CN109815967A - CNN ship seakeeping system and method based on Fusion Features - Google Patents
CNN ship seakeeping system and method based on Fusion Features Download PDFInfo
- Publication number
- CN109815967A CN109815967A CN201910148840.5A CN201910148840A CN109815967A CN 109815967 A CN109815967 A CN 109815967A CN 201910148840 A CN201910148840 A CN 201910148840A CN 109815967 A CN109815967 A CN 109815967A
- Authority
- CN
- China
- Prior art keywords
- feature
- cnn
- fusion
- features
- classification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of, and the CNN ship seakeeping system and method based on Fusion Features is used to carry out multilayer convolution to input naval vessel infrared image and pondization operates the system comprises CNN module is improved, automatic acquisition image high dimensional feature information；Improve CNN module and be based on LeNet-5 convolutional neural networks and be changed acquisitions: including use convolution kernel, the convolution number of plies of 3*3 for 6, for pond mode using maximum value pond mode, excitation function is Relu function；Fusion Features module for manually extracting the manual features information of the input naval vessel infrared image, and then obtains weighted feature sequence；Fused in tandem is carried out with based on the image high dimensional feature information for improving the acquisition of CNN module again, obtains fusion feature sequence；Classification output module realizes target identification for classifying using full articulamentum to fusion feature sequence.The present invention can automatically extract target signature, have good generalization ability and accuracy of identification；It makes characteristic of division have certain interpretation simultaneously.
Description
Technical field
The present invention relates to ship seakeeping technical field more particularly to a kind of CNN Ship Targets based on Fusion Features
Identifying system and method.
Background technique
Target Recognition Algorithms based on convolutional neural networks (Convolutional Neural Networks, CNN) are mesh
Preceding relatively advanced target identification method.With traditional algorithm such as SVM, K- nearest neighbor clustering algorithm etc. is compared, the target based on CNN
Recognizer has been abandoned the step of artificially carrying out feature extraction to target in traditional algorithm, automatic by profound convolution operation
It extracts the high dimensional feature of target image and classifies to it.CNN has preferably extensive compared with tional identification algorithm simultaneously
Ability and higher accuracy of identification.
Infrared imaging has many advantages, such as that high resolution, good concealment, climate adaptability are strong compared to other imaging modes.Cause
This is based on the advantages of infrared imaging, this research field of the target identification of naval vessel infrared image has extensive purposes, not only
It can be used for round-the-clock sea allocation of ships, ship rescue at sea searches aspect and also plays an important role, significant.
Currently, Infrared Targets image-recognizing method can be divided mainly into two major classes according to the difference of feature extraction mode: with people
Work extracts the conventional method such as support vector machines, clustering algorithm etc. that feature is basis of classification；And based on the automatic of deep learning
Target's feature-extraction classification method such as CNN etc..By taking support vector machines as an example, traditional sorting algorithm is dependent on people for target
Subjective experience manually extracts the statistical natures such as the structure of target, texture by the subjective understanding of people, and using data set to system
It is trained, finds the hyperplane that can distinguish two class target signatures, realize the purpose of classification.As used in tional identification algorithm
Feature joined artificial subjective understanding, have good controllability and interpretation, but cannot achieve the feature of automation
It extracts；On the other hand due to that can not extract target more high dimensional feature, traditional algorithm accuracy of identification is lower, and generalization ability is poor.Base
Target high dimensional feature can automatically be extracted by being handled by the convolution operation of deep layer image in the method for deep learning
And classify, realize full-automation of the algorithm from feature extraction to classification；Sorting algorithm based on deep learning has good
Good robustness and generalization ability, but its extracted high dimensional feature does not have interpretation, and for the warship of small data set
For ship infrared image, training set negligible amounts can also generate that discrimination is low and over-fitting.
Therefore, against the above deficiency, it is desirable to provide a kind of new ship seakeeping technology, by tional identification algorithm and depth
Degree learning algorithm is combined for ship seakeeping, so that target signature can either be automatically extracted, and has characteristic of division
Certain interpretation.
Summary of the invention
The technical problem to be solved in the present invention is that for the identification of ship target infrared image, adopting in the prior art
It cannot achieve the feature extraction of automation with traditional classification algorithm, and the extracted higher-dimension of the sorting algorithm based on deep learning is special
The defect without interpretation is levied, a kind of CNN ship seakeeping system and method based on Fusion Features is provided.
In order to solve the above-mentioned technical problems, the present invention provides a kind of CNN ship seakeeping system based on Fusion Features
System, comprising:
CNN module is improved, for carrying out multilayer convolution and pondization operation to input naval vessel infrared image, obtains image automatically
High dimensional feature information；The improvement CNN module is based on LeNet-5 convolutional neural networks and is changed acquisition: including using 3*3's
Convolution kernel, the convolution number of plies are 6, and pond mode uses maximum value pond mode, and excitation function is Relu function；
Fusion Features module, for manually extracting the manual features information of the input naval vessel infrared image, to artificial spy
Reference breath is handled, and is carried out linear weighted function and obtained weighted feature sequence；The weighted feature sequence with based on improve CNN mould
The image high dimensional feature information that block obtains carries out fused in tandem, obtains fusion feature sequence；
Classification output module realizes target identification for classifying using full articulamentum to fusion feature sequence.
