CN110008899A - A kind of visible remote sensing image candidate target extracts and classification method - Google Patents
A kind of visible remote sensing image candidate target extracts and classification method Download PDFInfo
- Publication number
- CN110008899A CN110008899A CN201910262483.5A CN201910262483A CN110008899A CN 110008899 A CN110008899 A CN 110008899A CN 201910262483 A CN201910262483 A CN 201910262483A CN 110008899 A CN110008899 A CN 110008899A
- Authority
- CN
- China
- Prior art keywords
- candidate
- candidate target
- target
- classification
- extracted
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of extraction of visible remote sensing image candidate target and classification methods, belong to remote sensing image processing and analysis technical field, solve the problems, such as that existing method extraction candidate target quantity is too many and nicety of grading is low.The following steps are included: extracting Large-scale areas in remote sensing images by the first sliding window, and it is input in trained candidate region identification model, obtains the candidate region comprising interesting target;Candidate target is extracted in above-mentioned candidate region by the second sliding window;Two step classification are carried out using candidate target of the trained candidate target disaggregated model to said extracted, determine the classification of candidate target.The present invention realizes under identical recall rate, the candidate target quantity of extraction far fewer than existing method, meanwhile, under conditions of using identical feature and identical classifier, classified by two steps, substantially increase candidate target nicety of grading.
Description
Technical field
The present invention relates to remote sensing image processing and analysis technical field more particularly to a kind of visible remote sensing image candidate mesh
Mark extracts and classification method.
Background technique
Candidate target extraction is the important foundation of visible remote sensing image target detection identification, is in detection identification process
Key link.Candidate target extraction refers to be scanned in the picture by specific method, removes non-targeted in image, guarantor
It stays suspected target as candidate target, is supplied to subsequent target detection identification link, reduces target and accurately detect cognitive phase
Data processing amount, improve detection recognition speed and precision.
Existing visible remote sensing image candidate target extracting method can be divided into two classes: method based on Threshold segmentation and
The method of view-based access control model conspicuousness.Wherein, the method based on Threshold segmentation is using target and background in gray value, color, texture
Etc. difference, find suitable threshold value and distinguish target and background.Main problem existing for such methods is segmentation energy
Power is limited, is generally only used for the relatively simple occasion of background, such as naval vessel detection etc..The method of view-based access control model conspicuousness is simulated
The target search procedure of the mankind, the information utilized include color, direction and gradient etc..Main problem existing for such methods is
It is directed to general target, all candidates for having the target of obvious geometric profile to be all taken as candidate target, therefore extracting
Destination number is more.
In addition, after completing candidate target extraction, it is also necessary to which judgement of further classifying to candidate target identifies target
Concrete type, such as identify that certain airplane is fighter plane or airline carriers of passengers etc..However, existing visible remote sensing image is candidate
Objective classification method is absorbed in the extraction or study of target signature, and classification essence is improved by extracting or learning different features
Degree, but the nicety of grading of this mode is still to be improved.
Summary of the invention
In view of above-mentioned analysis, the present invention is intended to provide a kind of visible remote sensing image candidate target extracts and classification side
Method, to solve the problems, such as that existing method extraction candidate target quantity is too many and nicety of grading is low.
The purpose of the present invention is mainly achieved through the following technical solutions:
A kind of visible remote sensing image candidate target is provided to extract and classification method, comprising the following steps:
Large-scale areas is extracted in remote sensing images by the first sliding window, and is input to trained candidate region identification mould
In type, the candidate region comprising interesting target is obtained;
Candidate target is extracted in above-mentioned candidate region by the second sliding window;
Two step classification are carried out using candidate target of the candidate target disaggregated model to said extracted, determine the class of candidate target
Not.
The present invention has the beneficial effect that: the present invention carries out candidate mesh for the aggregation feature of visible remote sensing image target
Mark extracts, and the classification of candidate target is accurately identified by the classification of two steps;Under identical recall rate, the candidate target number of extraction
It measures far fewer than existing method, meanwhile, under conditions of using identical feature and identical classifier, classified by two steps, greatly
Candidate target nicety of grading is improved greatly.In addition, application scenarios of the present invention are extensive, fly suitable for visible remote sensing image
The multi-class targets such as machine, naval vessel, vehicle;It is also expansible to be applied to infrared, EO-1 hyperion, the extraction of synthetic aperture radar candidate target and divide
Class can satisfy various environment occasions.
