CN108319949A - Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image - Google Patents
Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image Download PDFInfo
- Publication number
- CN108319949A CN108319949A CN201810075285.3A CN201810075285A CN108319949A CN 108319949 A CN108319949 A CN 108319949A CN 201810075285 A CN201810075285 A CN 201810075285A CN 108319949 A CN108319949 A CN 108319949A
- Authority
- CN
- China
- Prior art keywords
- ship target
- rectangle frame
- angle
- frame
- image
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes in a kind of high-resolution remote sensing image mostly towards Ship Target Detection and recognition methods, design and Implement under complex scene, in large scale environment mostly towards Ship Target Detection with identification network structure, the various dimensions features such as shallow-layer and the deep layer of Ship Target are automatically excavated based on convolutional neural networks, realize the quick accurate detection identification of the mostly Ship Target of direction, different scale and type in the high-resolution remote sensing image under small sample environment.The present invention compares the speed of traditional object detection method detection faster, and testing result is more accurate.
Description
Technical field
The present invention relates in technical field of earth observation more particularly to a kind of high-resolution remote sensing image mostly towards naval vessel mesh
Mark detection and recognition methods.
Background technology
With the continuous development of earth observation technology in recent years, the acquisition of high-resolution, high-precision remote sensing images becomes more
It is easy to add.High-resolution remote sensing image provides good basis for more accurate target detection and target identification.Make on naval vessel
For marine important goal, detection and identification all have great importance in civil and military field.Civilian aspect, naval vessel
Detection and identification can be monitored the position of offshore vessel, facilitate maritime search and rescue, owler search and marine site monitoring etc.；Army
With aspect, the form in battlefield can analyze etc. be of great significance with supplementary observation enemy's situation.
At present there is no perfect extensive Ship Target Detection and identification data set, existing data can not be directly utilized
With data processing method to being mostly trained towards Ship Target Detection and identification model based on deep learning.
Existing Remote Sensing Target detection, recognizer are broadly divided into conventional method and the method based on deep learning.
Steps are as follows for Ship Target Detection in conventional method：1) image preprocessing 2) 3) false-alarm go for candidate region extraction
It removes.Steps are as follows for ship seakeeping in conventional method：1) image preprocessing 2) Ship Target feature extraction 3) target type
Differentiate.Conventional method needs the priori using expert come design feature, and using small-scale training sample come training airplane
Device Study strategies and methods differentiate region to obtain final testing result.In addition the detection in conventional method and identification are two
It independent process and can not be carried out at the same time, need that the result of target detection is made further to differentiate just using Target Recognition Algorithms
It can obtain final detection and recognition result.
The target detection recognition methods based on deep learning used at present is mainly using detection and identification Integrated Model
Such as faster-rcnn, SSD (Single Shot Multibox Detector), DRBox (Detector using RBox).
These methods directly predict the position of target and classification in input picture using deep learning network.Detection and identification integration
The main process of method is as follows：1) deep learning network inputs image size is fixed, and input picture size is restricted.1)
It using unique characteristic pattern as feature extraction layer, is detected for simple target, the feature extraction layer of use is more single.
2) multiple feature extraction layers are designed in convolutional neural networks, and different size of priori frame is set for multiple feature extraction layers
To detect and identify the position of target and target type in priori frame corresponding region.3) priori frame is the object place of model conjecture
Position matches target using horizontal rectangle frame, and the length-width ratio of priori frame and the setting of scale are relatively easy.4) for
Each input picture carries out the process optimization network of propagated forward and reverse conduction, using stochastic gradient descent, Relu,
The technologies such as Dropout, BatchNorm accelerate the training process of network and improve the robustness of network.
For the detection and identification of high score remote sensing Ship Target, existing image pattern generates, augmentation method is most
It is to be based on small scene image (resolution ratio is in 1920x1080 or less), there is no the sample generating method for large scene image, it can not
Meet the generation of deep learning training sample and augmentation requirement in large scene remote sensing images.
Traditional Target detection and identification method needs to describe and set according to the priori of expert using expert features
Feature is counted, it is subjective to be easy to be influenced by expertise.In addition traditional machine learning classification model cannot be fully sharp
The precision of target detection and target identification is caused to drop it is difficult to obtain satisfactory classifying quality with large-scale data information
It is low.
Existing deep learning method cannot be satisfied mostly towards Ship Target Detection and the requirement of the technology of identification, main to show
：The first, the testing result of these methods describes the position of detection target using horizontal rectangle frame, for mostly towards warship
The description of ship target location is inaccurate, can include more background information for non-horizontal vertical Ship Target.The second, this
The feature extraction layer that a little methods use is more single, and the length-width ratio of priori frame and the setting of scale are also relatively easy.Each feature
Extract layer corresponds to different size of receptive field in artwork, and the feature extraction layer of the different scale of selection can detect different size of
Target, and the corresponding target for detecting different length-width ratios and size of setting of the length-width ratio of priori frame and scale.So existing god
Large scale, the aircraft carrier of more length-width ratios, Ship Target or small scaled target cannot be detected through network algorithm.Third, certain methods are (such as
DRBox) by design mostly towards priori frame come detect in remote sensing images mostly towards target, but it is made using unique characteristic pattern
It is characterized extract layer, does not make full use of the feature of different dimensions.In order to detect the target of different scale, by being carried out to image
The scaling of different scale establishes image pyramid, and is input to the target that different scale in same image is detected in network,
Computational complexity high efficiency lowly takes longer.4th, these methods are limited by input picture size, can not directly be detected big
Target in scene remote sensing images.Deep learning network inputs image size be it is fixed, and large scene remote sensing images size compared with
Greatly, resolution ratio is higher, and Ship Target therein can be scaled a point if directly being zoomed in and out to image, cause detection algorithm
Failure.5th, only simple target is detected, the classification information useless for making full use of different target.In the training process
There is no the losses of use classes to be modified the parameter of model, does not have the ability for identifying different classes of target.