In the CNN ship seakeeping system according to the present invention based on Fusion Features, the artificial extraction institute
The manual features information for stating input naval vessel infrared image includes extracting the HOG feature and SIFT feature of input naval vessel infrared image.
It is described to manual features in the CNN ship seakeeping system according to the present invention based on Fusion Features
Information carries out processing
Main feature is refined using PCA Principal Component Analysis to the HOG feature and SIFT feature, to the main feature
K- neighbour Cluster Classification is carried out, multiple tagsort groups are obtained；
The method for obtaining weighted feature sequence includes,
The error in classification for analyzing main feature described in each tagsort group, according to error in classification to each tagsort
Main feature in group carries out linear weighted function, obtains the weighted feature sequence of all main features corresponding to tagsort group.
In the CNN ship seakeeping system according to the present invention based on Fusion Features, the progress linearly adds
The weight w of poweriAre as follows:
In formula, i is characterized the ordinal number of sorting group, and M is characterized the total number of sorting group, eiFor error in classification.
In the CNN ship seakeeping system according to the present invention based on Fusion Features, the full articulamentum draws
Enter Dropout algorithm to classify to fusion feature sequence.
The CNN ship seakeeping method based on Fusion Features that the present invention also provides a kind of, comprising:
Multilayer convolution and pondization operation are carried out to input naval vessel infrared image using CNN module is improved, obtain image automatically
The step of high dimensional feature information；The improvement CNN module is based on LeNet-5 convolutional neural networks and is changed acquisition: including adopting
It is 6 with the convolution kernel of 3*3, the convolution number of plies, pond mode uses maximum value pond mode, and excitation function is Relu function；
The manual features information for manually extracting the input naval vessel infrared image, handles manual features information, and
Carry out the step of linear weighted function obtains weighted feature sequence；And using Fusion Features module to the weighted feature sequence and base
In the step of image high dimensional feature information for improving the acquisition of CNN module carries out fused in tandem, obtains fusion feature sequence；
The step of classifying to fusion feature sequence by the full articulamentum for output module of classifying, realize target identification.
In the CNN ship seakeeping method according to the present invention based on Fusion Features, the artificial extraction institute
The manual features information for stating input naval vessel infrared image includes extracting the HOG feature and SIFT feature of input naval vessel infrared image.
It is described to manual features in the CNN ship seakeeping method according to the present invention based on Fusion Features
Information carries out processing
Main feature is refined using PCA Principal Component Analysis to the HOG feature and SIFT feature, to the main feature
K- neighbour Cluster Classification is carried out, multiple tagsort groups are obtained；
The method for obtaining weighted feature sequence includes,
The error in classification for analyzing main feature described in each tagsort group, according to error in classification to each tagsort
Main feature in group carries out linear weighted function, obtains the weighted feature sequence of all main features corresponding to tagsort group.
In the CNN ship seakeeping method according to the present invention based on Fusion Features, the progress linearly adds
The weight w of poweriAre as follows:
In formula, i is characterized the ordinal number of sorting group, and M is characterized the total number of sorting group, eiFor error in classification.
In the CNN ship seakeeping method according to the present invention based on Fusion Features, the full articulamentum draws
Enter Dropout algorithm to classify to fusion feature sequence.
Implement the CNN ship seakeeping system and method for the invention based on Fusion Features, has the advantages that
The present invention is directed to this kind of small data set target of naval vessel infrared image, and tional identification algorithm and deep learning algorithm advantage are mutually melted
It closes, input naval vessel infrared image is handled and identified.Present system and method can automatically extract target signature, have
Good generalization ability and accuracy of identification；It makes characteristic of division have certain interpretation simultaneously, is the identification of naval vessel infrared image
One significant breakthrough in field.
Improvement CNN module in the present invention can effectively reduce the training parameter of convolutional neural networks, greatly shorten training
Time；It uses maximum value pond mode to reduce the complexity of network, improves training effectiveness；Used excitation function
Help to improve the convergence rate of system.
