CN107609601A - A kind of ship seakeeping method based on multilayer convolutional neural networks - Google Patents

Abstract

The invention discloses a kind of ship seakeeping method based on multilayer convolutional neural networks to include：S1, naval vessel Sample Storehouse are constantly enriched using existing image, parameter and model data structure, and in use by detecting target data collection；S2, Ship Target features training are under the framework of convolutional neural networks, by the recognition training to naval vessel Sample Storehouse, the naval vessel feature knowledge storehouse that visible ray/infrared and two-dimensional/three-dimensional merges are formed, for carrying out Ship target recognition identification；S3, Ship Target data acquisition are used for the collection that Real-time High Resolution rate is carried out to the visible ray or IR video stream of ShipTargets；S4, ShipTargets are detected；S5, to Ship Target image rough sort；The disaggregated classification that S6, the deep neural network model completed based on Ship Target features training carry out Ship Target identifies work, accurately identifies the type on naval vessel.Solves the problem of ship seakeeping.

Description

A kind of ship seakeeping method based on multilayer convolutional neural networks

Technical field

The present invention relates to a kind of target identification method, more particularly to the ship seakeeping based on multilayer convolutional neural networks Method.

Background technology

China possesses wide coastline, marine site and abundant marine resources, as economy continues to develop, marine vessel Quantity is more and more, and naval vessel detection has urgent actual demand；And the naval vessels of neighboring countries and regions, civilian boat etc. are often non- Method is engaged in the activities such as measurement, monitoring and fishing into the legal marine site in China so that the legal maritime rights and interests of country cannot be effective Ensure, ocean right-safeguarding situation becomes increasingly complex with maritime safety by serious threat, strengthened to warship in the range of the legal marine site in China The detection identification of the targets such as ship is of great practical significance.

Weaker to the image reconnaissance capability in China marine site at present, intelligence degree is relatively low.From aviation or satellite photography to obtaining Accurate information is taken to need the longer time cycle, its result is also not fully up to expectations, and with surface vessel, civilian boat and other are dry Disturb species, increasing, the complexity more and more higher of sea environment of quantity of target, the high speed renewal of naval vessel and civilian boat form Regenerate, the current many problems of Ship Recognition systems face, its recognition correct rate and availability have been difficult to meet to require.At present The ship seakeepings such as widely used constant false alarm rate (Constant False Alarm Rate, CFAR) algorithm, wavelet transformation Algorithm is 80%~85% for the accuracy of ship seakeeping；Modern sea environmental requirement naval vessel, civilian boat target identification are just True rate requirement reaches more than 95%, and traditional ship seakeeping algorithm scope of application is single, is identified when handling complex target Accuracy rate is relatively low, and far from meeting requirement of the modern sea environment to higher recognition correct rate.Current domestic Ship Target Detection identification is mostly based on the satellite remote sensing images data such as SAR, IR, but this method has the drawbacks of serious.Satellite data by In long with the distance on ground, cause transmission speed slow, Ship Target data are unable to real-time update, can not find and identify in time Specify the ship occurred on marine site；Satellite remote sensing can not clearly shoot sea level chart picture in thunderstorm cloudy weather, cause severe sea Ship seakeeping accuracy under environment reduces.Meanwhile all kinds of military, civilian naval vessel Modality rehabilitations are quickly, using general side The Ship Recognition image library of method generation needs frequently to upgrade naval vessel, adds the maintenance time and cost on naval vessel, intelligence It is not high degree to be changed.

The content of the invention

It is an object of the invention to provide a kind of ship seakeeping method based on multilayer convolutional neural networks, for solving The certainly problem of ship seakeeping.

A kind of ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, including：S1, using Some images, parameter and model data structure naval vessel Sample Storehouse, and improve naval vessel Sample Storehouse by detecting target data collection； S2, under the framework of convolutional neural networks, by the recognition training to naval vessel Sample Storehouse, form visible ray/infrared and two-dimentional/tri- The naval vessel feature knowledge storehouse of fusion is tieed up, for carrying out Ship target recognition identification；Including：Foundation is based on depth convolutional neural networks Ship Target features training framework, including feature share CNN layers, fine-tuning CNN layers, extracted region model, ROI ponds layer with And classification returns full articulamentum；Sharing feature CNN layers are that existing depth convolutional neural networks disaggregated model removes and finally connected entirely The part of layer is connect, before Ship Target Detection identification model training is carried out, in VOC categorized data sets or ImageNet classification numbers The extractability to characteristics of image is improved according to pre-training is carried out on collection, then recycles this disaggregated model trained to remove entirely The parameter for connecting layer segment carries out parameter initialization to sharing feature CNN layers；Fine-tuning CNN layers are used for online incremental learning；Area Domain extraction model is convolutional neural networks, after shared CNN layers and fine-tuning CNN layers, constructs a convolutional layer and two simultaneously The neutral net of the full articulamentum of row, by shared CNN layers feature and fine-tuning CNN layers feature come realize there may be it is to be checked The extraction in Ship Target region；ROI ponds layer, for behind ROI ponds, exporting the characteristic vector of fixed dimension；Naval vessel mesh Mark detection identification model training includes 4 step sections；First stage：Using the disaggregated model of the progress pre-training on Sample Storehouse to altogether Feature CNN layer parameters are enjoyed to be initialized, at the same using the Gauss normal distribution that zero-mean variance is σ to fine-tuning CNN layers with And extracted region model carries out the parameter initialization of weight and bias term, visible ray or infrared training set picture number are finally recycled It is finely adjusted according to extracted region neural network model；Second stage：Use the area of extracted region model extraction in the first stage Domain extraction training sharing feature CNN layers, fine-tuning CNN layers, at the same it is pre- using being carried out on Voc2012 or ImageNet data sets The disaggregated model initialization sharing feature CNN layer parameters of training；Phase III：The sharing feature trained using second stage CNN layers, fine-tuning CNN layer parameters, the sharing feature CNN layers of extracted region model, fine-tuning CNN layer parameters are reinitialized, Then sharing feature CNN layers, fine-tuning CNN layer parameters are fixed and extracted region neural network model is finely adjusted；Fourth order Section：The convolution layer parameter of extracted region neural network model in phase III is freezed and extracts extracted region, then shares spy Sign CNN layers, fine-tuning CNN layer parameters are finely adjusted；S3, visible ray or IR video stream to ShipTargets carry out real When high-resolution collection；S4, ShipTargets are detected；S5, Ship Target image rough sort are directed to naval vessel Major class is simply classified, and reduces the workload of follow-up Ship Target disaggregated classification identification；S6, based on Ship Target features training The deep neural network model of completion carries out the disaggregated classification identification work of Ship Target, accurately identifies the type on naval vessel.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S4, Ship Target Detection specifically includes：Calculated using SLIC image segmentation algorithms, SLIC algorithms concretely comprise the following steps：Set K Super-pixel point number initializes the center C of group_k=[l_k, a_k, b_k, x_k, y_k], k ∈ [1, K]；Upset in neighborhood in group The heart, cluster centre is moved to the minimum position of gradient；For the center of each group, according to apart from computing rule, around Best match pixel is distributed in the adjacent area at a group center；Calculate new group center and residual error；Until Residual error E, which is less than threshold values, to restrain.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S4 In image gradient calculate include：

