CN110516605A - Any direction Ship Target Detection method based on cascade neural network - Google Patents
Any direction Ship Target Detection method based on cascade neural network Download PDFInfo
- Publication number
- CN110516605A CN110516605A CN201910801199.0A CN201910801199A CN110516605A CN 110516605 A CN110516605 A CN 110516605A CN 201910801199 A CN201910801199 A CN 201910801199A CN 110516605 A CN110516605 A CN 110516605A
- Authority
- CN
- China
- Prior art keywords
- network
- ship target
- target detection
- neural network
- detection
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Any direction Ship Target Detection method based on cascade neural network that the invention discloses a kind of, comprising: satellite image is input to first area screening network, obtains Ship Target candidate region image;Ship Target candidate region image is input in preparatory trained cascade neural network naval vessel detection model and carries out target detection, obtains preliminary aim testing result;Duplicate removal is carried out to preliminary aim testing result using NMS method, obtains object detection results;Wherein, cascade neural network naval vessel detection model includes successively cascade second area screening network, network connectivity layer and target detection network.Any direction Ship Target Detection method provided by the invention based on cascade neural network, can effectively improve Ship Target Detection accuracy.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of appointing based on cascade neural network
Meaning direction Ship Target Detection method.
Background technique
In recent years, with the progress of aeronautical and space technology, the means that remote sensing image obtains are increasingly mature, the resolution of image
Rate, including temporal resolution, spatial resolution, radiometric resolution and spectral resolution are being continuously improved.Currently, remote sensing is
The bottleneck for breaking through data acquisition, is moving towards the new stage of overall application, has established data base for the extraction of ocean offshore target
Plinth.Naval vessel is the highest priority of marine monitoring and wartime strike as important targets in ocean, effective in real time to obtain naval vessel
Essential information, suffer from huge meaning in civilian and military field.In civil field, vessel in distress rescue, strike are assisted
The illegal activities such as smuggling, illegal dumping greasy dirt, illegal fishing and pirate, monitor the shipping vessels in specific harbour or sea area etc. all
Need to obtain naval vessel information;In military field, is detected, monitored and identified by the naval vessel to emphasis harbour and sea area, really
The important informations such as the model, type, position on naval vessel are determined, convenient for the analysis of naval battle field environmental situation, to grasp the sea of other side
Operating strength assesses wartime sea strike effect, forms naval warfare information etc., provides foundation for naval battle field decision support.
The naval vessel detection of early stage mainly uses SAR image, and comparative maturity, it is seen that the naval vessel mesh of light remote sensing images
Mark research is later, and related data is also less.And in optical imagery, the naval vessel detection under marine background also has been extensively studied,
Inshore ship detection is started late with respect to marine vessel detection.
Currently, the method for naval vessel detection is broadly divided into two major classes: conventional method and the method based on deep learning.Tradition
Satellite image object detection method mainly use multi-step strategy from coarse to fine, generally comprise Yunnan snub-nosed monkey, Hai Lufen
It cuts, Region Feature Extraction, target-recognition, conventional method needs artificial design features extracting method, adaptability is poor,
Cause testing result inaccurate.
Depth learning technology trains learning ability that it is examined in target due to its powerful character representation and end to end
Survey is widely used with identification field, and greatly improves detection performance.Object detection method based on deep learning is first
It is that candidate region is extracted from image, then using convolutional neural networks (Convolution Neural Network, CNN) etc.
Deep neural network identify to candidate region and bounding box returns, and realizes object detection and recognition.
The deficiency of existing Ship Detection specifically includes that the Calculation bottleneck of CNN essentially consists in the full articulamentum of higher-dimension,
Not only parameter is more, and computation complexity is high, is easy to cause over-fitting, and needs an equal amount of image input, and common target
Detection algorithm uses general convolution kernel, and aspect is more poor than target detection effect, causes testing result accuracy poor.
Therefore, how to provide a kind of method that Ship Target Detection result accuracy is high is that those skilled in the art need
It solves the problems, such as.
Summary of the invention
In view of this, the present invention provides a kind of any direction Ship Target Detection method based on cascade neural network,
Ship Target Detection accuracy can be effectively improved.
To achieve the goals above, the present invention adopts the following technical scheme:
A kind of any direction Ship Target Detection method based on cascade neural network, comprising:
Satellite image is input to first area screening network, obtains Ship Target candidate region image;
By Ship Target candidate region image be input in preparatory trained cascade neural network naval vessel detection model into
Row target detection obtains preliminary aim testing result;
Duplicate removal is carried out to preliminary aim testing result using NMS method, obtains object detection results;
Wherein, cascade neural network naval vessel detection model includes successively cascade second area screening network, network connection
Layer and target detection network.