On the basis of above scheme, the present invention has also done following improvement:
Further, the candidate target disaggregated model includes: the second convolutional neural networks and several second classifiers, each
Second classifier is corresponding with two possible classifications of candidate target respectively;
Second convolutional neural networks are VGGNet-16, include 13 convolutional layers, 2 full articulamentums and 1 output
Layer;The feature of candidate target is extracted from full articulamentum and output layer respectively, the feature that the output layer extracts is waited for determining
Select first M possible classification of target;
Second classifier is LIBSVM classifier, by the feature for the candidate target that above-mentioned full articulamentum extracts,
The final classification of candidate target is determined in above-mentioned first M possible classification.
It is further, described to carry out two step classification using candidate target of the candidate target disaggregated model to said extracted, comprising:
The candidate target of extraction is input to trained second convolutional neural networks;
The feature of candidate target is extracted by the output layer of second convolutional neural networks, and utilizes the spy of the extraction
Sign calculates class probability value;
The class probability value of above-mentioned calculating is ranked up, when the maximum value in the class probability value is greater than the threshold value of setting
When, then using the corresponding classification of maximum value as the classification of candidate target;Otherwise, second step classification is carried out;
The progress second step classification, comprising: will be extracted from first of second convolutional neural networks full articulamentum
Candidate target feature be input in corresponding second classifier, from the probability value of M before above-mentioned sequence it is corresponding M may classification
Classification belonging to middle selection candidate target.
Further, the candidate region identification model includes: the first convolutional neural networks and the first classifier;
First convolutional neural networks are VGGNet-16, include 13 convolutional layers, 2 full articulamentums and 1 output
Layer;The feature of Large-scale areas is extracted from first full articulamentum;
First classifier is LIBSVM classifier, by the Large-scale areas feature of said extracted to Large-scale areas
Classify.
Further, second sliding window that passes through carries out candidate target extraction in above-mentioned candidate region,
Include:
Several Small-scale spaces are extracted in above-mentioned candidate region using the second sliding window as candidate target, described second
For the size of sliding window less than the first sliding window, the degree of overlapping between adjacent Small-scale space is 25%.
Further, the calculating class probability value are as follows:
The feature of the extraction is normalized, the class probability value of candidate target is obtained.
Further, when being trained to second convolutional neural networks, training parameter setting are as follows: global cycle number is
10000, momentum 0.9, weight decays to 0.0005, and initial learning rate is 0.0001, and 4000 learning rates of every circulation become
Originally 1/10, the size of data block is 64.
Further, the quantity of second classifier is N × (N-1)/2, wherein N is the total of the possible classification of candidate target
Number.
Further, further include being screened using non-maxima suppression to obtained candidate region:
When there is IOU threshold value of multiple candidate region location overlapping degrees more than setting, non-maxima suppression is carried out, and
Candidate region is arranged from high to low according to score;Since the candidate region of highest scoring, successively with remaining all candidate regions
Domain is compared, and the candidate region that the area ratio of the candidate region of overlapping area and highest scoring is more than default ratio is given up
It abandons, the candidate region after obtaining one group of screening;Successively all candidate regions are equally handled, until traversing all candidates
Region obtains the set of candidate regions that overlapping area between any two is both less than default ratio.
Further, the size of first sliding window is 4 times of maximum target size in image, and the size of the second sliding window is figure
1 times of maximum target size as in.
It in the present invention, can also be combined with each other between above-mentioned each technical solution, to realize more preferred assembled schemes.This
Other feature and advantage of invention will illustrate in the following description, also, certain advantages can become from specification it is aobvious and
It is clear to, or understand through the implementation of the invention.The objectives and other advantages of the invention can by specification, claims with
And it is achieved and obtained in specifically noted content in attached drawing.
Detailed description of the invention
Attached drawing is only used for showing the purpose of specific embodiment, and is not to be construed as limiting the invention, in entire attached drawing
In, identical reference symbol indicates identical component.
Fig. 1 is that visible remote sensing image candidate target extracts and classification method flow chart in the embodiment of the present invention;
Fig. 2 is that candidate target extracts flow chart in the embodiment of the present invention;
Fig. 3 is that schematic diagram is extracted in Large-scale areas in the embodiment of the present invention;
Fig. 4 is that candidate target extracts schematic diagram in the embodiment of the present invention;
Fig. 5 is to carry out two step classification process figures to candidate target in the embodiment of the present invention.
Specific embodiment
Specifically describing the preferred embodiment of the present invention with reference to the accompanying drawing, wherein attached drawing constitutes the application a part, and
Together with embodiments of the present invention for illustrating the principle of the present invention, it is not intended to limit the scope of the present invention.