G (x, y)=| | I (x+1, y)-I (x-1, y) | |²+ | | I (x, y+1)-I (x, y-1) | |²(1)；

Wherein I (x, y) is lab vectors, represents the color vector of pixel space position (x, y), | | | | it is L2 norms；Setting step Long S=sqrt (N/K), N are sum of all pixels, and K is super-pixel number, is each cluster in the 2S*2S of cluster centre neighborhood Center distribute match point, calculated according to the setpoint distance of each point and cluster centre, each pixel in image finally with most Near cluster centre connects, and is covered by the region of search of the cluster centre, is associated in recently in all pixels Group center after, new center will be computed, and be the average value of all labxy vectors for belonging to the cluster, repeat with Nearest group's core contextual pixel and this process of group center is recalculated until convergence.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S4 In, the core in HOG extraction process includes：1) gradient in image abscissa and ordinate direction is calculated, and is calculated accordingly each The gradient direction value of location of pixels；2) image is evenly dividing fritter according to locus, according to set quantization in fritter The histogram of canonical statistics gradient direction, characteristic vector corresponding to fritter is obtained, then again all cell in a bulk Characteristic vector, which is together in series, just obtains the HOG features of bulk, wherein, each bulk is made up of m × m fritter, and fritter by n × N pixel composition.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, After each extracted region feature in training data, suitable grader is trained to complete the classification in region, is excluded with realizing Sea, the bulk region of sky, navigate to the target of area-of-interest.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S6 Including：The identification of warship and civilian boat；Note training set is { (x₁, t₁) ... ..., (x_N, t_N), wherein x_iIt is to input, t_iIt is that target is defeated Go out, N is the sample number of training set, and the integrated output of deep neural network is defined as：Wherein m is god Number through network, F_i(n) it is output of i-th of deep neural network on n-th of sample, and F (n) is then depth nerve net Network is integrated in the output on n-th of sample, and the ResNet deep neural networks used are used for integrated study, so as to reach to newly arriving The incremental learning of image data set, including four modules：Initial R esNet deep neural networks are integrated, current ResNet depth is refreshing Integrated through system integrating, the integrated and new ResNet deep neural networks of duplication ResNet deep neural networks；First with One warship and civilian boat view data set pair Initial R esNet deep neural networks, which integrate, to be trained to obtain current ResNet depths Spend Artificial neural network ensemble；Then, current ResNet deep neural networks are integrated to be replicated to obtain and replicates ResNet depth god Through system integrating；Next, when second warship and civilian boat image data set arrive, using the data set to replicating ResNet Deep neural network is integrated to be trained, and is obtained new ResNet deep neural networks and is integrated；Finally, learned using selectivity is negatively correlated Learning method is selected.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S6 Further comprise：When n-th of sample is used for training, the error function of i-th of neutral net is defined as：

Wherein p_i(n) it is penalty term, λ ∈ [0,1] are a control parameter, are referred to as punishing Penalty parameter, for controlling one between mean square error and penalty term to balance, penalty term p_i(n) may be defined as：p_i(n)=(F_i (n)-F(n))∑_j≠iF_j(n)-F (n), remaining neutral net during penalty term causes i-th of neutral net and integrated is negatively correlated, There is difference between individual neutral net so as to reach, as λ=0, obtained Artificial neural network ensemble is equivalent to stand-alone training one Group neutral net；And as λ continuous increase, the emphasis of training will gradually be adjusted to have in otherness between order individual.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S6 Further comprise：A) the Artificial neural network ensemble ens1 that a size is m is initialized；B) trained with S1 by negative correlation learning ens1；C) from second Sub Data Set start to finish, i.e. t=2 ..., T, following cyclic process is performed：Replicate Current neural net Network integrates enst-1, obtains a copy, is designated as enscopy；Enscopy is trained by negative correlation learning with St；Merge enst- 1 and enscopy, enscomb is integrated into after note merging；Selection course is used to enscomb, m neutral net is selected and forms Integrated enst；D) final Artificial neural network ensemble ensT is exported；For selection course in c), using genetic algorithm to enscomb Selected, it can be seen from the as above negatively correlated integrated learning approach of selectivity, selection course is selected from the enscomb that size is 2m M neutral net is selected to form enst, selection course can turn to the optimization problem of following belt restraining in the form of：

Wherein, J (ω) is predefined object function, and its quality designed will be directly connected to the nerve net chosen The integrated Generalization Capability of network.ω_iIt is the binary vector of 2m dimension, ω_i=1 i-th of neutral net of expression is selected for structure Into enst, ω_i=0 represents that i-th of neutral net is deleted；Carrying out solution using genetic algorithm includes：A) initialization algorithm Set：Integrated size m, GA Population Size pop_size, the Probability p of intersection_c, the Probability p of variation_mAnd fitness function J (ω)；B) initial population being made up of ω is randomly generated；C) repeat the steps of, until reaching defined number：Assess current The fitness of each individual in population；Intersect using roulette selection parent individuality, and with single-point with Probability p_cWith variation with general Rate p_mProduce offspring；Each individual ω is repaired with Greedy strategy, if ∑_iω_i＞ m, by the individual ω that one of them is 1_iIt is changed to 0, So that J (ω) is minimum, and the process is repeated until ∑_iω_i=m, if ∑_iω_i＜ m, by the individual ω that one of them is 0_iIt is changed to 1 so that J (ω) is minimum, and repeats the process until ∑_iω_i=m；D) output wopt selects as optimal solution, while according to wopt Neutral net corresponding to wopt=1 is selected to form final Artificial neural network ensemble.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, bag Include：Fine-tuning CNN layers are used for online incremental learning, when carrying out Ship Target Detection online, if the naval vessel mesh being detected Mark meets mark condition, then needs to carry out target detection model online incremental learning, needed during online incremental learning by Sharing feature CNN layers, extracted region model and the follow-up ROI ponds layer of fine-tuning CNN layers, full connection layer parameter are consolidated It is fixed, small parameter perturbations are then carried out to fine-tuning CNN layers by error in classification.

According to an embodiment of the ship seakeeping method based on multilayer convolutional neural networks of the present invention, wherein, S2 Also include：In the model reasoning stage, if the naval vessel identified using Ship Target Detection identification model meets that sample marks bar Part, then need to carry out online incremental learning to the fine-tuning CNN layers of Ship Target Detection identification model using the sample, including： The first step：The ship seakeeping threshold value predicted using target detection identification model is judged sample, if the threshold value Higher than σ₁Or less than σ₂Online incremental learning need not then be carried out；, whereas if the threshold value is less than σ₁And it is higher than σ₂Then need to carry out Once online incremental learning；Second step：Using meet sample mark condition view data to the fine-tuning of target detection model CNN layers carry out primary parameter fine setting.

The ship seakeeping method based on multilayer convolutional neural networks of the present invention, using Ship Target characteristics of image and Massive Sample view data, carries out the Ship Target artificial intelligence identification technology research based on big data, and invention is based on artificial intelligence The image procossing and ship seakeeping system of energy, improve ship seakeeping ability and accuracy.Present disclosure includes Naval vessel and civilian boat video image characteristic database sharing, propose naval vessel and civilian boat feature based on depth convolutional neural networks framework The fusion method of extraction and identification, the sorting technique on the naval vessel based on deep learning and civilian boat, the parameter instruction based on deep learning Practice method, establish image procossing and ship seakeeping system based on artificial intelligence.

Brief description of the drawings

Fig. 1 show image procossing and ship seakeeping system module figure based on artificial intelligence

Fig. 2 show image procossing and ship seakeeping system framework figure based on artificial intelligence；

Fig. 3 show the Ship Target features training Organization Chart based on depth convolutional neural networks

The schematic diagram of SLIC partitioning algorithms shown in Fig. 4；

Fig. 5 show complete HOG feature extraction algorithms and procedure chart；

Fig. 6 show rectangular element cell schematic diagram；

Fig. 7 show the schematic diagram of HOG 9 direction blocks；

Fig. 8 show ship, sky, the HOG feature schematic diagrames of ocean；

Fig. 9 show positioning exemplary plot at target；

Figure 10 show the flow chart of visible optical target rough sort；

Figure 11 show SIFT and describes subcharacter extraction and description flow chart；

Figure 12 show the feature extraction result figure of SIFT description；

Figure 13 show the feature extraction result figure of dense-SIFT description；

Figure 14 show the feature extraction figure of multiple dimensioned dense-SIFT descriptions；

Figure 15 show the depth ResNet integrated study frame diagrams based on selective negative correlation learning；

Figure 16 show Artificial neural network ensemble selection algorithm block diagram.