Preferably, the second area screening network and the multiple dimensioned convolutional layer of target detection Web vector graphic are examined
It surveys, is all made of VGG16 as backbone network, retains conv1_1 to conv5_3, last two layers full articulamentum is replaced with into convolution
Layer, in addition increases conv6_1, conv6_2, conv7_1, conv7_2, conv8_1 and conv8_2 totally 6 convolutional layers and one
The convolutional layer for detection of the long convolution karyogenesis of a 3x5;And second area screening network and target detection network are corresponding each
It is attached respectively by network connectivity layer between layer structure, and each network connectivity layer operates with height by deconvolution
The feature of layer network articulamentum.
Preferably, the training method of cascade neural network naval vessel detection model includes:
Training data generation step: the high-resolution satellite image marked in advance is cut into the sample of fixed size
Data;
Data augmentation step: network is generated based on sample data and super-resolution confrontation and generates multiple dimensioned high-resolution shadow
Picture, as training data;Using rotation, overturning and the further exptended sample data of luminance contrast method of adjustment, as training
Data;
Length-width ratio clustering step: carrying out the length and width and length-width ratio clustering of Ship Target to sample data,
Cluster result is obtained, length-width ratio parameter is set according to cluster result;
Training step: being based on length-width ratio parameter and preset default frame vertical shift, and training data is input to cascade mind
Duplicate removal is carried out through being trained in network naval vessel detection model, and using NMS algorithm, is based on loss function in the training process
Convergence judgement is carried out, trained cascade neural network naval vessel detection model is finally obtained.
Preferably, training data generation step specifically includes:
Using deep learning network inputs image size as the size of sliding window, there is the sliding on satellite image of overlapping
It is dynamic;
If including effective Ship Target in current sliding window mouth, the corresponding image cropping of current sliding window mouth is come out,
Meanwhile naval vessel in current window is updated to the coordinate in current sliding window mouth relative to the coordinate of whole picture satellite image, and
Save as the corresponding XML mark file of the image cut out.
Preferably, the method for discrimination of effective Ship Target includes:
Sliding window area and Ship Target quadrilateral area overlapping area are greater than with naval vessel quadrilateral area area ratio
0.5, then it is determined as effective Ship Target.
Preferably, loss function employed in training step are as follows:
Wherein, which default frame i indicates,Indicate the classification with i-th of default matched true frame of frame,Indicate with
I-th default the matched true frame of frame position and size, piIndicate confidence level, xiIt indicates to default in second area screening network
The coordinate of frame, ciIndicate prediction classification, tiIndicate the prediction coordinate information in target detection network;NrpnAnd NodnRespectively indicate
It screens the positive sample in network and target detection network and defaults frame quantity in two regions;LbIndicate two-value Classification Loss, LmIt indicates more
Classification loss, LrIt indicates to return loss,If indicating, the confidence level of negative sample is greater than a threshold value, returns to 1, otherwise returns
Return 0;If Nrpn=0, settingWithIf Nodn=0, then it is arrangedWith
Preferably, multiple dimensioned high resolution image includes: the image data and raw video 4 of 2 times of resolution ratio of raw video
The image data of times resolution ratio.
Preferably, the first area screening network is using the PNet network in MTCNN model.
It can be seen via above technical scheme that compared with prior art, the present disclosure provides be based on cascaded neural net
Any direction Ship Target Detection method of network has following technical advantage:
First, the present invention screens network using a kind of target detection frame end to end, by the region of conventional target detection
A network end to end is formed with target detection cascade, had both maintained two stages (region screening network and target detection
Network is independent) accuracy of object detection method, also maintain a stage (network directly carries out target detection) target detection side
The efficiency of method;
Second, a large amount of Ship Target data are marked, and clustering is carried out according to length-width ratio of the labeled data to naval vessel,
Design is suitable for the length-width ratio parameter of Ship Target Detection;
Third increases default frame density using long convolution kernel method, and in vertical direction, multi-direction preferably to adapt to
Detection realizes that aspect than any direction Ship Target Detection, is predicted four coordinate positions offset of target area, used
Four coordinates indicate Ship Target position more accurately.
In conclusion any direction Ship Target Detection method provided by the invention based on cascade neural network can
Effectively improve Ship Target Detection accuracy.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will to embodiment or
Attached drawing needed to be used in the description of the prior art is briefly described, it should be apparent that, the accompanying drawings in the following description is only
The embodiment of the present invention for those of ordinary skill in the art without creative efforts, can be with
Other attached drawings are obtained according to the attached drawing of offer.
Fig. 1 is the flow chart of any direction Ship Target Detection method provided by the invention based on cascade neural network;
Fig. 2 is the structural schematic diagram of cascade neural network naval vessel provided by the invention detection model;
Fig. 3 is the structural schematic diagram of network connectivity layer provided by the invention;
Fig. 4 is that default frame provided by the invention predicts process schematic;
Fig. 5 is default frame vertical shift schematic diagram provided by the invention;
Fig. 6 is MTCNN model PNet schematic network structure provided by the invention;
Fig. 7 is Ship Target Detection result schematic diagram provided by the invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts it is all its
His embodiment, shall fall within the protection scope of the present invention.