A specific embodiment of the invention discloses a kind of visible remote sensing image candidate target and extracts and classification side
Method.As shown in Figure 1, comprising the following steps:
Step S1, Large-scale areas is extracted in remote sensing images by the first sliding window, and is input to trained candidate regions
In the identification model of domain, the candidate region comprising interesting target is obtained;
Step S2, candidate target is extracted in above-mentioned candidate region by the second sliding window;
Step S3, two step classification are carried out using candidate target of the trained candidate target disaggregated model to said extracted,
Determine the classification of candidate target.
Compared with prior art, visible remote sensing image candidate target extraction provided in this embodiment and classification method, needle
Candidate target extraction is carried out to the aggregation feature of visible remote sensing image target, and candidate mesh is accurately identified by the classification of two steps
Mark classification;Under identical recall rate, the candidate target quantity of extraction far fewer than existing method, meanwhile, use identical spy
It seeks peace under conditions of identical classifier, is classified by two steps, substantially increase candidate target nicety of grading.In addition, of the invention
Application scenarios are extensive, the multi-class targets such as aircraft, naval vessel, vehicle suitable for visible remote sensing image;It is also expansible to be applied to
Infrared, EO-1 hyperion, synthetic aperture radar candidate target extract and classification, can satisfy various environment occasions;It efficiently solves existing
There is the candidate target extracting method segmentation ability based on Threshold segmentation limited, application scenarios are limited and the time of view-based access control model conspicuousness
The problem that the candidate target quantity for selecting target extraction method to extract is too many and candidate target nicety of grading is low.
Specifically, the visible remote sensing image of acquisition is handled in the present embodiment, including candidate target extract and
Candidate target is classified two stages, wherein carries out candidate target extraction, (step S1- step S2) as shown in Figure 2.
In step sl, Large-scale areas is extracted in remote sensing images by the first sliding window, and is input to trained time
In favored area identification model, the candidate region comprising interesting target is obtained;Specifically includes the following steps:
Step S101 extracts Large-scale areas by the first sliding window, wherein the size of the first sliding window is according to mesh in the picture
Size is marked to determine;
As shown in figure 3, since one end of input picture, Large-scale areas is sequentially cut using the method for sliding window, until
Throughout whole image region, completes Large-scale areas and extract.Illustratively, since the upper left corner of input picture, according to from a left side
Large-scale areas is cut to sequence right, from top to bottom.
The size of Large-scale areas determines that the method for the present invention is suitable for visual remote sensing figure according to the size of target in image
The multi-class targets such as aircraft, naval vessel, vehicle as in;It is also expansible to be applied to the candidate mesh such as infrared, EO-1 hyperion, synthetic aperture radar
Mark.In the present embodiment, it is only illustrated by taking Aircraft Targets as an example, extracts all aircrafts as candidate target, need
It is bright, due to including multiple types aircraft (in the present embodiment be 10 classes) in remote sensing images, and the size of different type aircraft
Difference need to determine the first sliding window with aircraft size maximum in image, illustratively, the size of maximum target in the picture
About 64 × 64 pixels, thereby determine that big ruler by 4 times of the length of the first sliding window (i.e. Large-scale areas) and wide substantially maximum target
The size for spending region is 256 × 256 pixels.
It should be noted that being divided in different Large-scale areas in order to avoid omitting target or target occur, give
Subsequent further object detection identification brings interference or can not recognize interested target.In the present embodiment, by first
The sliding step of sliding window is set smaller than the length and width dimensions of sliding window, to retain between the adjacent Large-scale areas divided
Certain degree of overlapping, it is preferred that the degree of overlapping between adjacent area is 25%.
Step S102, by the Large-scale areas of said extracted be sequentially inputted in trained candidate region identification model into
Row classification, obtains one or more candidate regions comprising interesting target.
Wherein, candidate region identification model is made of the first convolutional neural networks and the first classifier.
First convolutional neural networks are used to extract the feature of Large-scale areas, and the feature of extraction is transmitted to the first classification
Device is classified, using the Large-scale areas met as candidate region;First convolutional neural networks can be using a variety of convolution minds
Feature extraction is carried out through the different characteristic extract layer in network or network, can also use existing non-convolutional neural networks class
Feature extracting method extracts feature, can achieve the effect that in the present embodiment;Preferably, made in this example using VGGNet-16
For the first convolutional neural networks, and feature is extracted from first full articulamentum of VGGNet-16;The network includes 13 convolution
Layer, 2 full articulamentums and 1 output layer;
First classifier can realize classification feature using a variety of existing classifiers, use SVM in the present embodiment
(support vector machine, support vector machines), it is preferred that classified using LIBSVM.