Embodiment

To make the purpose of the present invention, content and advantage clearer, with reference to the accompanying drawings and examples, to the present invention's Embodiment is described in further detail.

Image procossing and detection target identification system framework based on artificial intelligence include naval vessel Sample Storehouse structure, naval vessel mesh Mark features training, Ship Target data acquisition, Ship Target Detection, Ship Target rough sort, Ship Target disaggregated classification identification six Individual part.

Fig. 1 show image procossing and ship seakeeping system module figure based on artificial intelligence, as shown in figure 1, base Include in the image procossing and ship seakeeping system of artificial intelligence：Network image database 1, Ship Target acquisition module 2, Naval vessel sample library module 3, Ship Target Detection module 4, ship seakeeping convolutional neural networks model 5, Ship Target rough segmentation Generic module 6, Ship Target subdivision generic module 7.

A kind of ship seakeeping method based on multilayer convolutional neural networks of the present invention includes：

S1, naval vessel Sample Storehouse in use, pass through inspection using existing image, parameter and model data structure Target data collection is surveyed constantly to be enriched；

S2, Ship Target features training are under the framework of convolutional neural networks, are instructed by the identification to naval vessel Sample Storehouse Practice, the naval vessel feature knowledge storehouse of visible ray/infrared and two-dimensional/three-dimensional fusion is formed, for carrying out Ship target recognition identification；

S3, Ship Target data acquisition are used to carry out the visible ray or IR video stream of ShipTargets height in real time The collection of resolution ratio；

S4, ShipTargets are detected；

S5, to Ship Target image rough sort；

S6, the deep neural network model completed based on Ship Target features training carry out the disaggregated classification identification of Ship Target Work, accurately identify the type on naval vessel.

As shown in figure 1, the operation of image procossing and detection target identification system based on artificial intelligence includes three phases： View data stage, model off-line training step and model On-line accoun stage are obtained, in each stage again comprising different Functional module.

1) the view data stage is obtained, including：

Including two groups of data, for training the data set of Ship Target Detection identification model and for warship and civilian boat point The data set of class, training dataset and test data of this two classes view data comprising infrared image and visible images Collection.Include Infrared Image Information acquisition module, image pre-processing module and extra large land separation module at this stage.

2) model off-line training step, including：

It is mainly used in the off-line training of Ship Target Detection identification model and warship and civilian boat disaggregated model.

3) the model On-line accoun stage, including：

Fig. 2 show image procossing and ship seakeeping system framework figure based on artificial intelligence, as shown in Fig. 2 should Stage is in addition to the detection identification applied to Ship Target and warship and civilian boat classification task, in addition to Ship Target Detection mould The online incremental detection identification of type.

As shown in Figure 1 and Figure 2, naval vessel sample base construction method specifically includes in S1：

The target detection based on supervised learning and the view data of disaggregated model is trained to include training set, checking collection and survey Examination collection, wherein checking collection can separate a part from training set and obtain, the ratio shared by them is 50%：25%：25%.With In the training of Ship Target Detection identification model and test image data set and the training for warship and civilian boat disaggregated model and Very big difference be present in test image data.It is right thereon that training and test image data for Ship Target Detection identification need Ship Target carries out label and confines and give the mark of naval vessel classification；And it is used for training and the test image of warship and civilian boat classification Data only need to give corresponding category label to warship image and civilian boat image.

The collection of visible images sample and the basis that Database is visible ray object detection and recognition, the number of image Amount, quality, goal congruence, rich very crucial effect all is established with the model of identification to final detect.

Obtained by multiple channel, contrast images data, the present invention is schemed using ImageNet, Voc2012 and Fleetmoom As database.ImageNet databases are set up by Stanford Univ USA, are the maximum numbers of current field of image recognition According to storehouse, about 22000 class targets, the type uncalibrated image of quantity about 15,000,000 or so are included altogether.

According to the demand analysis of scene and project, the foundation of database is built mainly around offshore and marine naval vessel It is vertical.By analysis, it is usually required mainly for establish the target database of three major types, i.e. military boats and ships, civilian naval vessel and jamming target.With Exemplified by ImageNet, the download of database and fileinfo generating process are as follows：

1) ship major classes are obtained from ImageNet；Utilize downloadstatus, releasedstatus, structure_ Released information acquisitions databseinfo.xml；

2) warship major classes are obtained from ImageNet；Also with downloadstatus, releasedstatus, Structure_released information acquisitions databasinfo.xml (gendatabase)；

3) nonwarship databaseinfo.xml is generated using the databaseinfo of the generation of two steps above (differencedatabase)；

4) classes such as aegeanisland, barrierreef, lighthouse are obtained from ImageNet.

The one-level screening of data：

Select principle：Screened according to title

Whether 30 classes such as warship all retain, and whether nonwarship 43 classes all retain, and refer to warship guarantor Stay information, nonwarhip reservation information (being preserved with xml text messages)；

Using artificial reservation information above, mobile file clips to warship V1 and Nonwarship V1, other In V1.

The two level screening of data：

Select principle:

1) side view is ensured as far as possible, to ensure fully obtain target signature.

2) image object is placed in the middle.

3) image object can not be too small, excludes the image less than 400*400 pixels.

4) ambient interferences of image are as far as possible few；

5) image object completely will be presented in image, partly should give exclusion outside field of view.

6) feature of target in itself can significantly be embodied.

After two-stage is screened, the naval vessel data that target is clear, angle is moderate can be obtained.Divide according to project demands Analysis and scene are it is assumed that interference class view data mainly have chosen reef, beacon and the class of island three.The other view data of three major types As shown in table 1, some groups are divided again in every one kind.

Table 1

Classification	Before screening	After screening
			Warship	30 classes	21 classes
Nonwarship	43 classes	11 classes
			Other	3 classes	3 classes

S2, Ship Target features training include：

It is the key link of Ship Recognition by Ship Target features training module, it is main to include being based on depth convolutional Neural The Ship Target features training architecture design of network model, two and three dimensions Ship Target fusion recognition, model training and The contents such as line incremental learning.

Fig. 3 show the Ship Target features training Organization Chart based on depth convolutional neural networks, as shown in figure 3, being based on The Ship Target features training framework of depth convolutional neural networks includes feature and shares CNN layers, fine-tuning CNN layers, extracted region Model, ROI ponds layer and classification return full articulamentum etc..

Sharing feature CNN layers can be the portion that existing depth convolutional neural networks disaggregated model removes last full articulamentum Point.The part is generally required in VOC categorized data sets or ImageNet before Ship Target Detection identification model training is carried out Pre-training is carried out on categorized data set to improve the extractability to characteristics of image, the disaggregated model for then recycling this to train The parameter for removing full connection layer segment carries out parameter initialization to sharing feature CNN layers.The classification of conventional sharing feature CNN layers Model is including AlexNet, GoogleNet, VGG16, VGG19 and ResNet etc..

Fine-tuning CNN layers are mainly used in online incremental learning.Detailed process is：When carrying out Ship Target Detection online, If the Ship Target being detected meets mark condition, need to carry out online incremental learning to target detection model.Online Needed during incremental learning by sharing feature CNN layers, extracted region model and the follow-up ROI ponds layer of fine-tuning CNN layers, Full connection layer parameter is fixed, and then carries out small parameter perturbations to fine-tuning CNN layers by error in classification, so as to reach from new figure Ability of the study to knowledge before retaining as data learning new knowledge and as far as possible.