Referring to attached drawing 1, any direction Ship Target inspection based on cascade neural network that the embodiment of the invention discloses a kind of
Survey method, comprising:
Satellite image is input to first area screening network, obtains Ship Target candidate region image;
By Ship Target candidate region image be input in preparatory trained cascade neural network naval vessel detection model into
Row target detection obtains preliminary aim testing result;
Duplicate removal is carried out to preliminary aim testing result using NMS method, obtains object detection results;
Wherein, cascade neural network naval vessel detection model includes successively cascade second area screening network, network connection
Layer and target detection network.
Preferably, the second area screening network and the multiple dimensioned convolutional layer of target detection Web vector graphic are examined
It surveys, is all made of VGG16 as backbone network, retains conv1_1 to conv5_3, last two layers full articulamentum is replaced with into convolution
Layer, in addition increases conv6_1, conv6_2, conv7_1, conv7_2, conv8_1 and conv8_2 totally 6 convolutional layers and one
The convolutional layer for detection of the long convolution karyogenesis of a 3x5;And second area screening network and target detection network are corresponding each
It is attached respectively by network connectivity layer between layer structure, and each network connectivity layer operates with height by deconvolution
The feature of layer network articulamentum.Each convolution layer parameter is as shown in the table.
Preferably, the training method of cascade neural network naval vessel detection model includes:
Training data generation step: the high-resolution satellite image marked in advance is cut into the sample of fixed size
Data;
Data augmentation step: network is generated based on sample data and super-resolution confrontation and generates multiple dimensioned high-resolution shadow
Picture, as training data;Using rotation, overturning and the further exptended sample data of luminance contrast method of adjustment, as training
Data;
The present invention generates network using super-resolution confrontation and generates multiple dimensioned high-definition picture, for expanding trained sample
Notebook data reduces requirement of the model to image resolution ratio, improves model generalization ability.
Length-width ratio clustering step: carrying out the length and width and length-width ratio clustering of Ship Target to sample data,
Cluster result is obtained, length-width ratio parameter is set according to cluster result;
Training step: being based on length-width ratio parameter and preset default frame vertical shift, and training data is input to cascade mind
Duplicate removal is carried out through being trained in network naval vessel detection model, and using NMS algorithm, is based on loss function in the training process
Convergence judgement is carried out, trained cascade neural network naval vessel detection model is finally obtained.
Preferably, training data generation step specifically includes:
Using deep learning network inputs image size as the size of sliding window, there is the sliding on satellite image of overlapping
It is dynamic;
If including effective Ship Target in current sliding window mouth, the corresponding image cropping of current sliding window mouth is come out,
Meanwhile naval vessel in current window is updated to the coordinate in current sliding window mouth relative to the coordinate of whole picture satellite image, and
Save as the corresponding XML mark file of the image cut out.
Preferably, the method for discrimination of effective Ship Target includes:
Sliding window area and Ship Target quadrilateral area overlapping area are greater than with naval vessel quadrilateral area area ratio
0.5, then it is determined as effective Ship Target.
Preferably, loss function employed in training step are as follows:
Wherein, which default frame i indicates,Indicate the classification with i-th of default matched true frame of frame,It indicates
Position and the size of the matched true frame of frame, p are defaulted with i-thiIndicate confidence level, xiIt indicates to write from memory in second area screening network
Recognize the coordinate of frame, ciIndicate prediction classification, tiIndicate the prediction coordinate information in target detection network;NrpnAnd NodnIt respectively indicates
Second area screens the positive sample in network and target detection network and defaults frame quantity;LbIndicate two-value Classification Loss, LmIt indicates
Multi-class loss, LrIt indicates to return loss,If indicating, the confidence level of negative sample is greater than a threshold value, returns to 1, otherwise
Return to 0;If Nrpn=0, settingWithIf Nodn=0, then it is arrangedWith
Preferably, multiple dimensioned high resolution image includes: the image data and raw video 4 of 2 times of resolution ratio of raw video
The image data of times resolution ratio.
Preferably, the first area screening network is using the PNet network in MTCNN model.
Ship Target Detection method provided by the invention devises the length-width ratio parameter for being suitable for Ship Target Detection, adopts
With the detection layers of long convolution kernel, full convolution detection network is constructed, multiscale target detection can be carried out, can be realized any direction
The detection of Ship Target, effectively solving the multi-direction, multiple dimensioned of Ship Target, aspect ratio and the characteristics such as scale is small causes
The not high problem of Detection accuracy.
Technical solution of the present invention is further elaborated below with reference to particular technique details.
1. the mark of satellite image naval vessel data
" four-point method " is used to carry out four sides the Ship Target in high-resolution satellite image using quadrangle annotation tool
Shape mark, and by the target informations marked all in satellite image with the format storage of XML file to local.The letter of storage
Breath includes the information such as the coordinate and ship type of four points of quadrangle.