Before the candidate region identification model using above-mentioned building carries out identification classification, need to establish large-scale image
Categorized data set is trained, so that model has powerful ability in feature extraction, improves the accuracy rate of classification.
Specifically, when constructing large-scale image categorized data set, can make by online public database or certainly
Mode obtains, it is preferred that the training set using Google Maps remote sensing image data as candidate region identification model, meanwhile, also
Data intensive data can be carried out to interference reinforcement (data level overturning, plus noise, random cropping), to improve the instruction of model
Practice effect.
It should be noted that in order to further increase the accuracy of model candidate region identification, in the training image of selection
In, a part of image need to include complete target.
It after collecting data set, is divided into for trained Large-scale areas, and each region is labeled, thoughts will be contained
The Large-scale areas of targets of interest and Large-scale areas without containing interesting target are classified as different classes: one kind is comprising feeling emerging
The Large-scale areas of interesting target, class label are set as 1;Another kind of is the Large-scale areas not comprising interesting target, class label
It is set as 0.
After completing data set, initial method, learning rate, optimizer and loss function are set, utilizes the big ruler of two classes
The image and corresponding class label for spending region are trained the first convolutional neural networks;From trained first convolution nerve net
The full articulamentum of first of network extracts the feature of two class Large-scale areas, utilizes the feature and corresponding class of two class Large-scale areas
Label is trained the first classifier;After training, the ideal candidate region identification model of recognition effect is obtained.
Each Large-scale areas extracted in above-mentioned steps S101 is sequentially inputted to above-mentioned trained candidate region identification
Model carries out identification classification, obtains the Large-scale areas comprising interesting target as candidate region.Specifically, the first volume is extracted
Feature of the full articulamentum feature of product neural network as Large-scale areas;By the feature of the Large-scale areas to be sorted of extraction
It is input in the first classifier and classifies, if the class label that the first classifier provides is 1, then it is assumed that the Large-scale areas is
Candidate region includes the region of interesting target;If the class label that the first classifier provides is 0, give up the large scale
Region.
In order to be further reduced the quantity of extra candidate region, reduces target detection and identify workload, improve subsequent detection
The speed and precision of identification;The present embodiment solves the problems, such as candidate region redundancy using non-maxima suppression, specifically, works as process
Occurring multiple candidate region location overlapping degrees in the set of candidate regions that candidate region identification model is classified is more than setting
IOU threshold value when, carry out non-maxima suppression, and arrange candidate region from high to low according to score;Then, from highest scoring
Candidate region starts, and is successively compared with remaining all candidate regions, by the candidate region of overlapping area and highest scoring
Area ratio be more than that default ratio (preferred, give up, the candidate regions after obtaining one group of screening by the candidate region being set as 0.7)
Domain set, is then successively equally handled, and until traversing all candidate regions, is obtained overlapping area between any two and is both less than
The set of candidate regions of default ratio.
In step s 2, candidate target is extracted in above-mentioned candidate region by the second sliding window, obtains candidate target.
As shown in figure 4, in obtained candidate region, using the method for sliding window, since the image upper left corner, according to from a left side
Small-scale space is cut to sequence right, from top to bottom, using the Small-scale space comprising part or all of target as candidate mesh
Mark.To avoid omitting target, adjacent interregional needs retain certain degree of overlapping, it is preferred that between adjacent Small-scale space
Degree of overlapping is 25%.
The size (size of Small-scale space) of second sliding window is determined according to the size of target in image, while the second sliding window
Size be less than above-mentioned first sliding window.In the present embodiment, still by taking Aircraft Targets as an example, maximum Aircraft Targets are in the picture
Size is about 64 × 64 pixels, thereby determines that the size of Small-scale space is 64 × 64 pixels.