Extracted region model is convolutional neural networks, after shared CNN layers and fine-tuning CNN layers, constructs a convolution The neutral net of layer and two full articulamentums arranged side by side.The effect of the model is by shared CNN layers feature and fine-tuning CNN Layer feature there may be the extraction in Ship Target region to be checked to realize.On the basis of the model, in order to solve Small object hardly possible Problem is detected, image context Context information can be added, so as to effectively solve the problem.At present, more commonly used area Domain extraction model include Faster RCNN proposed based on Attention mechanism RPN extracted regions model, based on neighborhood information And the automatic adaption object of AZ-Net models and energy of self adaptive pantographic Anchors strategies and its HAZN moulds of part yardstick Type, these extracted region models solve that picture resolution is relatively low, the difficult inspection of small size target by improving the quality of extracted region The problem of survey.

The mentality of designing of ROI ponds layer derives from Faster RCNN, and the design is to solve to detect target detection identification Model can receive the problem of any yardstick input picture.Behind ROI ponds, model can export the feature of a fixed dimension Vector, can connect thereafter Softmax or algorithm that correlation energy is classified, such as support vector machines, be made whether as warship The classification of ship target.

The training of Ship Target Detection identification model includes 4 step sections.First stage：Using carrying out pre-training on Sample Storehouse Disaggregated model sharing feature CNN layer parameters are initialized, while be σ's (typically taking 0.001) using zero-mean variance Gauss normal distribution carries out the parameter initialization of weight and bias term to fine-tuning CNN layers and extracted region model, again finally Extracted region neural network model is finely adjusted using visible ray or infrared training set view data；Second stage：Use Extracted region training sharing feature CNN layers, the fine-tuning CNN layers of extracted region model extraction in one stage, while utilize The disaggregated model initialization sharing feature CNN layer parameters of pre-training are carried out on Voc2012 or ImageNet data sets；3rd rank Section：Sharing feature CNN layers, the fine-tuning CNN layer parameters trained using second stage, reinitializes extracted region model Sharing feature CNN layers, fine-tuning CNN layer parameters, then fix sharing feature CNN layers, fine-tuning CNN layer parameters and region is carried Neural network model is taken to be finely adjusted；Fourth stage：By the convolution layer parameter of extracted region neural network model in the phase III Freeze and use the extracted region model extraction extracted region, then sharing feature CNN layers, fine-tuning CNN layer parameters are carried out micro- Adjust.

From training process as can be seen that the sharing feature CNN layers of Ship Target Detection identification model, fine-tuning CNN layers with Extracted region model needs to carry out alternately training, while needs in each training process to freeze a part of parameter, then trains Another part parameter, such benefit are the overall performances that can improve extracted region model and classification recurrence part simultaneously.

In the model reasoning stage, if the naval vessel identified using Ship Target Detection identification model meets that sample marks bar Part, then need to carry out online incremental learning to the fine-tuning CNN layers of Ship Target Detection identification model using the sample.Specifically do Method includes two steps, the first step：The ship seakeeping threshold value predicted using target detection identification model judged sample, If the threshold value is higher than σ₁Or less than σ₂(σ₁＞ σ₂, general σ₁Take 0.9, σ₂0.6) online incremental learning need not then be carried out by taking；Instead It, if the threshold value is less than σ₁And it is higher than σ₂Then need to carry out once online incremental learning；Second step：Using meet sample mark The view data of condition carries out primary parameter fine setting to the fine-tuning CNN layers of target detection model.

S3, Ship Target data acquisition include：

Connect the most using the Ship Target sample image that actual video sensor obtains and realistic objective detection environment-identification Closely, actual use scene can preferably be reacted.Therefore, the present invention will be monitored using full HD visible ray cmos image sensor Target is converted into picture signal, sends special image processing system to, and its function is completed by image data acquiring passage Collection to original image, and be input to processing computer and pre-processed.

S4, Ship Target Detection specifically include：

The detection of dynamic object is the basis of whole system, only detects accurately and in time and navigates in scene and occur Target, just can guarantee that follow-up target identification.Target detection in present invention application should ensure real-time, because battlefield wink Breath ten thousand becomes, and catching change not in time will bungle the chance of winning a battle；Ensure target detection and the accuracy rate of segmentation again, because the essence of target Really identification is only the core of whole system.

Therefore, the object detection method based on more algorithm fusions is used, plays each algorithm for above-mentioned requirements, the present invention Advantage, complete the task of detection.

(1) SLIC image segmentation algorithms

SLIC (simple linear iteration cluster) algorithm is a super-pixel segmentation algorithm based on clustering algorithm, empty by LAB Between and x, y pixel coordinate totally 5 dimension spaces calculate.The algorithm proposes a kind of brand-new distance calculating method and strengthens super picture It the regularity of plain shape, can not only split cromogram, compatible can also split gray-scale map, need to split it can in addition contain set Super-pixel quantity.SLIC efficiency when producing similar or more preferable segmentation can increase substantially.

SLIC algorithms need to set K super-pixel number when starting, uniformly divide in image according to the super-pixel number of setting With K cluster centre, cluster centre is upset in n*n (n typically takes 3) neighborhood, while moves cluster centre to neighborhood inside gradient Minimum position.The operation can not only avoid cluster centre be placed into edge can also reduce selection noise pixel point it is several Rate.

Image gradient calculation formula is as follows：

G (x, y)=| | I (x+1, y)-I (x-1, y) | |²+ | | I (x, y+1)-I (x, y-1) | |²(1)；

Wherein I (x, y) is lab vectors, represents the color vector of pixel space position (x, y), | | | | it is L2 norms.This Sample just contains color and strength information.

Step-length S=sqrt (N/K) is set, N is sum of all pixels, and K is super-pixel number.In the 2S*2S of cluster centre neighbour Match point is distributed for each cluster centre in domain, is calculated according to the setpoint distance of each point and cluster centre, it is each in image Pixel finally connects with nearest cluster centre, and is covered by the region of search of the cluster centre.In all pixels After point is associated in nearest group center, a new center will be computed, and the center is all to belong to the cluster The average value of lab xy vectors.Repeat with nearest group's core contextual pixel and to recalculate this process of group center straight To convergence.

After this process terminates, the color lost on a small quantity is present, i.e., a small amount of pixel in neighbouring larger segmentation section also There is same color, but be not associated with it.Such case is rarely found, but may increase, because the aggregation of SLIC algorithms does not have Have and clearly strengthen connectedness.So the final stage of algorithm needs to strengthen connectedness, nearby clustered by maximum to confirm not Intersecting segmentation section.

The time complexity of SLIC algorithms is O (N), relative to other image segmentation algorithms, it is only necessary to linear amount of calculation And amount of ram, segmentation efficiency is substantially increased, and the convenient deployment of the algorithm only needs 1 super-pixel number to join as input Number.

SLIC algorithms concretely comprise the following steps：

K super-pixel point number is set to initialize the center C of group_k=[l_k, a_k, b_k, x_k, y_k], k ∈ [1, K]；

Group center is upset in neighborhood, cluster centre is moved to the minimum position of gradient；

For the center of each group, according to apart from computing rule, it is being centered around in the adjacent area at a group center Distribute best match pixel；

Calculate new group center and residual error E (center in the past and present centre distance)；

3,4 operations are repeated, are restrained until residual error E is less than threshold values；

Force connective.

The schematic diagram of SLIC partitioning algorithms shown in Fig. 4, as shown in figure 4, segmentation caused by SLIC algorithms, size are equal It is even, the number of block is may specify, more flexibly.

(2) histograms of oriented gradients (HOG) character representation, including：

Histograms of oriented gradients (Histogram of Oriented Gradient, HOG) is characterized in one kind in computer It is used for carrying out the Feature Descriptor of object detection in vision and image procossing.The method has used the gradient direction of image in itself special Sign, similar to edge orientation histogram method, SIFT descriptions and context Method On Shape, but it is characterized in that it in a net Calculated on the unified pane location of the intensive size of lattice, and overlapping local contrast normalizing has been used in order to improve accuracy The method of change.