2. the generation and data augmentation of training data
The generation of 2.1 training datas
According to size needed for deep learning mode input, in conjunction with the XML file of mark, the large scene high score that will have been marked
Resolution satellite image is cut into the sample data of fixed size.The specific method is as follows: by deep learning network inputs image size
As the size of sliding window, there is sliding on substantially satellite image for overlapping.If in current sliding window mouth including effective naval vessel
Target then comes out the corresponding image cropping of current sliding window mouth, meanwhile, by naval vessel in current window relative to whole picture satellite
The coordinate of image is updated to the coordinate in current window, and saves as the corresponding XML mark file of the cutting image.
Naval vessel effective target discriminant approach is as follows: with sliding window area and Ship Target quadrilateral area overlapping area
It is criterion with naval vessel quadrilateral area area ratio, if area ratio is considered as effective Ship Target greater than 0.5.
2.2 data augmentation
Deep learning is the mode of learning of data driven type, in order to meet learning training requirement, improve target detection model
Generalization ability, prevent target detection model overfitting, take following method to training data carry out data augmentation.
(1) the multiple dimensioned high resolution image of network (SRGAN) generation is generated using based on super-resolution confrontation, generates 2
Again, 4 times of resolution images are to expand trained and verify data.
(2) using the methods of rotation, overturning, luminance contrast adjustment, further expand trained and verify data.Wherein,
When verify data refers to trained, for the data of test model accuracy rate.
3. Ship Target length-width ratio clustering
General target detection algorithm length-width ratio parameter setting is simple, length-width ratio biggish Ship Target changeable for direction
For be not appropriate for.Therefore, the present invention by above-mentioned data mark, data generate and etc. obtain Ship Target Detection data
After collection, clustering first is done to the length and width of Ship Target and length-width ratio, is suitable for according to the design of length-width ratio cluster result
The length-width ratio parameter of Ship Target.According to data clusters result by Ship Target Detection length-width ratio parameter setting are as follows: 3:1,5:1,
7:1,9:1,11:1.
4. Ship Target Detection modelling and training
4.1 Ship Target Detection modelling thinkings
The present invention designs Ship Target using a kind of target detection frame end to end, in conjunction with text scene detection thinking
Detection model, the detection network are divided into three parts: second area screens network, network connectivity layer and target detection network,
The second area screening network and target detection cascade of conventional target detection are formed into a network end to end, both protected
The accuracy for having held two stages (second area screens network and target detection network is independent) object detection method, also maintains
The efficiency of one stage (network directly carries out target detection) object detection method, meanwhile, according to the aspect on naval vessel ratio, in many ways
To, Analysis On Multi-scale Features, specific network structure and detection layers are designed, realize the efficiently and accurately detection of Ship Target.Based on grade
The Ship Target Detection network structure for joining neural network is as shown in Figure 2:
(1) position of frame is defaulted in the first adjustment for the first time of second area screening network, is allowed to provide for target detection network and be repaired
Default frame after just provides preferably initialization for subsequent detection and returns.Using VGG16 as backbone network, retain conv1_
1 to conv5_3, last two layers full articulamentum is replaced with convolutional layer, in addition increases by 6 layers of convolutional layer.Choose conv4_3, fc7
(having been replaced with convolutional layer), and the conv6_2 newly increased, conv7_2,5 convolutional layers such as conv8_2, then connect one
The convolutional layer for detection of 3x5 long convolution karyogenesis, exports whether each default frame includes target and rough quadrangle
Location bias information.
(2) network connecting part be by second area screen network in feature be transferred in target detection network to
Future position, size and class label.The output characteristic pattern of second area screening network is converted into target detection net
The input of network.Feature Conversion is carried out using structure as shown in Figure 3.
In order to establish the connection of second area screening network and target detection network, by network connecting part by the
Two regions are screened in the characteristic pattern converting into target detection network in network, and such target detection network can share the secondth area
The feature of domain screening network.Only with default the associated characteristic pattern of frame on using network connection.Network connection is high by addition
Grade feature integrates extensive context, to improve detection accuracy.For matching dimensionality, make high level using deconvolution operation
Characteristic pattern become larger, be added using the mode of Pixel-level, after summation add convolutional layer to ensure the feature for detection
Distinguishability.
(3) target detection network and second area screen network sharing features, and target detection network is sieved using second area
5 layers of characteristic pattern that network selection network generates carry out conversion and are used as input, merge the feature of different layers, further improve and return and predict
Multiclass label.
4.2 Ship Target Detection modellings and realization
(1) conv1_1 to conv5_3 is retained as backbone network using the VGG16 network of 300x300, chooses backbone
The conv4_3 of network, fc7 (have been replaced with convolutional layer), and the conv6_2 newly increased, conv7_2, conv8_2 etc. 5 volumes
For lamination for detecting, each difference detection layers can produce the characteristic pattern of different scale, and the characteristic pattern of generation is sieved for second area
Network selection network and target detection network share.
(2) second area screening network in, by above-mentioned 5 convolutional layers be used to default frame two classify (with/without target) with
And default frame position adjustment;In target detection, 5 convolutional layer features are converted by being connected to the network, are examined as target
The input of survey grid network is returned for multi-class prediction with target exact position.