In step s3, two steps point are carried out using candidate target of the trained candidate target disaggregated model to said extracted
Class determines the classification of candidate target.As shown in figure 5, specifically includes the following steps:
The candidate target extracted in step s 2 is input to trained candidate target disaggregated model by step S301;
Candidate target disaggregated model is made of the second convolutional neural networks and multiple second classifiers.Second convolution nerve net
Network is used to extract the feature (carrying out two step feature extractions) of candidate target, wherein the feature that the first step is extracted is general for calculating class
Rate value (see step S302) determines that first M of candidate target may classification;The feature that second step extracts is for being transmitted to corresponding the
Two classifiers are classified, and determine the final classification of candidate target;Preferably, still use VGGNet-16 as the second convolution
Neural network, and feature is extracted from the output layer of VGGNet-16 and first full articulamentum respectively;The network includes 13 convolution
Layer, 2 full articulamentums and 1 output layer;Second classifier still uses LIBSVM to classify, wherein the second classifier
Quantity determines (such as: the type summation of aircraft in the present embodiment) by the classification sum of candidate target, each second classifier and waits
Select two possible classifications of target corresponding, i.e., each second classifier is used to further confirm that possible class from two classifications
Not, and two classifications of the second different classifiers for differentiation are not exactly the same;Specifically, it may be expressed as: N × (N-1)/2
A linear one-to-one SVM, wherein N is the sum of classification.
Before carrying out actual classification, be still required to be trained candidate target disaggregated model by data set, i.e., it is sharp
The second convolutional neural networks are trained with candidate target image and corresponding class label;And from trained convolutional Neural net
Network extracts the full articulamentum feature of candidate target, using the full articulamentum feature and corresponding class label of candidate target to corresponding
Second classifier is trained.After completing data set and collecting (preferred, using Google Maps remote sensing image data), set
Initial method, learning rate, optimizer and loss function are trained the second convolutional neural networks and the second classifier, obtain
Obtain the ideal candidate target disaggregated model of classifying quality.
In the present embodiment, parameter setting that when the second convolution neural metwork training uses are as follows: global cycle number is 10000,
Momentum is 0.9, and weight decays to 0.0005, and initial learning rate is 0.0001, and every circulation 4000 times, learning rate becomes original
1/10, the size of data block is 64.
It should be noted that before carrying out actual classification, it is also necessary to according to classification task, by the second convolutional neural networks
Output layer in neuron the quantity classification number that is revised as candidate target by 1000.Illustratively, this example is directed to 10 classes
Aircraft Targets are classified, therefore N is 10;The number of output layer is changed to 10 by 1000.
After obtaining trained candidate target disaggregated model, the candidate target of extraction is input to trained second convolution
Neural network;And the feature of candidate target is extracted by the output layer of network, and utilize the feature calculation class probability of the extraction
Value;Specifically, for target image oi, need to determine class c belonging to the image from N number of classificationi, candidate target image is defeated
Enter to trained second convolutional neural networks, from the last layer of network, i.e., extracts clarification of objective f in output layeri:
fi={ fI, 1, fI, 2..., fI, N}
In formula, fiIt is the vector of a N-dimensional, in the present embodiment, N takes 10.
Features above is normalized, target image o is obtainediClass Probability pi={ pI, 1, pI, 2..., pI, N,
piIt is the vector of a N-dimensional, first of element therein represents the probability that target belongs to l class, and the sum of N number of element is 1.
Step S302 sorts to the class probability value calculated in step S301 according to descending sequence, and to sequence after
Class probability value judged;
When the maximum value in class probability value is greater than the threshold value of setting, then first step classification is executed, by maximum kind probability value
Corresponding classification is assigned to candidate target, completes candidate target classification;Otherwise step S303 is carried out.
Wherein, the class probability value that upper step obtains is ranked up according to sequence from big to small, obtains p 'i:
p′i={ p 'I, 1, p 'I, 2..., p 'I, N}
In formula, p 'iIt is the vector of a N-dimensional, first of element therein represents the probability that candidate target belongs to l class, and
p′I, 1≥p′I, 2≥…≥p′I, N。
If maximum probability value p 'I, 10.9) threshold value T greater than setting (preferably, is set as, then corresponds to the probability value
Classification be assigned to candidate target;Otherwise it is assumed that the classification of candidate target is { p 'I, 1, p 'I, 2..., p 'I, MCorresponding M classification
In certain is a kind of, wherein M < N.The size of M can be determines according to actual conditions, it is preferred that M is set as 3, i.e., from sorting first three
Candidate target generic is determined in classification.
Step S303 extracts the spy of candidate target from first full articulamentum of trained second convolutional neural networks
Sign, and be input in corresponding second classifier, carry out second step classification;Before sorting in the corresponding M class of class probability value of M
Select classification belonging to candidate target.