The core concept of histograms of oriented gradients feature is that the presentation of the object in piece image and shape can be by gradients Or the direction Density Distribution at edge describes well.Its implementation is first to divide the image into small to be called pane location connected region Domain；Then the gradient direction or edge orientation histogram of each pixel in pane location are gathered；Finally these histograms are combined Get up to can be formed by Feature Descriptor., can also be these local histograms in the bigger of image in order to improve accuracy Contrast normalization is carried out in section (block), the method is close in this section (block) by first calculating each histogram Degree, is then normalized according to this density value to each pane location in section., can be to illumination after this normalization Change and shade obtain more preferable stability.

Fig. 5 show complete HOG feature extraction algorithms and procedure chart, and Fig. 6 show rectangular element cell schematic diagram, As shown in figs.5 and 6,

Core procedure in HOG extraction process can letter include：

1) gradient in image abscissa and ordinate direction is calculated, and calculates the gradient direction of each location of pixels accordingly Value；

2) image is evenly dividing fritter according to locus, i.e. cell in Fig. 6, according to set quantization in cell The histogram of canonical statistics gradient direction, characteristic vector corresponding to the cell is obtained, then again institute in a block in Fig. 6 The characteristic vector for having cell, which is together in series, just obtains the HOG features of the block.

HOG features are to connect to obtain by the histograms of oriented gradients of multiple rectangle cell in a rectangle block, actual Block and cell can not be rectangle in, and only rectangle is more common one kind, and this HOG features can be referred to as Rectangle HOG, circular HOG and center ring actually also be present around HOG, what it is according to the consideration of project application demand is all rectangle.

Fig. 7 show the schematic diagram of HOG 9 direction blocks, and Fig. 8 show ship, sky, the HOG feature schematic diagrames of ocean, As shown in Figure 7 and Figure 8, for rectangle HOG using the rectangle cell grids for repeating traversal, Block blocks are to repeat to travel through by intensive Cell compositions, characteristic vector normalization is independently carried out in each block, to reduce the influence of illumination.General each block is by m × m cell grids composition, and cell is made up of n × n pixel.And each cell gradient direction is divided into z direction block, Projection is weighted to z direction using the gradient direction in cell and amplitude, last each cell produce the features of z dimensions to Amount.The HOG that Dalal etc. is used for human testing chooses z=9, i.e., is divided into 9 direction blocks by 360 degree, is subsequently used in direction gradient Projected.

Each cell just obtains the characteristic vector of a z dimension after being projected by the division of upper figure, the institute in block blocks There is characteristic vector corresponding to cell to be together in series and just constitute HOG features.

By contrasting textural characteristics and HOG features, it can be seen that ship, sky and the textural characteristics of ocean difference are little, i.e., Human eye is set to go to see, it is also difficult to distinguish, so using HOG features.

(3) training grader realizes target Primary Location

After each extracted region feature in training data, suitable grader is trained to complete the classification in region, With realize exclude sea, sky bulk region, navigate to the target of area-of-interest.

SVMs (Support Vector Machine, SVM) performance in the grader for have supervision protrudes, can be with Effectively solve the Machine Learning Problems under small sample, effectively solve the problems, such as evolvement problem and Nonlinear Classification, its main thought It may be summarized to be at 2 points：

1) it be linear can a point situation analyzed, in the case of linearly inseparable, reflected by using non-linear Penetrating algorithm the sample of low-dimensional input space linearly inseparable is converted into high-dimensional feature space makes its linear separability, so that high Dimensional feature space carries out linear analysis to the nonlinear characteristic of sample using linear algorithm and is possibly realized；

2) it is based on structural risk minimization theory the construction optimum segmentation hyperplane in feature space so that classification Device obtains global optimization, and meets certain upper bound in the expected risk of whole sample space with some probability.

Fig. 9 show at target positioning exemplary plot, as shown in figure 9, with reference to SVM and the feature of extraction, it is available can See optical target Primary Location result figure.

S5, Ship Target rough sort, including：

This item purpose final goal is desirable to detect and recognize from visible images the model of military boats and ships, but ring Some disturbing factors are had in border, such as civilian ship, reef, island, so must also be done these when being identified Disturb classification to be screened, to reduce the false alarm rate of identification.

If each model military boats and ships and interference classification (such as various civilian ships, beacon, island) together classified, So the between class distance of feature space can be different.Between class distance between military boats and ships, civilian ship, jamming target can be larger, And can be smaller as the between class distance between each model military boats and ships under same major class, these classes are mixed into progress Classification and Identification, the accuracy of all kinds of divisions in feature space is influenced whether, that is, have influence on the accuracy rate of final classification.

Figure 10 show the flow chart of visible optical target rough sort, as shown in Figure 10, considers, it is necessary to located Visible optical target carry out rough sort, complete the identification of major class, carry out the division of fine classification again on this basis.

For the feature extraction of visible ray/infrared target rough sort, a complete handling process is devised, and compare Two kinds of technical schemes：

1) character representation is used as using multiple dimensioned dense-SIFT and visual dictionary construction, SVM classifier is as modeling work Tool；

2) it is used as and builds as character representation, Linear SVM grader using multiple dimensioned dense-SIFT and Fisher Vector Die worker has；

(1) SIFT descriptions

Figure 11 show SIFT and describes subcharacter extraction and description flow chart, as shown in figure 11, describes son using SIFT and enters The extraction and description of row characteristics of image include following basic step：

1) build metric space and carry out extremum extracting；

2) key point is accurately positioned；

3) direction of key point is determined；

4) feature point description is generated.

The idiographic flow of brief description SIFT algorithms：

1) generation of metric space；

2) spatial extrema point detects；

3) extreme point position is determined；

4) removal of skirt response；

5) key point direction is distributed；

6) feature point description generation.

(2) multiple dimensioned dense-SIFT descriptions

SIFT description are extracted and described just for the invariant feature point of image, therefore it certainly exists information and lost Become estranged omission the problem of.The period of the day from 11 p.m. to 1 a.m is being described to image application SIFT, it is necessary to seek form of the image with more standard, such as image Size is sufficiently large, and critical object proportion is sufficiently large, so just can guarantee that enough characteristic points are detected and then used Subsequent match.During characteristic point is detected and is described, it is not difficult to find out, its complexity is very high, it is necessary to consume substantial amounts of The time is calculated, this is also the one side unfavorable to image recognition and calssification task.Carried when constructing Fisher Vector in progress feature After taking link, clustering method is applied to generate code book, so if can not provide what is enriched enough in feature extraction step Information, then the representativeness of the code book of generation can be directly affected, and then influence the follow-up classification degree of accuracy.Therefore, the present invention is comprehensive Close on the basis of considering above-mentioned factor, employ a kind of improved multiple dimensioned dense-SIFT feature extracting methods.

The method that dense-SIFT description take uniform sampling, feature is carried out to image by same pixel interval and taken out Take, its sampling density is controlled by parameter " step-length ", and here, the present invention is represented with step.So just obtain very intensive Characteristic point, therefore can ensure to better profit from the abundant information of image.Because it need not be by each key point and its Neighborhood and levels are compared to judge extreme point with 26 points of the band of position, therefore have been greatly reduced the complexity calculated Degree.It is the yardstick S of each key point distributing uniform after the sampling interval of base step sizes carries out the extraction of characteristic point, this In yardstick unified setting can be carried out according to actual conditions, avoid calculate yardstick large amount of complex computing.