In order to detect aspect than target, it is used for the convolutional layer of detection using 3x5 long convolution karyogenesis, is rolled up at above-mentioned 5
Lamination be followed by one for detection convolutional layer, for predict output, obtain it is each default frame classification and position believe
Breath, the location information are the offset information of the coordinate of four points of Ship Target.
Detection layers are the cores of network, and default frame is rectangle, and output is quadrangle prediction block, prediction be relative to
Default the offset information of frame.It is as shown in Figure 4 to default frame learning process.Solid white line is true frame, and white dashed line matches
Default frame, white arrow indicates learning process.
Detailed process is as follows:
1) b is assumed initially that0={ x0,y0,w0,h0Indicate that default frame, corresponding quadrangle representation method areWherein, (x0,y0) indicate to default the central point of frame, (w0,h0) indicate default
The width and height of frame, then shown in the calculation method such as formula (1) that quadrangle indicates.
2) detection layers after 5 convolutional layers will predict the Ship Target probability and location bias of each default frame, defeated
It is worth out are as follows: (Δ x, Δ y, Δ w, Δ h, Δ x1,Δy1,Δx2,Δy2,Δx3,Δy3,Δx4,Δy4, c), in the training stage,
True value is calculated in default frame and mark quadrangle, then calculates penalty values by the difference of true value and predicted value.
(3) vertical shift is set, and it is elongated shape that the present invention, which defaults frame, and it is close in the horizontal direction to may cause default frame in this way
Collect sparse in vertical direction, so as to cause detection inaccuracy.Therefore, frame vertical shift is defaulted in setting in vertical direction, makes
It must default that frame is intensive in vertical direction, the only solid white line frame of vertical shift, will not miss many continuous Vertical Squares
To target.After white dashed line frame joined vertical shift, Ship Target information can be all enclosed into, such as Fig. 5 institute
Show.
(4) NMS (non-maxima suppression) algorithm is used, the candidate target region that different characteristic layers predicts is carried out
Duplicate removal obtains the final prediction result of Ship Target position and classification information.Firstly, all testing results are big according to probability
Small sequence, and traversal prediction block from high to low.For each prediction block, remove big with current predictive frame IOU under same category
Testing result in 0.5 other prediction blocks, after obtaining duplicate removal.
4.3 Ship Target Detection loss functions
Shown in Ship Target Detection loss function such as formula (2).
Wherein, which default frame i indicates,Indicate the classification with i-th of default matched true frame of frame,It indicates
Position and the size of the matched true frame of frame, p are defaulted with i-thiIndicate confidence level, xiIt indicates to write from memory in second area screening network
Recognize the coordinate of frame, ciIndicate prediction classification, tiIndicate the prediction coordinate information in target detection network.NrpnAnd NodnIt respectively indicates
Second area screens the positive sample in network and target detection network and defaults frame quantity.LbExpression two-value Classification Loss (with/without
Target), LmIndicate multi-class loss, LrIt indicates to return loss,If indicating, the confidence level of negative sample is greater than a threshold value,
1 is so returned, otherwise returns to 0.If Nrpn=0, settingWithIf Nodn=0, then setting
It setsWith
5. substantially satellite image Ship Target Detection
Substantially the flow chart of satellite image Ship Target Detection is as shown in Figure 1.
The screening of 5.1 candidate regions
Traditional slip window sampling has the sliding of overlapping on substantially satellite image, then using sliding window region as
The input of target detection model carries out target detection, needs to be traversed for whole picture image, and computational efficiency is low, therefore the present invention is using real
When Face datection MTCNN model PNet network, as first area screen network, screen Ship Target candidate region, accelerate
Search speed.Its network structure is as shown in Figure 6.MTCNN model by three different scales small-sized convolution neural network group at,
Respectively PNet, RNet and ONet.PNet is that network is suggested in region, for generating candidate target.PNet is the complete of a shallow-layer
Convolutional network, includes three convolutional layers and a pond layer, and input picture size is 12 pixels × 12 pixels.Full convolutional Neural
Network is free of full articulamentum, can satisfy the image input of arbitrary size, therefore the target of a wide range of remote sensing image may be implemented
Search.PNet is substantially a kind of slip window sampling accelerated using GPU, to each candidate while choosing candidate window
Window carries out classification judgement.
5.2 Ship Target Detections and repetition object removal
(1) candidate region that screening network obtains is screened into as detection model input in region, detect in the region whether
There is Ship Target, if so, four coordinates of prediction target, and the coordinate by Ship Target in the candidate region is mapped to substantially
On satellite image, all candidate regions repeat the process.
(2) after finishing to the detection of all candidate regions, the detection target area obtained on substantially satellite image may
Overlapping is had, therefore unique object detection area in order to obtain, duplicate removal is carried out using NMS algorithm, obtains last target inspection
Survey result.Ship Target Detection testing result schematic diagram is as shown in Figure 7.