Specifically, in the present embodiment, extraction is characterized in the vector of 4096 dimensions, and the full articulamentum feature of extraction is inputted
To trained corresponding second classifier, class c belonging to target is determined in the corresponding M class of probability value of M before sortingi。
In order to verify the effect of visible remote sensing image candidate target extraction and classification method in the present embodiment, with remote sensing shadow
Aircraft, naval vessel etc. be respectively as candidate target as in, with existing view-based access control model conspicuousness and the candidate mesh based on Threshold segmentation
Mark extracting method is compared, the results showed that under identical recall rate, candidate target quantity that the present invention extracts far fewer than
Existing method.Specifically, aircraft candidate target is extracted, compared to the method for view-based access control model conspicuousness, time that the present invention extracts
Destination number is selected to reduce about 40%;Naval vessel candidate target is extracted, compared to the method based on Threshold segmentation, what the present invention extracted
Candidate target quantity reduces about 30%.
In addition, two step classification methods of the invention can incite somebody to action under conditions of use identical feature and identical classifier
Nicety of grading improves 2 to 3 percentage points.
It will be understood by those skilled in the art that realizing all or part of the process of above-described embodiment method, meter can be passed through
Calculation machine program instruction relevant hardware is completed, and the program can be stored in computer readable storage medium.Wherein, described
Computer readable storage medium is disk, CD, read-only memory or random access memory etc..
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.
Claims (10)
1. a kind of visible remote sensing image candidate target extracts and classification method, which comprises the following steps:
Large-scale areas is extracted in remote sensing images by the first sliding window, and is input to trained candidate region identification model
In, obtain the candidate region comprising interesting target;
Candidate target is extracted in above-mentioned candidate region by the second sliding window;
Two step classification are carried out using candidate target of the trained candidate target disaggregated model to said extracted, determine candidate target
Classification.
2. the method according to claim 1, wherein the candidate target disaggregated model includes: the second convolution mind
Through network and several second classifiers, each second classifier is corresponding with two possible classifications of candidate target respectively;
Second convolutional neural networks are VGGNet-16, include 13 convolutional layers, 2 full articulamentums and 1 output layer;Point
The feature of candidate target is extracted not from full articulamentum and output layer, the feature that the output layer extracts is for determining candidate target
First M may classification;
Second classifier is LIBSVM classifier, by the feature for the candidate target that above-mentioned full articulamentum extracts, upper
The final classification of candidate target is determined before stating in M possible classification.
3. according to the method described in claim 2, it is characterized in that, it is described using candidate target disaggregated model to said extracted
Candidate target carries out two step classification, comprising:
The candidate target of extraction is input to trained second convolutional neural networks;
The feature of candidate target is extracted by the output layer of second convolutional neural networks, and utilizes the feature meter of the extraction
Calculate class probability value;
The class probability value of above-mentioned calculating is ranked up, when the maximum value in the class probability value is greater than the threshold value of setting, then
Using the corresponding classification of maximum value as the classification of candidate target;Otherwise, second step classification is carried out;
The progress second step classification, comprising: the time that will be extracted from first of second convolutional neural networks full articulamentum
It selects target signature to be input in corresponding second classifier, may be selected in classification from the probability value of M before above-mentioned sequence corresponding M
Select classification belonging to candidate target.
4. method described in one of -3 according to claim 1, which is characterized in that the candidate region identification model includes: first
Convolutional neural networks and the first classifier;
First convolutional neural networks are VGGNet-16, include 13 convolutional layers, 2 full articulamentums and 1 output layer;From
The feature of Large-scale areas is extracted in first full articulamentum;
First classifier is LIBSVM classifier, is carried out by the Large-scale areas feature of said extracted to Large-scale areas
Classification.
5. according to the method described in claim 4, it is characterized in that, second sliding window that passes through carries out in above-mentioned candidate region
Candidate target extracts, comprising:
Several Small-scale spaces are extracted in above-mentioned candidate region using the second sliding window as candidate target, second sliding window
Size less than the first sliding window, the degree of overlapping between adjacent Small-scale space is 25%.
6. according to the method described in claim 3, it is characterized in that, the calculating class probability value are as follows: to the feature of the extraction
It is normalized, obtains the class probability value of candidate target.
7. according to the method described in claim 6, it is characterized in that, when being trained to second convolutional neural networks,
Training parameter setting are as follows: global cycle number is 10000, and momentum 0.9, weight decays to 0.0005, and initial learning rate is
0.0001,4000 learning rates of every circulation become original 1/10, and the size of data block is 64.