Detailed description of the invention
Fig. 1 is the exemplary block diagram of the CNN ship seakeeping system according to the present invention based on Fusion Features；
Fig. 2 is the exemplary block diagram according to the present invention for improving CNN module；
Fig. 3 is the exemplary algorithm block diagram of Fusion Features module according to the present invention；
Fig. 4 is the exemplary block diagram of the CNN ship seakeeping method according to the present invention based on Fusion Features.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiments of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill people
Member's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
Specific embodiment one, the first aspect of the present invention provide a kind of CNN Ship Target knowledge based on Fusion Features
Other system, in conjunction with shown in Fig. 1 to Fig. 3, comprising:
CNN module is improved, for carrying out multilayer convolution and pondization operation to input naval vessel infrared image, obtains image automatically
High dimensional feature information；The improvement CNN module is based on LeNet-5 convolutional neural networks and is changed acquisition: including using 3*3's
Convolution kernel, the convolution number of plies are 6, and pond mode uses maximum value pond mode, and excitation function is Relu function；
Fusion Features module, for manually extracting the manual features information of the input naval vessel infrared image, to artificial spy
Reference breath is handled, and is carried out linear weighted function and obtained weighted feature sequence；The weighted feature sequence with based on improve CNN mould
The image high dimensional feature information that block obtains carries out fused in tandem, obtains fusion feature sequence；
Classification output module realizes target identification for classifying using full articulamentum to fusion feature sequence.
Present embodiment is by three modules, to realize that the identification to different Ship Targets is classified.The naval vessel can wrap
Include warship, merchant ship, freighter etc..In order to separate ship different classes of under IR Scene, traditional machine learning algorithm is usual
Using the skeleton or contour feature of ship as classification foundation, this method needs artificial extraction feature, and process is cumbersome, and by
In the textural characteristics for having ignored different classes of ship, cause classification accuracy low.Present invention employs improved LeNet-5 convolution
Neural network automatically extracts its higher-dimension textural characteristics to naval vessel infrared image, then blends with the feature manually extracted, and improves
The degree of automation of Ship Recognition and the accuracy rate for increasing classification.
It is described improve CNN module based on classical LeNet-5 convolutional neural networks carry out structure and parameter improvement and
It obtains, relates generally to convolutional layer and pond layer, system parameter table is as follows:
| Parameter | Improve CNN module |
| Convolution kernel | 3*3 |
| The convolution number of plies | 6 |
| Excitation function | ReLu |
| Pond mode | Maximum value |
In conjunction with shown in upper table and Fig. 1 and Fig. 2, in improving CNN module, in order to reduce training parameter, network system is reduced
Complexity improves its stability, has selected smaller 3*3 convolution kernel to replace original 5*5 convolution kernel, by 3 layers of former LeNet-5
Convolution operation deepens to be 6 layers；The method for deepening the convolution number of plies while by reducing convolution kernel size, by LeNet-5 network institute
It needs trained parameter to be reduced to 98 by original 156, effectively reduces the training parameter of convolutional neural networks, significantly
Shorten the training time.
Present embodiment carries out convolution operation to input naval vessel infrared image using the template of 3*3, carries out 6 layers of convolution altogether
Operation.After every layer of convolution operation, maximum value pondization is carried out to gained characteristic pattern using 3*3 template and is operated, the pondization operation
The high dimensional feature information of image can be obtained automatically, and prevents the loss of target information well.Excitation function uses ReLu letter
Number, can prevent the gradient saturation effect generated in training process.
Since the infrared picture data of Ship Target first passes through dividing processing in advance, only retain target area.Exist in order to prevent
Target signature information is lost while pond, by mean value pond mode maximum value in former LeNet-5 network in present embodiment
Pond mode replaces, as shown in Figure 2.Maximum value pond carries out every layer of characteristic pattern by way of choosing template area maximum value
Dimension-reduction treatment reduces the complexity of network, improves training effectiveness.G1 in Fig. 1 and Fig. 2 indicates first convolutional layer,
G2 indicates second convolutional layer, and V1 indicates that first pond layer, G2 indicate second pond layer, up to the 6th convolutional layer and the
Six pond layers.
In order to further speed up the rate of convergence of network system, prevent network system from occurring gradient during training full
And the phenomenon that, the Sigmoid function in LeNet-5 network is replaced in present embodiment with Relu function.Relu function is one
Piecewise function, the gradient of Relu function is 1 in x >=0 compared with Sigmoid function, otherwise is 0.Disappear completely in the part x >=0
In addition to the gradient saturation effect of Sigmoid function；On computation complexity, Relu function is also relatively easy.Relu function simultaneously
The convergence rate of system is helped to improve, convergence rate is relative to about fast 6 times or so of Sigmoid function.
Simoid function expression are as follows:
σ (x) indicates the neuron output response of Simoid function in formula, and x indicates the input of neuron；
Relu function expression are as follows:
ReLu (x) indicates the neuron output response of Relu function in formula.
The Fusion Features module obtains weighted feature sequence by manually extracting characteristics of image and carrying out linear weighted function to it
Column finally carry out fused in tandem with the mentioned feature of improvement CNN module；This is not only able to improve the accuracy of identification to image, also exists
Add somewhat to the interpretation of characteristic of division.Finally artificial extraction feature is merged with the feature obtained automatically
Fusion feature sequence is obtained as characteristic of division.