Second area is screened network using a kind of target detection frame end to end by technical solution provided by the invention
With target detection network by network connection, cascade forms a network end to end, both maintains two stages target detection side
The accuracy of method also maintains the efficiency of a stage object detection method;Meanwhile a large amount of Ship Targets are marked using four-point method
Data, and clustering is carried out, the length-width ratio of Ship Target is obtained, design is suitable for default frame of the aspect than Ship Target,
Improve Ship Target Detection accuracy;Character detecting method is used for reference, long convolution kernel is designed, increases vertical direction default frame density,
Any direction Ship Target Detection is realized, in addition, indicating prediction regression result, expression that can be more accurate using four-point method
Target position.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with its
The difference of his embodiment, the same or similar parts in each embodiment may refer to each other.For being filled disclosed in embodiment
For setting, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is referring to method portion
It defends oneself bright.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, defined herein
General Principle can realize in other embodiments without departing from the spirit or scope of the present invention.Therefore, originally
Invention is not intended to be limited to the embodiments shown herein, and is to fit to special with principles disclosed herein and novelty
The consistent widest scope of point.
Claims (8)
1. a kind of any direction Ship Target Detection method based on cascade neural network characterized by comprising
Satellite image is input to first area screening network, obtains Ship Target candidate region image;
Ship Target candidate region image is input in preparatory trained cascade neural network naval vessel detection model and carries out mesh
Mark detection, obtains preliminary aim testing result;
Duplicate removal is carried out to preliminary aim testing result using NMS method, obtains object detection results;
Wherein, cascade neural network naval vessel detection model include successively cascade second area screening network, network connectivity layer and
Target detection network.
2. any direction Ship Target Detection method according to claim 1 based on cascade neural network, feature exist
In the second area screening network and the multiple dimensioned convolutional layer of target detection Web vector graphic are detected, and VGG16 is all made of
As backbone network, retains conv1_1 to conv5_3, last two layers full articulamentum is replaced with into convolutional layer, is in addition increased
Conv6_1, conv6_2, conv7_1, conv7_2, conv8_1 and the conv8_2 long paper of totally 6 convolutional layers and a 3x5
The convolutional layer for detection of product karyogenesis;And second area screening network and target detection network correspond to and divide between each layer structure
It is not attached by network connectivity layer, and each network connectivity layer operates with upper layer network articulamentum by deconvolution
Feature.
3. any direction Ship Target Detection method according to claim 1 based on cascade neural network, feature exist
In the training method of cascade neural network naval vessel detection model includes:
Training data generation step: the high-resolution satellite image marked in advance is cut into the sample data of fixed size;
Data augmentation step: network is generated based on sample data and super-resolution confrontation and generates multiple dimensioned high resolution image, is made
For training data;Using rotation, overturning and the further exptended sample data of luminance contrast method of adjustment, as training data;
Length-width ratio clustering step: the length and width and length-width ratio clustering of Ship Target are carried out to sample data, is gathered
Class is as a result, set length-width ratio parameter according to cluster result;
Training step: it is based on length-width ratio parameter and preset default frame vertical shift, training data is input to cascaded neural net
It is trained in network naval vessel detection model, and carries out duplicate removal using NMS algorithm, received in the training process based on loss function
Judgement is held back, trained cascade neural network naval vessel detection model is finally obtained.
4. any direction Ship Target Detection method according to claim 3 based on cascade neural network, feature exist
In training data generation step specifically includes:
Using deep learning network inputs image size as the size of sliding window, there is sliding on satellite image for overlapping;
If including effective Ship Target in current sliding window mouth, the corresponding image cropping of current sliding window mouth is come out, meanwhile,
Naval vessel in current window is updated to the coordinate in current sliding window mouth relative to the coordinate of whole picture satellite image, and is saved as
The corresponding XML of the image cut out marks file.
5. any direction Ship Target Detection method according to claim 4 based on cascade neural network, feature exist
In the method for discrimination of effective Ship Target includes:
Sliding window area and Ship Target quadrilateral area overlapping area and naval vessel quadrilateral area area ratio are greater than 0.5,
Then it is determined as effective Ship Target.
6. any direction Ship Target Detection method according to claim 3 based on cascade neural network, feature exist
In loss function employed in training step are as follows:
Wherein, which default frame i indicates,Indicate the classification with i-th of default matched true frame of frame,It indicates and i-th
Default position and the size of the matched true frame of frame, piIndicate confidence level, xiIndicate the seat that frame is defaulted in second area screening network
Mark, ciIndicate prediction classification, tiIndicate the prediction coordinate information in target detection network;NrpnAnd NodnRespectively indicate second area
The positive sample screened in network and target detection network defaults frame quantity;LbIndicate two-value Classification Loss, LmIndicate multi-class damage
It loses, LrIt indicates to return loss,If indicating, the confidence level of negative sample is greater than a threshold value, returns to 1, otherwise returns to 0;If
Nrpn=0, settingWithIf Nodn=0, then it is arrangedWith
7. any direction Ship Target Detection method according to claim 3 based on cascade neural network, feature exist
In multiple dimensioned high resolution image includes: the shadow of 4 times of resolution ratio of image data and raw video of 2 times of resolution ratio of raw video
As data.