When key point is described, in order to ensure the rotational invariance of feature, adjusted first to 0 °, then with Key point is the center of circle, constructs border circular areas as radius using pre-assigned unified yardstick S, will fall the pixel in the border circular areas Be divided into 4 × 4 nonoverlapping subregions, each region calculate 8 directions (0 °, 45 °, 90 °, 135 °, 180 °, 225 °, 270 °, 315 °) gradient accumulated value.Different from SIFT, SIFT is tired out using the weighting that gradient is carried out using Gauss function Product, and in dense-SIFT, the present invention replaces Gauss function with rectangular window, and crucial neighborhood of a point is carried out uniformly to add Power, rather than Gauss weighting, after gradient accumulation is completed to key point, carried out using the Gaussian function average of unit where it Weighting.This approximation method both improves speed, in turn ensure that performance is not suffered a loss.Each characteristic area is still with 128 dimensions Vector represents.

Because multiple dimensioned dense-SIFT is using the uniform method for extracting key point, therefore the Scale invariant of its feature Property performance can by certain destruction, in order to ensure scale invariability, the present invention using multiple dimensioned extraction method, to each Key point is extracted and described with 4 different yardsticks.Large scale represents the general picture feature of image, and small yardstick represents figure The minutia of picture.The feature so obtained equally can guarantee that the consistency of yardstick.

Figure 12 show the feature extraction result figure of SIFT description, and Figure 13 show the feature of dense-SIFT description Result figure is extracted, Figure 14 show the feature extraction figure of multiple dimensioned dense-SIFT descriptions, as shown in Figure 12 to Figure 14,

(3) the Fisher Vector constructions of target rough sort

In image classification problem, the method for most representational description piece image is to image zooming-out local feature (SIFT), then these local features are encoded into the global characteristics expression of a high dimension vector, the most the width image.It is the most frequently used Coding techniques be to quantify local feature set to time on the visual dictionary of an off-line training, occurred according to each feature Number is characteristic vector of the word frequency by iamge description into regular length, and visual dictionary can be mixed by k averages (k-means) or Gauss The clustering methods such as matched moulds type (GMM) training obtain, this coding techniques be referred to as bag of words (Bag-of-visual-words, BOV).But bag of words are only a kind of organizational form of discrete disorder feature point set, and quantizing process damages, Therefore the information that can be described is not comprehensive.

This project uses the coding techniques based on Fisher cores (Fisher Kernel), and the vector for encoding generation is Fisher Vector.Image overall feature representation based on Fisher Vector can be regarded as bag of words (BOV) expansion Exhibition, the two is all based on a kind of intermediate representation of image, and Fisher Vector are due to having merged discriminate and production model Advantage, not only embodies the frequency of occurrences of each vision word, also encodes and is distributed on local feature in vision word Different information, therefore Fisher Vector can characterize more rich characteristics of image, and its higher-dimension characteristic than bag of words So that combined with simply and effectively linear classifier with regard to good effect can be obtained.

Image Fisher Vector constitution steps are as described below：

If I=(x₁... ..., x_n) be one group of D dimension characteristic vector, such as the SIFT extracted in image description.If Θ =(μ_k, ∑_k, π_k：K=1 ... ..., K) be a mixed Gauss model GMM parameter, be fitted description son distribution.GMM moulds Type is to each characteristic vector x_iWith each mixed state in GMM, an intensity (posterior probability) is assigned：

For each mode k, average and variance vectors are considered：

Wherein j=1,2 ... ..., D are the dimension subscripts of characteristic vector.Each mode in gauss hybrid models u_kAnd v_kIt is together in series, the Fisher Vector for just constituting image are represented：

If the length of local feature description's is D, the quantity of mixed Gauss model (clusters for K with Gaussian Mixture Algorithm produce one be K code book), vision bag of words (BOV) can obtain K dimension histogram vectors, and construct Fisher Vector is (2D+1) K-1 dimension histogram vectors.Therefore an equal amount of sample, Fisher Vector dimension are general remote high In BOV, so to generate similarly sized histogram vectors, Fisher Vector need smaller code book, i.e., smaller meter It is counted as this.

The present invention is improved Fisher Vector, employs a kind of image based on space Fisher Vector Sorting technique.Implementation step is summarised as：

1) the multiple dimensioned dense-SIFT scale invariant features converting characteristic point of all images in image library is extracted.

2) in the feature space of points of image, characteristic point is clustered using Gaussian Mixture clustering algorithm, obtains code book.

3) gradient vector and Ke Liesiji components are utilized, generates the FisherVector of each image.

4) 2 × 2 Spacial domain decompositions are carried out to each image, counts the feature point number and coordinate of each block of cells respectively.

5) the Fisher Vector of each block of cells, the space Fisher Vector of splicing generation each image are utilized.

After Fisher Vector are constructed using the above method, each image will be by a Fisher Vector table Show, be exactly to train grader in next step, image object is classified, to complete the rough sort target of visible optical target.

S6, the identification of Ship Target disaggregated classification include：

The ship images for splitting to obtain by carrying out Ship Target image, it is also necessary to carry out the classification of warship and civilian boat. Image classification problem is by the analysis to image, and image is incorporated into as a certain kind in several classifications, is mainly emphasized to figure As overall semanteme is judged.For warship and civilian boat classification task, carried out using ResNet depth convolutional neural networks real It is existing.

(1) the depth CNN naval vessels Classification and Identification framework negative correlation learning based on the negatively correlated integrated study of selectivity is a kind of The method for training Artificial neural network ensemble, note training set is { (x₁, t₁) ... ..., (x_N, t_N), wherein x_iIt is to input, t_iIt is that target is defeated Go out, N is the sample number of training set.In negative correlation learning, the integrated output of deep neural network is defined as：

Wherein m be neutral net number, F_i(n) it is i-th of deep neural network in n-th of sample Output in sheet, and F (n) is then the output that deep neural network is integrated on n-th of sample.The ResNet that this project uses is deep Degree neutral net is used for integrated study, so as to reach the incremental learning to image data set of newly arriving.Figure 15 is shown based on selection The depth ResNet integrated study frame diagrams of property negative correlation learning, as shown in figure 15, mainly include four modules：Initial R esNet Deep neural network is integrated, current ResNet deep neural networks are integrated, it is integrated and new to replicate ResNet deep neural networks ResNet deep neural networks integrate.

As shown in figure 15, the interactive relation between each module is：First with first warship and civilian boat image data set It is integrated to Initial R esNet deep neural networks be trained to obtain current ResNet deep neural networks integrate；Then, ought Preceding ResNet deep neural networks are integrated to be replicated to obtain and replicates ResNet deep neural networks and integrate；Next, when second When individual warship and civilian boat image data set arrives, using the data set, to replicating, ResNet deep neural networks are integrated to instruct Practice, obtain new ResNet deep neural networks and integrate；Finally, in order that ResNet deep neural networks are integrated can be new in study The knowledge learnt can largely be retained while view data knowledge, while keep the size of Deep integrating model It is constant, it is therefore desirable to be selected using selective negative correlation learning method, the selection algorithm used herein arrived and negative correlation Learning algorithm will be illustrated in model training part.

(2) model training

The integrated training of ResNet deep neural networks is related to selection algorithm and Negative Correlation Training Method.Herein below will Negative Correlation Training Method and the negatively correlated integrated learning approach of selectivity are specifically described.

Negative Correlation Training Method uses back-propagation algorithm, each neutral net individual of parallel training.Negative correlation learning It can to have differences between neutral net individual, while in turn ensure that each individual performance.Why is negative correlation learning There can be the characteristics of such, mainly in the design of its error function.The error function of negative correlation learning is by mean square error and one Individual penalty term is formed.When n-th of sample is used for training, the error function of i-th of neutral net is defined as：

Wherein p_i(n) it is penalty term, λ ∈ [0,1] are a control parameter, are referred to as punishing Penalty parameter, for controlling one between mean square error (having influence on individual precision) and penalty term (having influence on the difference between individual) Balance.Penalty term p_i(n) may be defined as：p_i(n)=(F_i(n)-F(n))∑_j≠iF_j(n)-F(n).As can be seen that penalty term causes I-th neutral net with it is integrated in remaining neutral net it is negatively correlated, so as to reach between individual neutral net have it is discrepant Purpose.Simultaneously it can also be seen that as λ=0, obtained Artificial neural network ensemble is equivalent to one group of neutral net of stand-alone training；And With λ continuous increase, the emphasis of training will gradually be adjusted to have in otherness between order individual.