8. the method according to the description of claim 7 is characterized in that the quantity of second classifier be N × (N-1)/2,
In, N is the sum of the possible classification of candidate target.
9. method according to claim 1 or 8, which is characterized in that further include using non-maxima suppression to obtained time
Favored area is screened:
When there is IOU threshold value of multiple candidate region location overlapping degrees more than setting, progress non-maxima suppression, and according to
Score arranges candidate region from high to low;Since the candidate region of highest scoring, successively with remaining all candidate regions into
Row compares, and the candidate region that the area ratio of the candidate region of overlapping area and highest scoring is more than default ratio is given up, is obtained
Candidate region to after one group of screening;Successively all candidate regions are equally handled, until traversing all candidate regions,
Obtain the set of candidate regions that overlapping area between any two is both less than default ratio.
10. according to the method described in claim 9, it is characterized in that, the size of first sliding window is maximum target in image
4 times of size, the size of the second sliding window are 1 times of maximum target size in image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910262483.5A CN110008899B (en) | 2019-04-02 | 2019-04-02 | Method for extracting and classifying candidate targets of visible light remote sensing image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910262483.5A CN110008899B (en) | 2019-04-02 | 2019-04-02 | Method for extracting and classifying candidate targets of visible light remote sensing image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110008899A true CN110008899A (en) | 2019-07-12 |
CN110008899B CN110008899B (en) | 2021-02-26 |
Family
ID=67169597
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910262483.5A Active CN110008899B (en) | 2019-04-02 | 2019-04-02 | Method for extracting and classifying candidate targets of visible light remote sensing image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110008899B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110517575A (en) * | 2019-08-21 | 2019-11-29 | 华北水利水电大学 | A kind of surface water body drafting method and device |
CN111582176A (en) * | 2020-05-09 | 2020-08-25 | 湖北同诚通用航空有限公司 | Visible light remote sensing image withered and dead wood recognition software system and recognition method |
CN116229280A (en) * | 2023-01-09 | 2023-06-06 | 广东省科学院广州地理研究所 | Method and device for identifying collapse sentry, electronic equipment and storage medium |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113780256B (en) * | 2021-11-12 | 2022-03-15 | 科大讯飞(苏州)科技有限公司 | Image target detection method combining thickness classification and related device |
Citations (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105528078A (en) * | 2015-12-15 | 2016-04-27 | 小米科技有限责任公司 | Method and device controlling electronic equipment |
CN105589929A (en) * | 2015-12-09 | 2016-05-18 | 东方网力科技股份有限公司 | Image retrieval method and device |
CN106503742A (en) * | 2016-11-01 | 2017-03-15 | 广东电网有限责任公司电力科学研究院 | A kind of visible images insulator recognition methods |
CN106778835A (en) * | 2016-11-29 | 2017-05-31 | 武汉大学 | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic |
US20170213071A1 (en) * | 2016-01-21 | 2017-07-27 | Samsung Electronics Co., Ltd. | Face detection method and apparatus |
CN107122733A (en) * | 2017-04-25 | 2017-09-01 | 西安电子科技大学 | Hyperspectral image classification method based on NSCT and SAE |
CN107437083A (en) * | 2017-08-16 | 2017-12-05 | 上海荷福人工智能科技(集团)有限公司 | A kind of video behavior recognition methods of adaptive pool |
CN108805039A (en) * | 2018-04-17 | 2018-11-13 | 哈尔滨工程大学 | The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features |
CN108875794A (en) * | 2018-05-25 | 2018-11-23 | 中国人民解放军国防科技大学 | Image visibility detection method based on transfer learning |
CN108875667A (en) * | 2018-06-27 | 2018-11-23 | 北京字节跳动网络技术有限公司 | target identification method, device, terminal device and storage medium |
CN108960338A (en) * | 2018-07-18 | 2018-12-07 | 苏州科技大学 | The automatic sentence mask method of image based on attention-feedback mechanism |
CN109241817A (en) * | 2018-07-02 | 2019-01-18 | 广东工业大学 | A kind of crops image-recognizing method of unmanned plane shooting |
-
2019
- 2019-04-02 CN CN201910262483.5A patent/CN110008899B/en active Active
Patent Citations (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105589929A (en) * | 2015-12-09 | 2016-05-18 | 东方网力科技股份有限公司 | Image retrieval method and device |
CN105528078A (en) * | 2015-12-15 | 2016-04-27 | 小米科技有限责任公司 | Method and device controlling electronic equipment |