8. any direction Ship Target Detection side described in any one based on cascade neural network according to claim 1~7
Method, which is characterized in that the first area screening network is using the PNet network in MTCNN model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910801199.0A CN110516605A (en) | 2019-08-28 | 2019-08-28 | Any direction Ship Target Detection method based on cascade neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910801199.0A CN110516605A (en) | 2019-08-28 | 2019-08-28 | Any direction Ship Target Detection method based on cascade neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110516605A true CN110516605A (en) | 2019-11-29 |
Family
ID=68628639
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910801199.0A Pending CN110516605A (en) | 2019-08-28 | 2019-08-28 | Any direction Ship Target Detection method based on cascade neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110516605A (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111178267A (en) * | 2019-12-30 | 2020-05-19 | 成都数之联科技有限公司 | Video behavior identification method for monitoring illegal fishing |
CN111353459A (en) * | 2020-03-10 | 2020-06-30 | 北京观微科技有限公司 | Ship target detection method under resource-limited condition |
CN111414821A (en) * | 2020-03-11 | 2020-07-14 | 北京市商汤科技开发有限公司 | Target detection method and related device |
CN112084869A (en) * | 2020-08-10 | 2020-12-15 | 北京航空航天大学 | Compact quadrilateral representation-based building target detection method |
CN112270259A (en) * | 2020-10-28 | 2021-01-26 | 中国人民解放军战略支援部队航天工程大学 | SAR image ship target rapid detection method based on lightweight convolutional neural network |
CN112651671A (en) * | 2021-01-18 | 2021-04-13 | 中国民航信息网络股份有限公司 | Flight space adjusting method and related equipment |
CN112735600A (en) * | 2020-12-30 | 2021-04-30 | 华南师范大学 | Early warning method based on big data monitoring and deep learning cascade prediction |
CN112766108A (en) * | 2021-01-08 | 2021-05-07 | 西安电子科技大学 | SAR image target detection method based on context information |
CN113420745A (en) * | 2021-08-25 | 2021-09-21 | 江西中业智能科技有限公司 | Image-based target identification method, system, storage medium and terminal equipment |
CN114049478A (en) * | 2022-01-13 | 2022-02-15 | 武汉理工大学 | Infrared ship image rapid identification method and system based on improved Cascade R-CNN |
CN114898308A (en) * | 2022-07-14 | 2022-08-12 | 上海鹰觉科技有限公司 | Ship cockpit position detection method and system based on deep convolutional neural network |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108460382A (en) * | 2018-03-26 | 2018-08-28 | 西安电子科技大学 | Remote sensing image Ship Detection based on deep learning single step detector |
CN108647665A (en) * | 2018-05-18 | 2018-10-12 | 西安电子科技大学 | Vehicle real-time detection method of taking photo by plane based on deep learning |
CN108846446A (en) * | 2018-07-04 | 2018-11-20 | 国家新闻出版广电总局广播科学研究院 | The object detection method of full convolutional network is merged based on multipath dense feature |
CN109447078A (en) * | 2018-10-23 | 2019-03-08 | 四川大学 | A kind of detection recognition method of natural scene image sensitivity text |
CN110084292A (en) * | 2019-04-18 | 2019-08-02 | 江南大学 | Object detection method based on DenseNet and multi-scale feature fusion |
-
2019
- 2019-08-28 CN CN201910801199.0A patent/CN110516605A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108460382A (en) * | 2018-03-26 | 2018-08-28 | 西安电子科技大学 | Remote sensing image Ship Detection based on deep learning single step detector |
CN108647665A (en) * | 2018-05-18 | 2018-10-12 | 西安电子科技大学 | Vehicle real-time detection method of taking photo by plane based on deep learning |
CN108846446A (en) * | 2018-07-04 | 2018-11-20 | 国家新闻出版广电总局广播科学研究院 | The object detection method of full convolutional network is merged based on multipath dense feature |
CN109447078A (en) * | 2018-10-23 | 2019-03-08 | 四川大学 | A kind of detection recognition method of natural scene image sensitivity text |
CN110084292A (en) * | 2019-04-18 | 2019-08-02 | 江南大学 | Object detection method based on DenseNet and multi-scale feature fusion |
Non-Patent Citations (3)
Title |
---|
SHIFENG ZHANG ET AL: "Single-Shot Refinement Neural Network for Object Detection", 《HTTPS://ARXIV.ORG/ABS/1711.06897》 * |
李健伟等: "基于深度学习的SAR图像舰船检测数据集及性能分析", 《第五届高分辨率对地观测学术年会论文集》 * |
邱世赟: "基于深度学习的高分辨率光学遥感图像舰船检测", 《中国优秀硕士学位论文全文数据库 工程科技II辑》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111178267A (en) * | 2019-12-30 | 2020-05-19 | 成都数之联科技有限公司 | Video behavior identification method for monitoring illegal fishing |
CN111353459A (en) * | 2020-03-10 | 2020-06-30 | 北京观微科技有限公司 | Ship target detection method under resource-limited condition |
CN111414821A (en) * | 2020-03-11 | 2020-07-14 | 北京市商汤科技开发有限公司 | Target detection method and related device |
CN111414821B (en) * | 2020-03-11 | 2023-12-19 | 北京市商汤科技开发有限公司 | Target detection method and related device |