Figure 16 show Artificial neural network ensemble selection algorithm block diagram, and as shown in figure 16, selective algorithm is mainly used in figure 15 selection link, there are two purposes：One allows in integrated study, and not all individual is all beneficial to improve collection Into the Generalization Capability of model；Two are so that Artificial neural network ensemble will not be excessively huge.The selectivity used in this project is negatively correlated Integrated learning approach is screened using genetic algorithm to ResNet deep neural networks are integrated.The whole process of this method is such as (assuming that each Sub Data Set is S1, S2 ..., ST) shown in lower：

A) the Artificial neural network ensemble ens1 that a size is m is initialized；

B) ens1 is trained by negative correlation learning with S1；

C) from second Sub Data Set start to finish, i.e. t=2 ..., T, following cyclic process is performed：Replicate current god Through system integrating enst-1, a copy is obtained, is designated as enscopy；Enscopy is trained by negative correlation learning with St；Merge Enst-1 and enscopy, enscomb is integrated into after note merging；Selection course is used to enscomb, selects m neutral net Form integrated enst；

D) final Artificial neural network ensemble ensT is exported.

For selection course in c), this project is selected enscomb using genetic algorithm.It is negative according to as above selectivity Related integrated learning approach understands that selection course selects m neutral net to form enst from the enscomb that size is 2m. Therefore, selection course can turn to the optimization problem of following belt restraining in the form of：

Wherein, J (ω) is predefined object function, and its quality designed will be directly connected to the nerve net chosen The integrated Generalization Capability of network.ω_iIt is the binary vector of 2m dimension, ω_i=1 i-th of neutral net of expression is selected for structure Into enst, ω_i=0 represents that i-th of neutral net is deleted.Because this project is only classified and ResNet to warship and civilian boat Depth convolutional neural networks also have stronger classification capacity, so m can take a relatively small value, such as m=5.

For the optimization problem of as above belt restraining, can be solved using genetic algorithm (GA).Based on genetic algorithm Integrated selection algorithm steps are as follows：

A) initialization algorithm is set：Integrated size m, GA Population Size pop_size, the Probability p of intersection_c, variation it is general Rate p_mAnd fitness function J (ω)；

B) initial population being made up of ω is randomly generated；

C) repeat the steps of, until reaching defined number：Assess the adaptation of each individual (i.e. ω) in current population Degree；Intersect using roulette selection parent individuality, and with single-point with Probability p_cWith variation with Probability p_mProduce offspring；With greedy plan Each individual ω is slightly repaired, if ∑_iω_i＞ m, by the individual ω that one of them is 1_iIt is changed to 0 so that J (ω) is minimum, and repeats The process is until ∑_iω_i=m, if ∑_iω_i＜ m, by the individual ω that one of them is 0_iIt is changed to 1 so that J (ω) is minimum, lays equal stress on The multiple process is until ∑_iω_i=m；

D) wopt is exported as optimal solution, while neutral net is final to form according to corresponding to wopt selects wopt=1 Artificial neural network ensemble.

Described above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For member, without departing from the technical principles of the invention, some improvement and deformation can also be made, these are improved and deformation Also it should be regarded as protection scope of the present invention.

Claims (10)

1. A kind of 1. ship seakeeping method based on multilayer convolutional neural networks, it is characterised in that including：

   S1, naval vessel Sample Storehouse is built using existing image, parameter and model data, and improved by detecting target data collection Naval vessel Sample Storehouse；

   S2, under the framework of convolutional neural networks, by the recognition training to naval vessel Sample Storehouse, form visible ray/infrared and two The naval vessel feature knowledge storehouse of dimension/three-dimensional fusion, for carrying out Ship target recognition identification；

   Including：Establish the Ship Target features training framework based on depth convolutional neural networks, including feature share CNN layers, can Finely tune CNN layers, extracted region model, ROI ponds layer and classification and return full articulamentum；

   Sharing feature CNN layers are the parts that existing depth convolutional neural networks disaggregated model removes last full articulamentum, are being entered Before the training of row Ship Target Detection identification model, instructed in advance on VOC categorized data sets or ImageNet categorized data sets Practice the ginseng for improve the extractability to characteristics of image, then recycling this disaggregated model trained to remove full connection layer segment It is several that parameter initialization is carried out to sharing feature CNN layers；

   Fine-tuning CNN layers are used for online incremental learning；

   Extracted region model is convolutional neural networks, after shared CNN layers and fine-tuning CNN layers, construct a convolutional layer and The neutral net of two full articulamentums arranged side by side, it may be deposited by shared CNN layers feature and fine-tuning CNN layers feature to realize Extraction in Ship Target region to be checked；

   ROI ponds layer, for behind ROI ponds, exporting the characteristic vector of fixed dimension；

   The training of Ship Target Detection identification model includes 4 step sections；First stage：Use point that pre-training is carried out on Sample Storehouse Class model initializes to sharing feature CNN layer parameters, while can be micro- using the Gauss normal distribution pair that zero-mean variance is σ Adjust CNN layers and extracted region model to carry out the parameter initialization of weight and bias term, finally recycle visible ray or infrared instruction Practice collection view data to be finely adjusted extracted region neural network model；Second stage：Use extracted region mould in the first stage Extracted region training sharing feature CNN layers, the fine-tuning CNN layers of type extraction, while using in Voc2012 or ImageNet data The disaggregated model initialization sharing feature CNN layer parameters of pre-training are carried out on collection；Phase III：Trained using second stage Sharing feature CNN layers, fine-tuning CNN layer parameters, reinitialize sharing feature CNN layers, the fine-tuning CNN of extracted region model Layer parameter, then fix sharing feature CNN layers, fine-tuning CNN layer parameters and extracted region neural network model is finely adjusted； Fourth stage：The convolution layer parameter of extracted region neural network model in phase III is freezed and extracts extracted region, then Sharing feature CNN layers, fine-tuning CNN layer parameters are finely adjusted；

   S3, visible ray or IR video stream to ShipTargets carry out the collection of Real-time High Resolution rate；

   S4, ShipTargets are detected；

   S5, Ship Target image rough sort are directed to simply is classified to the major class on naval vessel, reduces follow-up Ship Target subdivision The workload of class identification；

   The disaggregated classification that S6, the deep neural network model completed based on Ship Target features training carry out Ship Target identifies work Make, accurately identify the type on naval vessel.
2. 2. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 1, it is characterised in that S4, Ship Target Detection specifically includes：Calculated using SLIC image segmentation algorithms, SLIC algorithms concretely comprise the following steps：

   K super-pixel point number is set to initialize the center C of group_k=[l_k, a_k, b_k, x_k, y_k], k ∈ [1, K]；

   Group center is upset in neighborhood, cluster centre is moved to the minimum position of gradient；

   For the center of each group, according to apart from computing rule, distributed being centered around in the adjacent area at a group center Best match pixel；

   Calculate new group center and residual error；

   Restrained until residual error E is less than threshold values.
3. 3. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 1, it is characterised in that in S4 Image gradient calculate include：

   G (x, y)=| | I (x+1, y)-I (x-1, y) | |²+ | | I (x, y+1)-I (x, y-1) | |²(1)；

   Wherein I (x, y) is lab vectors, represents the color vector of pixel space position (x, y), | | | | it is L2 norms；