US20170213071A1 (en) * | 2016-01-21 | 2017-07-27 | Samsung Electronics Co., Ltd. | Face detection method and apparatus |
CN106503742A (en) * | 2016-11-01 | 2017-03-15 | 广东电网有限责任公司电力科学研究院 | A kind of visible images insulator recognition methods |
CN106778835A (en) * | 2016-11-29 | 2017-05-31 | 武汉大学 | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic |
CN107122733A (en) * | 2017-04-25 | 2017-09-01 | 西安电子科技大学 | Hyperspectral image classification method based on NSCT and SAE |
CN107437083A (en) * | 2017-08-16 | 2017-12-05 | 上海荷福人工智能科技(集团)有限公司 | A kind of video behavior recognition methods of adaptive pool |
CN108805039A (en) * | 2018-04-17 | 2018-11-13 | 哈尔滨工程大学 | The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features |
CN108875794A (en) * | 2018-05-25 | 2018-11-23 | 中国人民解放军国防科技大学 | Image visibility detection method based on transfer learning |
CN108875667A (en) * | 2018-06-27 | 2018-11-23 | 北京字节跳动网络技术有限公司 | target identification method, device, terminal device and storage medium |
CN109241817A (en) * | 2018-07-02 | 2019-01-18 | 广东工业大学 | A kind of crops image-recognizing method of unmanned plane shooting |
CN108960338A (en) * | 2018-07-18 | 2018-12-07 | 苏州科技大学 | The automatic sentence mask method of image based on attention-feedback mechanism |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110517575A (en) * | 2019-08-21 | 2019-11-29 | 华北水利水电大学 | A kind of surface water body drafting method and device |
CN110517575B (en) * | 2019-08-21 | 2021-03-02 | 华北水利水电大学 | Method and device for mapping surface water body |
CN111582176A (en) * | 2020-05-09 | 2020-08-25 | 湖北同诚通用航空有限公司 | Visible light remote sensing image withered and dead wood recognition software system and recognition method |
CN116229280A (en) * | 2023-01-09 | 2023-06-06 | 广东省科学院广州地理研究所 | Method and device for identifying collapse sentry, electronic equipment and storage medium |
CN116229280B (en) * | 2023-01-09 | 2024-06-04 | 广东省科学院广州地理研究所 | Method and device for identifying collapse sentry, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110008899B (en) | 2021-02-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2016037300A1 (en) | Method and system for multi-class object detection | |
CN108830285B (en) | Target detection method for reinforcement learning based on fast-RCNN | |
CN110008899A (en) | A kind of visible remote sensing image candidate target extracts and classification method | |
CN110826379B (en) | Target detection method based on feature multiplexing and YOLOv3 | |
CN110619369A (en) | Fine-grained image classification method based on feature pyramid and global average pooling | |
Xu et al. | Scale-aware feature pyramid architecture for marine object detection | |
CN108009509A (en) | Vehicle target detection method | |
CN106845430A (en) | Pedestrian detection and tracking based on acceleration region convolutional neural networks | |
CN110837836A (en) | Semi-supervised semantic segmentation method based on maximized confidence | |
CN106557778A (en) | Generic object detection method and device, data processing equipment and terminal device | |
CN110008900A (en) | A kind of visible remote sensing image candidate target extracting method by region to target | |
CN110018453A (en) | Intelligent type recognition methods based on aircraft track feature | |
CN110210431A (en) | A kind of point cloud classifications method based on cloud semantic tagger and optimization | |
CN114648665A (en) | Weak supervision target detection method and system | |
CN112766170B (en) | Self-adaptive segmentation detection method and device based on cluster unmanned aerial vehicle image | |
CN114332473A (en) | Object detection method, object detection device, computer equipment, storage medium and program product | |
CN110020669A (en) | A kind of license plate classification method, system, terminal device and computer program | |
CN112686902A (en) | Two-stage calculation method for brain glioma identification and segmentation in nuclear magnetic resonance image | |
CN113159215A (en) | Small target detection and identification method based on fast Rcnn | |
CN114241250A (en) | Cascade regression target detection method and device and computer readable storage medium | |
Fan et al. | A novel sonar target detection and classification algorithm | |
Balasubramanian et al. | Open-set recognition based on the combination of deep learning and ensemble method for detecting unknown traffic scenarios | |
CN113128564B (en) | Typical target detection method and system based on deep learning under complex background | |
CN117542082A (en) | Pedestrian detection method based on YOLOv7 | |
CN116311387B (en) | Cross-modal pedestrian re-identification method based on feature intersection |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |