fully utilizes the multi-view data and the information among the multi-views, and simultaneously uses the multilayer random forest structure, so that deep characteristic information can be effectively developed, and higher online classification accuracy is obtained.
(3) The online multi-view classification method based on the deep forest does not need to store all data, and can effectively save computing resources.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings in which:
FIG. 1 is a schematic flow chart of an online multi-view classification method based on deep forests according to the invention;
FIG. 2 is a schematic diagram of a fully-polarized synthetic aperture radar image and corresponding superpixel segmentation map and real object class label map according to an embodiment of the online multi-view classification method based on depth forest;
FIG. 3 is a comparison graph of classification results of an embodiment of the online multi-view classification method based on deep forests according to the present invention and the prior art.
Detailed Description
The present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present application will be described in detail below with reference to the embodiments with reference to the attached drawings.
The invention discloses an online multi-view classification method based on a deep forest, which comprises the following steps:
step S10, acquiring the characteristics of the same target with different attributes or the data of different sensors as input multi-view data;
step S20, based on the input multi-view data, obtaining a classification result graph of the input multi-view data through a multi-view depth forest network; the multi-view depth forest network comprises a first preset number of layers, and each layer comprises a second preset number of random forests;
step S30, arranging the prediction labels of the classification result picture in a column with the same size as the input multi-view data picture;
in step S40, the arranged classification result maps are colored into gray classification result maps according to the size of the label.
In order to more clearly describe the online multi-view classification method based on deep forests, the following describes the steps in the embodiment of the method in detail with reference to fig. 1.
The online multi-view classification method based on the deep forest comprises the steps of S10-S40, and the steps are described in detail as follows:
step S10, acquiring features of different attributes of the same object or data of different sensors as input multi-view data.
In the big data era, the data amount is increased, the representation forms of the data are more and more diversified, and the data of the same object in different representation forms are multi-view data. Multi-perspective is understood to mean a plurality of data sources (e.g. different sensor data of an object, reports from different institutions of the same event, etc.), a plurality of representations (e.g. a picture can be represented in pixels, can be represented in features, can be interpreted in different languages in the same article, etc.), a plurality of modalities (e.g. a person's data can be represented as images, sounds, words, etc., and a photograph and a tag, comment, etc. attached to the photograph). Each source or type of multi-view data may be complementary, e.g., an untagged picture may be tagged by the picture; it may also be redundant, such as reports from different entities of the same event being largely different in emphasis and style.
In the embodiment of the invention, the acquired multi-view data is polarized SAR covariance data and is subjected to super-pixel segmentation. The method is characterized in that a self-adaptive local iterative clustering method designed for polarized SAR data is adopted for clustering, firstly, a Sphere Invariant Random Vector (SIRV) product model is used for normalizing a covariance matrix, and then an edge graph is calculated; setting the number of super pixels and initializing a clustering center; and then carrying out local iterative clustering on the self-adaptive distance designed based on polarization, texture and spatial information, and finally eliminating unconnected pixels.
Extracting polarization features from each superpixelColor characteristicsTexture featuresAveraging the data in the super pixels to serve as three visual angle data; the polarization characteristics are 45-dimensional and comprise original characteristics extracted from polarization SAR data and characteristics based on polarization decomposition; 34 dimensions of color features including grayscale image elements, dominant color weights, HSV images and histograms thereof; and the 86-dimensional texture features comprise local binary pattern histograms, gray level co-occurrence matrixes, Gabor and wavelet transform coefficients.
Step S20, based on the input multi-view data, obtaining a classification result graph of the input multi-view data through a multi-view depth forest network; the multi-view depth forest network comprises a first preset number of layers, and each layer comprises a second preset number of random forests.
The training method of the multi-view depth forest network comprises the following steps:
and selecting the layer number L of the multi-view depth forest network in a cross validation mode. In the embodiment of the present invention, the number of random trees in each random forest is m-100, and the minimum number of samples required for splitting the tree is min _ samples _ split-10.
And B10, initializing weights of different visual angles of the multi-visual angle depth forest network, and obtaining the initialized multi-visual angle depth forest network.