   Step-length S=sqrt (N/K) is set, N is sum of all pixels, and K is super-pixel number, in the 2S*2S of cluster centre neighborhood Match point is distributed for each cluster centre, is calculated according to the setpoint distance of each point and cluster centre, each pixel in image Point finally connects with nearest cluster centre, and is covered by the region of search of the cluster centre, in all pixel quilts After being associated in nearest group center, new center will be computed, and be the flat of all lab xy vectors for belonging to the cluster Average, repeat with nearest group's core contextual pixel and recalculate this process of group center until convergence.
4. 4. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 1, it is characterised in that S4 In, the core in HOG extraction process includes：

   1) gradient in image abscissa and ordinate direction is calculated, and calculates the gradient direction value of each location of pixels accordingly；

   2) image is evenly dividing fritter according to locus, according to set quantitative criteria statistical gradient direction in fritter Histogram, characteristic vector corresponding to fritter is obtained, then the characteristic vector of all cell in a bulk is together in series just again The HOG features of bulk are obtained, wherein, each bulk is made up of m × m fritter, and fritter is made up of n × n pixel.
5. 5. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 4, it is characterised in that right After each extracted region feature in training data, suitable grader is trained to complete the classification in region, and sea is excluded to realize Face, the bulk region of sky, navigate to the target of area-of-interest.
6. 6. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 4, it is characterised in that S6 bags Include：The identification of warship and civilian boat；

   Note training set is { (x₁, t₁) ... ..., (x_N, t_N), wherein x_iIt is to input, t_iIt is target output, N is the sample of training set Number, the integrated output of deep neural network are defined as：Wherein m be neutral net number, F_i(n) It is output of i-th of deep neural network on n-th of sample, and F (n) is then deep neural network is integrated in n-th of sample On output, the ResNet deep neural networks used are used for integrated study, so as to reach the increment to image data set of newly arriving Study, including four modules：Initial R esNet deep neural networks are integrated, current ResNet deep neural networks are integrated, replicate The integrated and new ResNet deep neural networks of ResNet deep neural networks integrate；

   It is trained first with first warship and civilian boat view data set pair Initial R esNet deep neural networks are integrated Integrated to current ResNet deep neural networks；Then, integrated replicated of current ResNet deep neural networks is answered ResNet deep neural networks processed integrate；Next, when second warship and civilian boat image data set arrive, the data are utilized Set pair duplication ResNet deep neural networks are integrated to be trained, and is obtained new ResNet deep neural networks and is integrated；Finally, apply Selective negative correlation learning method is selected.
7. 7. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 5, it is characterised in that S6 enters One step includes：

   When n-th of sample is used for training, the error function of i-th of neutral net is defined as：

   Wherein p_i(n) it is penalty term, λ ∈ [0,1] are a control parameter, are referred to as punishing Penalty parameter, for controlling one between mean square error and penalty term to balance, penalty term p_i(n) may be defined as：p_i(n)=(F_i (n)-F(n))∑_j≠iF_j(n) remaining neutral net during-F (n) penalty terms cause i-th of neutral net and integrated is negatively correlated, from And there is difference between reaching individual neutral net, as λ=0, obtained Artificial neural network ensemble is equivalent to one group of stand-alone training Neutral net；And as λ continuous increase, the emphasis of training will gradually be adjusted to have in otherness between order individual.
8. 8. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 5, it is characterised in that S6 enters One step includes：

   A) the Artificial neural network ensemble ens1 that a size is m is initialized；

   B) ens1 is trained by negative correlation learning with S1；

   C) from second Sub Data Set start to finish, i.e. t=2 ..., T, following cyclic process is performed：Replicate Current neural net Network integrates enst-1, obtains a copy, is designated as enscopy；Enscopy is trained by negative correlation learning with St；Merge enst- 1 and enscopy, enscomb is integrated into after note merging；Selection course is used to enscomb, m neutral net is selected and forms Integrated enst；

   D) final Artificial neural network ensemble ensT is exported；

   For selection course in c), enscomb is selected using genetic algorithm, integrated learned according to as above selectivity is negatively correlated Learning method understands that selection course selects m neutral net to form enst from the enscomb that size is 2m, and selection course can The optimization problem of following belt restraining is turned in the form of：

   <mrow> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <mi>min</mi> <mi> </mi> <mi>J</mi> <mrow> <mo>(</mo> <mi>&amp;omega;</mi> <mo>)</mo> </mrow> <mo>,</mo> <mi>w</mi> <mo>=</mo> <mo>&amp;lsqb;</mo> <msub> <mi>&amp;omega;</mi> <mn>1</mn> </msub> <msub> <mi>&amp;omega;</mi> <mn>2</mn> </msub> <mo>...</mo> <msub> <mi>&amp;omega;</mi> <mrow> <mn>2</mn> <mi>m</mi> </mrow> </msub> <mo>&amp;rsqb;</mo> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mi>s</mi> <mo>.</mo> <mi>t</mi> <mo>.</mo> <msub> <mi>&amp;omega;</mi> <mi>i</mi> </msub> <mo>&amp;Element;</mo> <mo>{</mo> <mn>0</mn> <mo>,</mo> <mn>1</mn> <mo>}</mo> <mo>,</mo> <mo>&amp;ForAll;</mo> <mi>i</mi> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>&amp;Sigma;</mi> <mi>i</mi> </msub> <msub> <mi>&amp;omega;</mi> <mi>i</mi> </msub> <mo>=</mo> <mi>m</mi> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>3</mn> <mo>)</mo> </mrow> <mo>;</mo> </mrow>

   Wherein, J (ω) is predefined object function, and its quality designed will be directly connected to the neutral net collection chosen Into Generalization Capability.ω_iIt is the binary vector of 2m dimension, ω_i=1 i-th of neutral net of expression is selected for forming Enst, ω_i=0 represents that i-th of neutral net is deleted；

   Carrying out solution using genetic algorithm includes：

   A) initialization algorithm is set：Integrated size m, GA Population Size pop_size, the Probability p of intersection_c, the Probability p of variation_m And fitness function J (ω)；

   B) initial population being made up of ω is randomly generated；

   C) repeat the steps of, until reaching defined number：Assess the fitness of each individual in current population；Use wheel disc Gambling selection parent individuality, and intersected with single-point with Probability p_cWith variation with Probability p_mProduce offspring；Repaired with Greedy strategy per each and every one Body ω, if ∑_iω_i＞ m, by the individual ω that one of them is 1_iIt is changed to 0 so that J (ω) is minimum, and repeats the process until ∑_i ω_i=m, if ∑_iω_i＜ m, by the individual ω that one of them is 0_iBe changed to 1 so that J (ω) is minimum, and repeat the process until ∑_iω_i=m；

   D) wopt is exported as optimal solution, while the neutral net according to corresponding to wopt selections wopt=1 forms final god Through system integrating.
9. 9. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 1, it is characterised in that bag Include：Fine-tuning CNN layers are used for online incremental learning, when carrying out Ship Target Detection online, if the naval vessel mesh being detected Mark meets mark condition, then needs to carry out target detection model online incremental learning, needed during online incremental learning by Sharing feature CNN layers, extracted region model and the follow-up ROI ponds layer of fine-tuning CNN layers, full connection layer parameter are consolidated It is fixed, small parameter perturbations are then carried out to fine-tuning CNN layers by error in classification.
10. 10. the ship seakeeping method based on multilayer convolutional neural networks as claimed in claim 1, it is characterised in that S2 Also include：

    In the model reasoning stage, if the naval vessel identified using Ship Target Detection identification model meets that sample marks condition, Then need to carry out online incremental learning to the fine-tuning CNN layers of Ship Target Detection identification model using the sample, including：First Step：The ship seakeeping threshold value predicted using target detection identification model is judged sample, if the threshold value is higher than σ₁Or less than σ₂Online incremental learning need not then be carried out；, whereas if the threshold value is less than σ₁And it is higher than σ₂Then need to carry out once Online incremental learning；Second step：Utilize fine-tuning CNN layer of the view data to target detection model for meeting sample mark condition Carry out primary parameter fine setting.