CN112084869B (en) * | 2020-08-10 | 2023-04-07 | 北京航空航天大学 | Compact quadrilateral representation-based building target detection method |
CN112084869A (en) * | 2020-08-10 | 2020-12-15 | 北京航空航天大学 | Compact quadrilateral representation-based building target detection method |
CN112270259A (en) * | 2020-10-28 | 2021-01-26 | 中国人民解放军战略支援部队航天工程大学 | SAR image ship target rapid detection method based on lightweight convolutional neural network |
CN112270259B (en) * | 2020-10-28 | 2023-09-15 | 中国人民解放军战略支援部队航天工程大学 | SAR image ship target rapid detection method based on lightweight convolutional neural network |
CN112735600A (en) * | 2020-12-30 | 2021-04-30 | 华南师范大学 | Early warning method based on big data monitoring and deep learning cascade prediction |
CN112766108A (en) * | 2021-01-08 | 2021-05-07 | 西安电子科技大学 | SAR image target detection method based on context information |
CN112651671A (en) * | 2021-01-18 | 2021-04-13 | 中国民航信息网络股份有限公司 | Flight space adjusting method and related equipment |
CN113420745B (en) * | 2021-08-25 | 2021-12-24 | 江西中业智能科技有限公司 | Image-based target identification method, system, storage medium and terminal equipment |
CN113420745A (en) * | 2021-08-25 | 2021-09-21 | 江西中业智能科技有限公司 | Image-based target identification method, system, storage medium and terminal equipment |
CN114049478A (en) * | 2022-01-13 | 2022-02-15 | 武汉理工大学 | Infrared ship image rapid identification method and system based on improved Cascade R-CNN |
CN114898308A (en) * | 2022-07-14 | 2022-08-12 | 上海鹰觉科技有限公司 | Ship cockpit position detection method and system based on deep convolutional neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110516605A (en) | Any direction Ship Target Detection method based on cascade neural network | |
CN110516606A (en) | High-resolution satellite image any direction Ship Target Detection method | |
Cheng et al. | FusionNet: Edge aware deep convolutional networks for semantic segmentation of remote sensing harbor images | |
CN109711295B (en) | Optical remote sensing image offshore ship detection method | |
CN109299688B (en) | Ship detection method based on deformable fast convolution neural network | |
CN110378308B (en) | Improved port SAR image near-shore ship detection method based on fast R-CNN | |
CN102867196B (en) | Based on the complicated sea remote sensing image Ship Detection of Gist feature learning | |
CN109271856A (en) | Remote sensing image object detection method based on expansion residual error convolution | |
CN109740460B (en) | Optical remote sensing image ship detection method based on depth residual error dense network | |
CN109117802B (en) | Ship detection method for large-scene high-resolution remote sensing image | |
CN107247930A (en) | SAR image object detection method based on CNN and Selective Attention Mechanism | |
CN109766830A (en) | A kind of ship seakeeping system and method based on artificial intelligence image procossing | |
CN108304873A (en) | Object detection method based on high-resolution optical satellite remote-sensing image and its system | |
CN109427055B (en) | Remote sensing image sea surface ship detection method based on visual attention mechanism and information entropy | |
CN109657541A (en) | A kind of ship detecting method in unmanned plane image based on deep learning | |
CN104331886B (en) | Port region ship and warship detection method based on high resolution SAR image | |
CN103020975A (en) | Wharf and ship segmentation method combining multi-source remote sensing image characteristics | |
CN110008900B (en) | Method for extracting candidate target from visible light remote sensing image from region to target | |
CN108229342A (en) | A kind of surface vessel target automatic testing method | |
CN110414509A (en) | Stop Ship Detection in harbour based on the segmentation of extra large land and feature pyramid network | |
CN107704865A (en) | Fleet Targets Detection based on the extraction of structure forest edge candidate region | |
CN112149591A (en) | SSD-AEFF automatic bridge detection method and system for SAR image | |
CN110516560A (en) | Remote sensing image object detection method based on FPGA isomery deep learning | |
CN108596952A (en) | Fast deep based on candidate region screening learns Remote Sensing Target detection method | |
He et al. | Ship detection without sea-land segmentation for large-scale high-resolution optical satellite images |
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 | ||
CB03 | Change of inventor or designer information | ||
CB03 | Change of inventor or designer information |
Inventor after: Wang Lei Inventor after: Yu Jintao Inventor after: Zhao Xiuyun Inventor before: Wang Lei Inventor before: Yu Jintao Inventor before: Guo Haitao Inventor before: Zhao Xiuyun |
|
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191129 |