The method for initializing the weights among different visual angles of the multi-visual angle depth forest network comprises the following steps:
in the embodiment of the invention, the multi-view data comprises three views, and the weight among the three views is beta1、β2、β3The weights between different viewing angles satisfy the relationship shown in equation (1):
the online training multi-view depth forest network is of an L-layer structure, each layer is provided with 3 random forests, data of 3 different views are processed respectively, each forest adopts a random forest based on a Mondrian process, the structure of each forest is continuously updated along with data flow in the online training process, and the weight between views of each layer is continuously updated along with the prediction accuracy of the random forest of each view on the data.
And B20, updating the multi-view depth forest network structure at the t moment based on the multi-view depth forest network at the t-1 moment, the multi-view data at the t moment and the corresponding real label, acquiring a prediction label of the multi-view data at the t moment, calculating a loss function of the comparison between the prediction label and the real label, and updating the weights of different views.
There are three online updating methods for the random forest structure: firstly, introducing a new split "on top" of the existing split; secondly, expanding the existing split; thirdly, splitting an existing leaf node into two child nodes.
The method for updating the structure of the multi-view depth forest network comprises the following steps:
step B211, the structure of the current tree in the current layer is shown in formula (2):
t ═ T (T, δ, ξ, τ) formula (2)
Where T denotes the current tree, δ denotes the dimension in which node splitting of the tree occurs, ξ denotes the location in which node splitting of the tree occurs, and τ denotes the time at which node splitting of the tree occurs.
Step B212, for the input sample D of the ith view angle(i)=(x(i),y),el、euAs shown in formulas (3) and (4):
wherein the content of the first and second substances,are input samples x respectively(i)The lower and upper bounds of the element-by-element calculation of the space in which the device is located; e.g. of the typel、euAre input samples x respectively(i)And j is the jth node of the tree when the difference falls outside the given interval and is the difference between the lower bound and the upper bound.
Step B213, from the parametersThe method comprises the following steps of sampling in exponential distribution, obtaining splitting time E which needs to be increased when nodes of the tree are split, and updating as follows:
if τparen(t)j+E<τjThen the sampling yields the probability and of the splitting dimension δ, δ ═ dIn direct proportion, a set sampling interval is selected for sampling to obtain a splitting position, a new father node is inserted above the current node, and a new leaf node is generated below the father node; if τparent(j)+E≥τjUpdating the upper and lower bounds of the input space, and judging whether the current node j is a leaf node, if so, stopping updating the random forest of the current layer, and if not, continuing to iterate downwards.
Wherein, taujIs the split time of internal node j, parent (j) is the parent of node j.
The method for selecting and setting the sampling interval comprises the following steps:
Wherein the content of the first and second substances,input sample x for ith view(i)The value at the dimension delta is taken from,as a lower boundary of the sample spaceThe value at dimension δ.
And step B214, after the random forest of the current layer is updated, updating the random forest of the next layer according to the mode from the step B201 to the step B204 until the structure updating of the whole multi-view depth forest network is completed.
The method for updating the weights of different views comprises the following steps:
step B221, after the updating of the random forest structure of a certain layer is completed, the prediction labels corresponding to the random forests of all the viewing angles are respectively calculatedAs shown in formula (5):
wherein f isiA prediction vector for view i;
step B222, respectively corresponding the predicted labels of the I visual anglesComparing with the data real label y and updating each viewThe output weight of the corner is shown in equation (6):
step B223, normalizing the weight of each view angle, as shown in formula (7):
βi'=βi/(β1+β2+β3) Formula (7)
After normalization, the weight β1'+β2'+β3'=1。
Each layer in the depth forest updates the weight, and the final output vector f is the weight of the output result of the last three layers of the views, as shown in formula (8):
the loss function is shown in equation (9):
And step B30, repeating step B20 until a preset training time is reached or the loss function value of the prediction tag is lower than a set threshold value, wherein t is t + 1.
Step S30, arranging the prediction labels of the classification result picture in a column with the same size as the input multi-view data picture.
And step S40, coloring the arranged classification result graph into a gray classification result graph according to the size of the label.
In the embodiment of the invention, real polarized SAR data is used for a test experiment, and the data is L-band four-view full-polarization data of the San Francisco (San Francisco) region in the United states, which is acquired by an AIRSAR sensor. As shown in fig. 2, which is a schematic diagram of a fully-polarized synthetic aperture radar image and corresponding superpixel segmentation maps and real object class label maps according to an embodiment of the online multi-view classification method based on a deep forest of the present invention, the left diagram of fig. 2 is a Pauli decomposition gray scale image of the data, the size of which is 900 × 1024, the diagram of fig. 2 is the superpixel segmentation map, the right diagram of fig. 2 is a corresponding real object class map, the region includes 4 classes of ground objects, and the object types represented by the regions with different gray scales are given below fig. 2.
In the simulation experiment of the embodiment of the invention, software is used: python 2.7, processor: intel (R) core (TM) i7-6700HQ, memory: 16.0GB, operating System: 64-bit Windows 10.
Analysis of experimental contents and results:
in order to evaluate the effect of the on-line multi-view classification method based on the deep forest, the method for carrying out experimental comparison comprises the following steps: prior art Online Multi-view learning algorithms (OMPA), and prior art Online Mondrian Forest algorithms (OMF). The parameters of these methods were selected by cross-validation, with the optimal parameters as follows: the penalty parameter c of OMPA is 0.1, and the balance parameter lambda1=λ2=1,λ31.5, 1e-4, a coupling parameter d, and a weighting parameter r1=r2=0.3,r30.4. The number of random trees in the OMF parameter is 100 and the minimum sample split is set to 2. The number of network layers is 3, the number of random trees in each forest is 100, and the minimum sample split is set to be 10.
As shown in fig. 3, which is a comparison graph of classification results of an embodiment of the online multi-view classification method based on deep forest according to the present invention and the prior art, from left to right are graphs of classification results using the OMPA, the OMF and the method of the present invention, respectively. As shown in Table 1, the classification error rate comparison results of these methods on the test set are given. As can be seen from fig. 3 and table 1, the classification result of the method of the present invention is significantly better than the results of OMPA and OMF, the classification accuracy of the four subclasses is higher than that of OMPA and OMF, and the total accuracy is improved by nearly 2% compared with OMF, indicating that the method of the present invention can realize a classification result with higher accuracy.
TABLE 1
Precision of classification (%) | Building construction | Grass land | Oceans | Mountain range | Total accuracy |
OMPA | 95.54 | 82.26 | 74.16 | 93.20 | 90.13 |
OMF | 98.03 | 73.58 | 78.41 | 95.72 | 92.30 |
The method of the invention | 98.12 | 89.98 | 78.97 | 96.25 | 94.02 |
The online multi-view classification system based on the deep forest comprises an input module, a multi-view classification module, an arrangement module, a coloring module and an output module, wherein the input module is used for inputting a plurality of images;
the input module is configured to acquire and input multi-view data which are characteristics of the same target with different attributes or data of different sensors;
the multi-view classification module is configured to obtain a classification result graph of the input multi-view data through a multi-view depth forest network based on the input multi-view data;
the arrangement module is configured to arrange the prediction labels of the classification result graph into the same size as the input multi-view data picture in a column;
the coloring module is configured to color the arranged classification result graph into a gray classification result graph according to the size of the label;
the output module is configured to output the acquired gray level classification result image.
It can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working process and related description of the system described above may refer to the corresponding process in the foregoing method embodiments, and will not be described herein again.
It should be noted that, the online multi-view classification system based on deep forest provided in the above embodiment is only illustrated by the division of the above functional modules, and in practical applications, the above functions may be allocated to different functional modules according to needs, that is, the modules or steps in the embodiment of the present invention are further decomposed or combined, for example, the modules in the above embodiment may be combined into one module, or may be further split into multiple sub-modules, so as to complete all or part of the above described functions. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not to be construed as unduly limiting the present invention.
A storage device according to a third embodiment of the invention has stored therein a plurality of programs adapted to be loaded and executed by a processor to implement the above-described method for online multi-perspective classification based on deep forests.