For present embodiment after improving CNN resume module, the image high dimensional feature of acquisition is the interpretable of image format
Spend very low abstract characteristics；The characteristic information that Fusion Features module is manually extracted mainly passes through the HOG after PCA principal component analysis
Feature and SIFT feature；HOG feature, SIFT feature are input to full articulamentum together with image high dimensional feature, to reach string
Join the purpose of fusion.
Present embodiment has the function that Automatic Feature Extraction and classification are carried out based on naval vessel infrared signature, has good
Discrimination, generalization ability and a degree of feature interpretation can be ShipTargets early warning detection, rescue at sea etc.
Strong technical support is provided.
As an example, the manual features information for manually extracting the input naval vessel infrared image includes extracting input
The HOG feature and SIFT feature of naval vessel infrared image.HOG feature is the operation acquisition on the local pane location of image, so
It is able to maintain good invariance to the geometry and optical deformation of image；And key can be detected according to SIFT feature in the picture
Point, so SIFT feature can be used as a kind of local feature description's.This has good mutual with the high dimensional feature that CNN is extracted
Benefit property.
Further, as shown in connection with fig. 3, it is described to manual features information carry out processing include:
Main feature is refined using PCA Principal Component Analysis to the HOG feature and SIFT feature, to the main feature
K- neighbour Cluster Classification is carried out, multiple tagsort groups are obtained；
The method for obtaining weighted feature sequence includes,
The error in classification for analyzing main feature described in each tagsort group, according to error in classification to each tagsort
Main feature in group carries out linear weighted function, obtains the weighted feature sequence of all main features corresponding to tagsort group.
In present embodiment, Fusion Features module carries out HOG feature to input naval vessel infrared image and SIFT feature mentions
After taking, dimensionality reduction operation is carried out to it using PCA, only retains main feature, it is described to be mainly characterized by from HOG feature and SIFT
Feature after the dimensionality reduction of feature.
Further, the weight w for carrying out linear weighted functioniAre as follows:
In formula, i is characterized the ordinal number of sorting group, and M is characterized the total number of sorting group, eiFor error in classification.
Error in classification eiIt can be used as weighting foundation, according to error in classification and calculate the weight obtained, it can be to corresponding
Main feature in tagsort group is weighted.
Further, referring to figs. 1 and 2, the full articulamentum introduces Dropout algorithm and carries out to fusion feature sequence
Classification.
In present embodiment, according to actual needs, for the objective factor of naval vessel infrared image negligible amounts, connecting entirely
Dropout algorithm is introduced in layer, be can effectively improve the generalization ability of neural network, is prevented from drawing because training set of images is less
The training system over-fitting risen.Meanwhile the Dropout algorithm of full articulamentum can select during training parameter each time
Property ignores partial nerve member parameter.Finally, be input to fusion feature sequence as characteristic of division in full articulamentum, it can be quick
Precisely obtain the output category of target image.
In present embodiment, system parameter can be carried out using error backpropagation algorithm and stochastic gradient descent method
Training, algorithm parameter initialization use Xavier parameter initialization method.
The present invention is further detailed combined with specific embodiments below:
Embodiment: the ship seakeeping system includes improving CNN module, Fusion Features module and classification output mould
Block, as shown in Figure 1.
In improving CNN module, convolution operation is carried out to input picture using the template of 3*3, each layer of convolution can all obtain
To the target signature more abstract than upper one layer of convolution more higher-dimension, 6 layers of convolution are carried out altogether.After every layer of convolution operation, to gained
Characteristic pattern carries out the operation of maximum value pondization, and maximum value pondization traverses gained characteristic pattern using 3*3 template, chooses in ergodic process
Maximum value is exported as sampled value in template area, finally obtains the input figure by the characteristic pattern of dimensionality reduction as next convolutional layer
Picture.This pond mode can retain to the greatest extent target signature information while to characteristic pattern dimensionality reduction.Wherein motivate letter
Number uses ReLu function, can prevent the gradient saturation effect generated in training process.
In Fusion Features module, manual features extraction is carried out to input picture, PCA principal component is used to extracted feature
Analysis, then target classification is carried out by K- neighbour clustering method with gained feature, the error in classification for analyzing every kind of feature, which is used as, to be added
Power foundation is weighted feature.Gained feature is finally subjected to Fusion Features of connecting with the high dimensional feature that CNN is automatically extracted out.
Fusion Features modular algorithm block diagram is as shown in Figure 3.
Classification output module in introduce Dropout algorithm, by the above-mentioned full articulamentum of gained fusion feature sequence inputting into
Row classification output, while while carrying out parameter training using error backpropagation algorithm, it is selective to partial nerve member
Parameter remains unchanged, and improves the generalization ability of network.Finally obtain the classification output of naval vessel infrared image.
Specific embodiment two, another aspect of the present invention additionally provide a kind of CNN Ship Target based on Fusion Features
Recognition methods, in conjunction with shown in Fig. 1 to Fig. 4, comprising:
Multilayer convolution and pondization operation are carried out to input naval vessel infrared image using CNN module is improved, obtain image automatically
The step of high dimensional feature information；The improvement CNN module is based on LeNet-5 convolutional neural networks and is changed acquisition: including adopting
It is 6 with the convolution kernel of 3*3, the convolution number of plies, pond mode uses maximum value pond mode, and excitation function is Relu function；
The manual features information for manually extracting the input naval vessel infrared image, handles manual features information, and
Carry out the step of linear weighted function obtains weighted feature sequence；And using Fusion Features module to the weighted feature sequence and base
In the step of image high dimensional feature information for improving the acquisition of CNN module carries out fused in tandem, obtains fusion feature sequence；
The step of classifying to fusion feature sequence by the full articulamentum for output module of classifying, realize target identification.
Present embodiment realizes that the identification to different Ship Targets is classified by three steps.The naval vessel may include
Warship, merchant ship, freighter etc..In order to separate ship different classes of under IR Scene, traditional machine learning algorithm is usually adopted
The skeleton or contour feature for using ship are as classification foundation, and this method needs artificial extraction feature, and process is cumbersome, and due to
The textural characteristics for having ignored different classes of ship cause classification accuracy low.Present embodiment uses improved LeNet-5 convolution
Neural network automatically extracts its higher-dimension textural characteristics to naval vessel infrared image, then blends with the feature manually extracted, and improves
The degree of automation of Ship Recognition and the accuracy rate for increasing classification.
Present embodiment carries out convolution operation to input naval vessel infrared image using the template of 3*3, carries out 6 layers of convolution altogether
Operation.After every layer of convolution operation, maximum value pondization is carried out to gained characteristic pattern using 3*3 template and is operated, the pondization operation
The high dimensional feature information of image can be obtained automatically, and prevents the loss of target information well.Excitation function uses ReLu letter
Number, can prevent the gradient saturation effect generated in training process.
Since the infrared picture data of Ship Target first passes through dividing processing in advance, only retain target area.Exist in order to prevent
Target signature information is lost while pond, by mean value pond mode maximum value in former LeNet-5 network in present embodiment
Pond mode replaces, as shown in Figure 2.Maximum value pond carries out every layer of characteristic pattern by way of choosing template area maximum value
Dimension-reduction treatment reduces the complexity of network, improves training effectiveness.G1 in Fig. 1 and Fig. 2 indicates first convolutional layer,
G2 indicates second convolutional layer, and V1 indicates that first pond layer, G2 indicate second pond layer, up to the 6th convolutional layer and the
Six pond layers.
In order to further speed up the rate of convergence of network system, prevent network system from occurring gradient during training full
And the phenomenon that, the Sigmoid function in LeNet-5 network is replaced in present embodiment with Relu function.Relu function is one
Piecewise function, the gradient of Relu function is 1 in x >=0 compared with Sigmoid function, otherwise is 0.Disappear completely in the part x >=0
In addition to the gradient saturation effect of Sigmoid function；On computation complexity, Relu function is also relatively easy.Relu function simultaneously
The convergence rate of system is helped to improve, convergence rate is relative to about fast 6 times or so of Sigmoid function.
Simoid function expression are as follows:
σ (x) indicates the neuron output response of Simoid function in formula, and x indicates the input of neuron；
Relu function expression are as follows:
ReLu (x) indicates the neuron output response of Relu function in formula.
The step of acquisition fusion feature sequence, is obtained by manually extracting characteristics of image and carrying out linear weighted function to it
Weighted feature sequence finally carries out fused in tandem with the extracted feature of improvement CNN module；This is not only able to improve the knowledge to image
Other precision, also increases the interpretation of characteristic of division to a certain extent.Finally obtained by artificial extraction feature and automatically
Feature carries out fusion and obtains fusion feature sequence as characteristic of division.
For present embodiment by improving the step of CNN module is to image procossing, the image high dimensional feature of acquisition is image shape
The very low abstract characteristics of the interpretable degree of formula；The characteristic information manually extracted in Fusion Features step mainly pass through PCA it is main at
HOG feature and SIFT feature after analysis；HOG feature, SIFT feature are input to full connection together with image high dimensional feature
Layer, to achieve the purpose that fused in tandem.
Present embodiment has the function that Automatic Feature Extraction and classification are carried out based on naval vessel infrared signature, has good
Discrimination, generalization ability and a degree of feature interpretation can be ShipTargets early warning detection, rescue at sea etc.
Strong technical support is provided.
Further, as shown in connection with fig. 1, the manual features packet for manually extracting the input naval vessel infrared image
It includes, extracts the HOG feature and SIFT feature of input naval vessel infrared image.HOG feature is grasped on the local pane location of image
It obtains, so it is able to maintain good invariance to the geometry and optical deformation of image；And it can be in image according to SIFT feature
In detect key point, so SIFT feature can be used as a kind of local feature description.This has with the high dimensional feature that CNN is extracted
There is good complementarity.
Further, as shown in connection with fig. 3, it is described to manual features information carry out processing include:
Main feature is refined using PCA Principal Component Analysis to the HOG feature and SIFT feature, to the main feature
K- neighbour Cluster Classification is carried out, multiple tagsort groups are obtained；
The method for obtaining weighted feature sequence includes,
The error in classification for analyzing main feature described in each tagsort group, according to error in classification to each tagsort
Main feature in group carries out linear weighted function, obtains the weighted feature sequence of all main features corresponding to tagsort group.
In present embodiment, Fusion Features module carries out HOG feature to input naval vessel infrared image and SIFT feature mentions
After taking, dimensionality reduction operation is carried out to it using PCA, only retains main feature, it is described to be mainly characterized by from HOG feature and SIFT
Feature after the dimensionality reduction of feature.
Further, the weight w for carrying out linear weighted functioniAre as follows:
In formula, i is characterized the ordinal number of sorting group, and M is characterized the total number of sorting group, eiFor error in classification.
Error in classification eiIt can be used as weighting foundation, according to error in classification and calculate the weight obtained, it can be to corresponding
Main feature in tagsort group is weighted.
Further, referring to figs. 1 and 2, the full articulamentum introduces Dropout algorithm and carries out to fusion feature sequence
Classification.
In present embodiment, according to actual needs, for the objective factor of naval vessel infrared image negligible amounts, connecting entirely
Dropout algorithm is introduced in layer, be can effectively improve the generalization ability of neural network, is prevented from drawing because training set of images is less
The training system over-fitting risen.Meanwhile the Dropout algorithm of full articulamentum can select during training parameter each time
Property ignores partial nerve member parameter.Finally, be input to fusion feature sequence as characteristic of division in full articulamentum, it can be quick
Precisely obtain the output category of target image.
In present embodiment, system parameter can be carried out using error backpropagation algorithm and stochastic gradient descent method
Training, algorithm parameter initialization use Xavier parameter initialization method.
In conclusion the present invention may be implemented to carry out fast accurate identification to different Ship Targets, possess better identification
Rate and better generalization ability.
Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used
To modify the technical solutions described in the foregoing embodiments or equivalent replacement of some of the technical features；
And these are modified or replaceed, technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution spirit and
Range.
Claims (10)
1. a kind of CNN ship seakeeping system based on Fusion Features, characterized by comprising:
CNN module is improved, it is automatic to obtain image higher-dimension for carrying out multilayer convolution and pondization operation to input naval vessel infrared image
Characteristic information；The improvement CNN module is based on LeNet-5 convolutional neural networks and is changed acquisition: including the convolution using 3*3
Core, the convolution number of plies are 6, and pond mode uses maximum value pond mode, and excitation function is Relu function；
Fusion Features module believes manual features for manually extracting the manual features information of the input naval vessel infrared image
Breath is handled, and is carried out linear weighted function and obtained weighted feature sequence；The weighted feature sequence is obtained with based on improvement CNN module
The image high dimensional feature information obtained carries out fused in tandem, obtains fusion feature sequence；
Classification output module realizes target identification for classifying using full articulamentum to fusion feature sequence.
2. the CNN ship seakeeping system according to claim 1 based on Fusion Features, it is characterised in that:
The manual features information for manually extracting the input naval vessel infrared image includes extracting input naval vessel infrared image
HOG feature and SIFT feature.
3. the CNN ship seakeeping system according to claim 2 based on Fusion Features, it is characterised in that: described right
Manual features information carries out processing
Main feature is refined using PCA Principal Component Analysis to the HOG feature and SIFT feature, the main feature is carried out
K- neighbour's Cluster Classification obtains multiple tagsort groups；
The method for obtaining weighted feature sequence includes,
The error in classification for analyzing main feature described in each tagsort group, according to error in classification in each tagsort group
Main feature carry out linear weighted function, obtain corresponding to tagsort group all main features weighted feature sequence.
4. the CNN ship seakeeping system according to claim 3 based on Fusion Features, it is characterised in that: it is described into
The weight w of row linear weighted functioniAre as follows:
In formula, i is characterized the ordinal number of sorting group, and M is characterized the total number of sorting group, eiFor error in classification.
5. the CNN ship seakeeping system according to any one of claim 1 to 4 based on Fusion Features, feature
Be: the full articulamentum introduces Dropout algorithm and classifies to fusion feature sequence.
6. a kind of CNN ship seakeeping method based on Fusion Features, characterized by comprising:
Multilayer convolution and pondization operation are carried out to input naval vessel infrared image using CNN module is improved, it is automatic to obtain image higher-dimension
The step of characteristic information；The improvement CNN module is based on LeNet-5 convolutional neural networks and is changed acquisition: including using 3*3
Convolution kernel, the convolution number of plies be 6, pond mode use maximum value pond mode, excitation function be Relu function；
The manual features information for manually extracting the input naval vessel infrared image, is handled manual features information, and carry out
Linear weighted function obtains the step of weighted feature sequence；And using Fusion Features module to the weighted feature sequence be based on change
The step of image high dimensional feature information obtained into CNN module carries out fused in tandem, obtains fusion feature sequence；
The step of classifying to fusion feature sequence by the full articulamentum for output module of classifying, realize target identification.
7. the CNN ship seakeeping method according to claim 6 based on Fusion Features, it is characterised in that:
The manual features information for manually extracting the input naval vessel infrared image includes extracting input naval vessel infrared image
HOG feature and SIFT feature.
8. the CNN ship seakeeping method according to claim 7 based on Fusion Features, it is characterised in that: described right
Manual features information carries out processing
Main feature is refined using PCA Principal Component Analysis to the HOG feature and SIFT feature, the main feature is carried out
K- neighbour's Cluster Classification obtains multiple tagsort groups；
The method for obtaining weighted feature sequence includes,
The error in classification for analyzing main feature described in each tagsort group, according to error in classification in each tagsort group
Main feature carry out linear weighted function, obtain corresponding to tagsort group all main features weighted feature sequence.
9. the CNN ship seakeeping method according to claim 8 based on Fusion Features, it is characterised in that: it is described into
The weight w of row linear weighted functioniAre as follows:
In formula, i is characterized the ordinal number of sorting group, and M is characterized the total number of sorting group, eiFor error in classification.
10. the CNN ship seakeeping method according to any one of claims 6 to 9 based on Fusion Features, feature
Be: the full articulamentum introduces Dropout algorithm and classifies to fusion feature sequence.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910148840.5A CN109815967A (en) | 2019-02-28 | 2019-02-28 | CNN ship seakeeping system and method based on Fusion Features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910148840.5A CN109815967A (en) | 2019-02-28 | 2019-02-28 | CNN ship seakeeping system and method based on Fusion Features |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109815967A true CN109815967A (en) | 2019-05-28 |
Family
ID=66607664
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910148840.5A Pending CN109815967A (en) | 2019-02-28 | 2019-02-28 | CNN ship seakeeping system and method based on Fusion Features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109815967A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110782481A (en) * | 2019-10-18 | 2020-02-11 | 华中光电技术研究所（中国船舶重工集团有限公司第七一七研究所） | Unmanned ship intelligent decision method and system |
| CN111442827A (en) * | 2020-04-08 | 2020-07-24 | 南京艾森斯智能科技有限公司 | Optical fiber passive online monitoring system and method for transformer winding vibration |
| CN111563553A (en) * | 2020-05-08 | 2020-08-21 | 华北理工大学 | Method for predicting alkalinity of pellet ore phase based on PCA-CNN classification recognition |
| CN111639513A (en) * | 2019-12-10 | 2020-09-08 | 珠海大横琴科技发展有限公司 | Ship shielding identification method and device and electronic equipment |
| CN111695397A (en) * | 2019-12-20 | 2020-09-22 | 珠海大横琴科技发展有限公司 | Ship identification method based on YOLO and electronic equipment |
| CN111931558A (en) * | 2020-06-22 | 2020-11-13 | 武汉第二船舶设计研究所(中国船舶重工集团公司第七一九研究所) | Ship category identification method and system |
| CN112069958A (en) * | 2020-08-27 | 2020-12-11 | 广西柳工机械股份有限公司 | Material identification method, device, equipment and storage medium |
| CN112329835A (en) * | 2020-10-30 | 2021-02-05 | 天河超级计算淮海分中心 | Image processing method, electronic device, and storage medium |
| CN113365382A (en) * | 2021-08-10 | 2021-09-07 | 深圳市信润富联数字科技有限公司 | Light control method and device, electronic equipment and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105787458A (en) * | 2016-03-11 | 2016-07-20 | 重庆邮电大学 | Infrared behavior identification method based on adaptive fusion of artificial design feature and depth learning feature |
| CN107909082A (en) * | 2017-10-30 | 2018-04-13 | 东南大学 | Sonar image target identification method based on depth learning technology |
| CN108921030A (en) * | 2018-06-04 | 2018-11-30 | 浙江大学 | A kind of SAR automatic target recognition method of Fast Learning |
-
2019
- 2019-02-28 CN CN201910148840.5A patent/CN109815967A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105787458A (en) * | 2016-03-11 | 2016-07-20 | 重庆邮电大学 | Infrared behavior identification method based on adaptive fusion of artificial design feature and depth learning feature |
| CN107909082A (en) * | 2017-10-30 | 2018-04-13 | 东南大学 | Sonar image target identification method based on depth learning technology |
| CN108921030A (en) * | 2018-06-04 | 2018-11-30 | 浙江大学 | A kind of SAR automatic target recognition method of Fast Learning |
Non-Patent Citations (7)
| Title |
|---|
| 刘建 等，: "卷积神经网络综述", 《计算机时代》 * |
| 刘建伯 等，: "基于改进LeNet-5的油井井号识别方法", 《自动化技术与应用》 * |
| 周奇，: "基于多特征的轮船运动目标跟踪及轨迹获取方法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 周飞燕 等，: "卷积神经网络研究综述", 《计算机学报》 * |
| 张建虎，: "面向目标识别的多特征融合研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 王佩军 等，: "《摄影测量学》", 31 May 2016, 武汉大学出版社 * |
| 苏欣 等，: "《Android手机应用网络流量分析与恶意行为检测研究》", 31 October 2016, 湖南大学出版社 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110782481A (en) * | 2019-10-18 | 2020-02-11 | 华中光电技术研究所（中国船舶重工集团有限公司第七一七研究所） | Unmanned ship intelligent decision method and system |
| CN110782481B (en) * | 2019-10-18 | 2023-09-05 | 华中光电技术研究所（中国船舶重工集团有限公司第七一七研究所） | Unmanned ship intelligent decision-making method and system |
| CN111639513A (en) * | 2019-12-10 | 2020-09-08 | 珠海大横琴科技发展有限公司 | Ship shielding identification method and device and electronic equipment |
| CN111695397A (en) * | 2019-12-20 | 2020-09-22 | 珠海大横琴科技发展有限公司 | Ship identification method based on YOLO and electronic equipment |
| CN111442827A (en) * | 2020-04-08 | 2020-07-24 | 南京艾森斯智能科技有限公司 | Optical fiber passive online monitoring system and method for transformer winding vibration |
| CN111442827B (en) * | 2020-04-08 | 2022-07-19 | 南京艾森斯智能科技有限公司 | Optical fiber passive online monitoring system for transformer winding vibration |
| CN111563553A (en) * | 2020-05-08 | 2020-08-21 | 华北理工大学 | Method for predicting alkalinity of pellet ore phase based on PCA-CNN classification recognition |
| CN111931558A (en) * | 2020-06-22 | 2020-11-13 | 武汉第二船舶设计研究所(中国船舶重工集团公司第七一九研究所) | Ship category identification method and system |
| CN112069958A (en) * | 2020-08-27 | 2020-12-11 | 广西柳工机械股份有限公司 | Material identification method, device, equipment and storage medium |
| CN112329835A (en) * | 2020-10-30 | 2021-02-05 | 天河超级计算淮海分中心 | Image processing method, electronic device, and storage medium |
| CN113365382A (en) * | 2021-08-10 | 2021-09-07 | 深圳市信润富联数字科技有限公司 | Light control method and device, electronic equipment and storage medium |
| CN113365382B (en) * | 2021-08-10 | 2021-11-09 | 深圳市信润富联数字科技有限公司 | Light control method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109815967A (en) | CNN ship seakeeping system and method based on Fusion Features | |
| CN106203395B (en) | Face attribute recognition method based on multitask deep learning | |
| CN106599854B (en) | Automatic facial expression recognition method based on multi-feature fusion | |
| CN108615010B (en) | Facial expression recognition method based on parallel convolution neural network feature map fusion | |
| Hertel et al. | Deep convolutional neural networks as generic feature extractors | |
| CN106650806B (en) | A kind of cooperating type depth net model methodology for pedestrian detection | |
| US5245672A (en) | Object/anti-object neural network segmentation | |
| Burie et al. | ICFHR2016 competition on the analysis of handwritten text in images of balinese palm leaf manuscripts | |
| CN106709568A (en) | RGB-D image object detection and semantic segmentation method based on deep convolution network | |
| Delakis et al. | Text detection with convolutional neural networks | |
| CN107808132A (en) | A kind of scene image classification method for merging topic model | |
| Ahranjany et al. | A very high accuracy handwritten character recognition system for Farsi/Arabic digits using convolutional neural networks | |
| CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
| CN111414862A (en) | Expression recognition method based on neural network fusion key point angle change | |
| CN104992223A (en) | Intensive population estimation method based on deep learning | |
| CN110188615A (en) | A kind of facial expression recognizing method, device, medium and system | |
| CN109344856B (en) | Offline signature identification method based on multilayer discriminant feature learning | |
| CN111145145B (en) | Image surface defect detection method based on MobileNet | |
| CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
| CN110427912A (en) | A kind of method for detecting human face and its relevant apparatus based on deep learning | |
| CN106874929A (en) | A kind of pearl sorting technique based on deep learning | |
| Sarigül et al. | Comparison of different deep structures for fish classification | |
| CN105956570A (en) | Lip characteristic and deep learning based smiling face recognition method | |
| CN113221956A (en) | Target identification method and device based on improved multi-scale depth model | |
| CN110991554B (en) | Improved PCA (principal component analysis) -based deep network image classification method |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190528 |