A processing apparatus according to a fourth embodiment of the present invention includes a processor, a storage device; a processor adapted to execute various programs; a storage device adapted to store a plurality of programs; the program is adapted to be loaded and executed by a processor to implement the above-described depth forest based online multi-view classification method.
It can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working processes and related descriptions of the storage device and the processing device described above may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.
Those of skill in the art would appreciate that the various illustrative modules, method steps, and modules described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that programs corresponding to the software modules, method steps may be located in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. To clearly illustrate this interchangeability of electronic hardware and software, various illustrative components and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as electronic hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The terms "first," "second," and the like are used for distinguishing between similar elements and not necessarily for describing or implying a particular order or sequence.
The terms "comprises," "comprising," or any other similar term are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.
Claims (8)
1. An online multi-view classification method based on a deep forest is characterized by comprising the following steps:
step S10, acquiring the characteristics of the same target with different attributes or the data of different sensors as input multi-view data;
step S20, based on the input multi-view data, obtaining a classification result graph of the input multi-view data through a multi-view depth forest network; the multi-view depth forest network comprises a first preset number of layers, and each layer comprises a second preset number of random forests;
step S30, arranging the prediction labels corresponding to the pixels of the classification result image into the same size as the input multi-view data image according to the corresponding positions of the pixels in the input multi-view data image;
step S40, coloring the arranged classification result graph into a gray classification result graph according to different classes of the prediction labels corresponding to the pixels;
the training method of the multi-view depth forest network comprises the following steps:
step B10, initializing weights of different visual angles of the multi-visual angle depth forest network, and obtaining the initialized multi-visual angle depth forest network;
step B20, updating the multi-view depth forest network structure at the time t on line based on the multi-view depth forest network at the time t-1, the multi-view data at the time t and corresponding real labels:
step B211, the structure of the current tree in the current layer is T ═ (T, δ, ξ, τ), where T denotes the current tree, δ denotes the dimension of the occurrence of node splitting of the tree, ξ denotes the position of the occurrence of node splitting of the tree, τ denotes the time of occurrence of node splitting of the tree;
step B212, for the input sample D of the ith view angle(i)=(x(i)Y), settingWherein the content of the first and second substances,are input samples x respectively(i)The lower and upper bounds of the element-by-element calculation of the space in which the device is located; e.g. of the typel、euAre input samples x respectively(i)The difference between the current node and the lower and upper bounds when the current node falls outside the given interval, wherein j is the jth node of the tree;
step B213, from the parametersThe method comprises the following steps of sampling in exponential distribution, obtaining splitting time E which needs to be increased when nodes of the tree are split, and updating as follows: if τparent(j)+E<τjThen the sampling yields the probability and of the splitting dimension δ, δ ═ dIn direct proportion, a set sampling interval is selected for sampling to obtain a splitting position, a new father node is inserted above the current node, and a new leaf node is generated below the father node; if τparent(j)+E≥τjUpdating the upper and lower bounds of the input space,judging whether the current node j is a leaf node, if so, stopping updating the random forest of the current layer, and if not, continuing to iterate downwards; wherein, taujIs the split time of internal node j, parent (j) is the parent of node j;
step B214, after the random forest of the current layer is updated, updating the random forest of the next layer according to the modes from step B211 to step B213 until the structure updating of the whole depth forest is completed;
acquiring a prediction label of the multi-view data at the time t, calculating a loss function of the comparison between the prediction label and a real label, and updating weights of different views;
and step B30, repeating step B20 until a preset training time is reached or the loss function value of the prediction tag is lower than a set threshold value, wherein t is t + 1.
2. The method for online multi-view classification based on depth forest according to claim 1, wherein step B10 "initialize weights between different views of the multi-view depth forest network" by:
3. the method for on-line multi-view classification based on deep forest as claimed in claim 1, wherein in step B213 "select set sampling interval" is performed by:
4. The online multi-view classification method based on deep forest as claimed in claim 1, wherein in step B20 "update weights of different views" the method is:
step B221, after the updating of the random forest structure of a certain layer is completed, the prediction labels corresponding to the random forests of all the viewing angles are respectively calculated
Wherein f isiA prediction vector for view i;
step B222, respectively corresponding the predicted labels of the I visual anglesComparing with the data real label y, and updating the output weight of each visual angle:
step B223, normalizing each view angle weight:
wherein I is the number of viewing angles.
6. An online multi-view classification system based on a deep forest is characterized by comprising an input module, an online multi-view classification module, an arrangement module, a coloring module and an output module;
the input module is configured to acquire and input multi-view data which are characteristics of the same target with different attributes or data of different sensors;
the online multi-view classification module is configured to obtain a classification result graph of the input multi-view data through a multi-view depth forest network based on the input multi-view data;
the arrangement module is configured to arrange the prediction labels corresponding to the pixels of the classification result image into the same size as the input multi-view data image according to the corresponding positions of the pixels in the multi-view data image;
the coloring module is configured to color the arranged classification result images into gray classification result images according to different classes of the prediction labels corresponding to the pixels;
the output module is configured to output the obtained gray level classification result graph;
the training method of the multi-view depth forest network comprises the following steps:
step B10, initializing weights of different visual angles of the multi-visual angle depth forest network, and obtaining the initialized multi-visual angle depth forest network;
step B20, updating the t-moment multi-view depth forest network structure based on the t-1-moment multi-view depth forest network, the t-moment multi-view data and the corresponding real labels:
step B211, the structure of the current tree in the current layer is T ═ (T, δ, ξ, τ), where T denotes the current tree, δ denotes the dimension of the occurrence of node splitting of the tree, ξ denotes the position of the occurrence of node splitting of the tree, τ denotes the time of occurrence of node splitting of the tree;
step B212, for the input sample D of the ith view angle(i)=(x(i)Y), settingWherein the content of the first and second substances,are input samples x respectively(i)The lower and upper bounds of the element-by-element calculation of the space in which the device is located; e.g. of the typel、euAre input samples x respectively(i)The difference between the current node and the lower and upper bounds when the current node falls outside the given interval, wherein j is the jth node of the tree;
step B213, from the parametersThe method comprises the following steps of sampling in exponential distribution, obtaining splitting time E which needs to be increased when nodes of the tree are split, and updating as follows: if τparent(j)+E<τjThen the sampling yields the probability and of the splitting dimension δ, δ ═ dIn direct proportion, a set sampling interval is selected for sampling to obtain a splitting position, a new father node is inserted above the current node, and a new leaf node is generated below the father node; if τparent(j)+E≥τjUpdating the upper and lower bounds of the input space,judging whether the current node j is a leaf node, if so, stopping updating the random forest of the current layer, and if not, continuing to iterate downwards; wherein, taujIs the split time of internal node j, parent (j) is the parent of node j;
step B214, after the random forest of the current layer is updated, updating the random forest of the next layer according to the modes from step B211 to step B213 until the structure updating of the whole depth forest is completed;
acquiring a prediction label of the multi-view data at the time t, calculating a loss function of the comparison between the prediction label and a real label, and updating weights of different views;
and step B30, repeating step B20 until a preset training time is reached or the loss function value of the prediction tag is lower than a set threshold value, wherein t is t + 1.
7. A storage device having a plurality of programs stored therein, wherein the programs are adapted to be loaded and executed by a processor to implement the method for deep forest based online multi-view classification of any one of claims 1 to 5.
8. A treatment apparatus comprises
A processor adapted to execute various programs; and
a storage device adapted to store a plurality of programs;
wherein the program is adapted to be loaded and executed by a processor to perform:
the method for on-line multi-view classification based on depth forest as claimed in any one of claims 1-5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910502702.2A CN110222778B (en) | 2019-06-11 | 2019-06-11 | Online multi-view classification method, system and device based on deep forest |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910502702.2A CN110222778B (en) | 2019-06-11 | 2019-06-11 | Online multi-view classification method, system and device based on deep forest |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110222778A CN110222778A (en) | 2019-09-10 |
CN110222778B true CN110222778B (en) | 2021-04-13 |
Family
ID=67816562
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910502702.2A Active CN110222778B (en) | 2019-06-11 | 2019-06-11 | Online multi-view classification method, system and device based on deep forest |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110222778B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113378941B (en) * | 2021-06-16 | 2022-05-03 | 中国石油大学(华东) | Multi-decision fusion small sample image classification method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108205805A (en) * | 2016-12-20 | 2018-06-26 | 北京大学 | The dense corresponding auto-creating method of voxel between pyramidal CT image |
CN108388907A (en) * | 2017-12-29 | 2018-08-10 | 中国科学院自动化研究所 | Polarization SAR data sorter real time updating method based on various visual angles study |
US10311341B1 (en) * | 2015-08-27 | 2019-06-04 | Hrl Laboratories, Llc | System and method for online deep learning in an ultra-low power consumption state |
-
2019
- 2019-06-11 CN CN201910502702.2A patent/CN110222778B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10311341B1 (en) * | 2015-08-27 | 2019-06-04 | Hrl Laboratories, Llc | System and method for online deep learning in an ultra-low power consumption state |
CN108205805A (en) * | 2016-12-20 | 2018-06-26 | 北京大学 | The dense corresponding auto-creating method of voxel between pyramidal CT image |
CN108388907A (en) * | 2017-12-29 | 2018-08-10 | 中国科学院自动化研究所 | Polarization SAR data sorter real time updating method based on various visual angles study |
Non-Patent Citations (2)
Title |
---|
Balaji Lakshminarayanan etal..Mondrian Forests: Efficient Online Random Forests.《www.arXiv:1406.2673v2》.2015, * |
基于深度森林模型的火焰检测;朱晓好等;《计算机工程》;20180715;第44卷(第07期);第2.1节第2段和图6 * |
Also Published As
Publication number | Publication date |
---|---|
CN110222778A (en) | 2019-09-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Cheng et al. | Pest identification via deep residual learning in complex background | |
Hu et al. | Fast forest fire smoke detection using MVMNet | |
Li et al. | Fast and accurate green pepper detection in complex backgrounds via an improved Yolov4-tiny model | |
Byeon et al. | Scene labeling with lstm recurrent neural networks | |
CN107066559B (en) | Three-dimensional model retrieval method based on deep learning | |
CN112639828A (en) | Data processing method, method and equipment for training neural network model | |
CN113627472B (en) | Intelligent garden leaf feeding pest identification method based on layered deep learning model | |
CN110222718B (en) | Image processing method and device | |
Feng et al. | Bag of visual words model with deep spatial features for geographical scene classification | |
Palomo et al. | Learning topologies with the growing neural forest | |
Xu et al. | Weakly supervised deep semantic segmentation using CNN and ELM with semantic candidate regions | |
CN109840518B (en) | Visual tracking method combining classification and domain adaptation | |
CN110889865A (en) | Video target tracking method based on local weighted sparse feature selection | |
CN112560966A (en) | Polarimetric SAR image classification method, medium and equipment based on scattergram convolution network | |
Wang et al. | Joint-learning segmentation in Internet of drones (IoD)-based monitor systems | |
Chodey et al. | Hybrid deep learning model for in-field pest detection on real-time field monitoring | |
Niu | Semantic segmentation of remote sensing image based on convolutional neural network and mask generation | |
Bowley et al. | An analysis of altitude, citizen science and a convolutional neural network feedback loop on object detection in unmanned aerial systems | |
CN110222778B (en) | Online multi-view classification method, system and device based on deep forest | |
CN112418256A (en) | Classification, model training and information searching method, system and equipment | |
CN112668662A (en) | Outdoor mountain forest environment target detection method based on improved YOLOv3 network | |
Rashno et al. | Mars image segmentation with most relevant features among wavelet and color features | |
CN113963178A (en) | Method, device, equipment and medium for detecting infrared dim and small target under ground-air background | |
Huo et al. | Local graph regularized coding for salient object detection | |
Song et al. | CNN based Wildlife Recognition with Super-pixel Segmentation for Ecological Surveillance |
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 |