Invention content
The technical problem to be solved by the present invention is to be provided in a kind of high-resolution remote sensing image mostly towards Ship Target Detection
With recognition methods, realize multi-direction, different scale Ship Target fast in the high-resolution remote sensing image under small sample environment
Fast accurate detection identification.
The technical solution adopted by the present invention is, mostly towards Ship Target Detection and identification in the high-resolution remote sensing image
Method, including：
Step 1, multi-angle rectangle frame is marked to the Ship Target in remote sensing images, and preserves the original mark of Ship Target
Information, the original markup information include：The center position of the multi-angle rectangular shaped rim, length and width and towards angle；
Step 2, to carrying out data augmentation processing by the remote sensing images of mark, institute is inputted using deep learning network model
Sliding of the sliding window of size on the remote sensing images by the processing of data augmentation overlappingly is needed, will include effective Ship Target
The image cropping of sliding window get off, and based on the original markup information to the Ship Target in the described image of cutting from
Center position of the dynamic multi-angle rectangle frame for marking the Ship Target in the described image of cutting, length and width and towards angle
Degree, obtains the opposite markup information of the Ship Target；
Step 3, the opposite markup information and depth of the Ship Target in described image and described image based on cutting
Degree learning network model training goes out Ship Target Detection and identification model；
Step 4, the remote sensing images of the sliding window in step 1 of required size are inputted using deep learning network model
On sliding overlappingly, the image cropping of the sliding window comprising effective Ship Target is got off, the image cut is defeated
Enter to predict in the Ship Target Detection and identification model in the image multi-angle rectangle frame position of each Ship Target and
Classification confidence level, according to position of the image on the remote sensing images of step 1 by the image each Ship Target it is more
Angle rectangle frame position is mapped in the remote sensing images of step 1；By identical Ship Target in all images cut
Corresponding multi-angle rectangle frame position and classification confidence level merge to obtain final testing result.
Further, the step 1, including：
To any Ship Target, multi-angle of the line segment as mark is drawn first along a long side of the Ship Target
Then the long side of rectangle frame takes up an official post in another long side of ship and takes some width as mark rectangle frame, and according to the length marked
Multi-angle rectangle frame completion is obtained into original annotation results with width；
The original markup information of each Ship Target includes the following attribute of multi-angle rectangle frame：Multi-angle rectangle frame
The rectangular co-ordinate of center position, the length of multi-angle rectangle frame and wide and multi-angle rectangle frame long side and rectangular coordinate system x-axis
Angle i.e. towards angle, optionally, the original markup information of each target also includes four vertex of multi-angle rectangle frame
Rectangular co-ordinate.
Further, in the step 2, data augmentation processing is carried out to the remote sensing images by mark, including：
To carrying out one or more of operation by the remote sensing images of mark：Rotation, scaling, affine transformation, color tune
Whole, setting contrast, saturation degree adjustment；
The deep learning network model, including：VGG16, GoogLeNet or ResNet.
Further, in the step 2, the judgement side of effective Ship Target whether is included in the image of sliding window
Formula, including：
Judge the area double ratio R=S of any Ship Target j in sliding windowjWhether/S is more than the ratio of setting, if so,
Then judge to include effective Ship Target in the image of sliding window, effective naval vessel mesh is not otherwise included in the image of sliding window
Mark；
Wherein, SjIndicate that region area of the multi-angle rectangle frame of any Ship Target j in sliding window, S indicate to appoint
The area of the multi-angle rectangle frame of one Ship Target j.
Further, the step 3, including：
Step 31：The full articulamentum of deep learning network model is changed to convolutional layer, deep learning network model is to input
The described image of cutting repeat convolution sum down-sampling to extract characteristics of image, obtain the characteristic pattern of various dimensions；According to institute
Feature extraction layer of corresponding size is arranged in the characteristic pattern for stating various dimensions, and the characteristic pattern of each dimension corresponds to a feature extraction layer；
Step 32：According to the size generation pair for each putting the receptive field in the described image of cutting in the feature extraction layer
The priori frame answered, the information of priori frame include the center position of priori frame, length and width and towards angle；By the way that the feature is carried
The characteristics of image in the characteristic pattern of the corresponding various dimensions of layer is taken to be input in the full articulamentum of deep learning network model, to be directed to
Ship Target is predicted included in the receptive field that each pair of point is answered in the feature extraction layer, is predicted the feature and is carried
Take multi-angle rectangle frame position and the classification confidence level of the Ship Target that each pair of point is answered in layer；
Step 33, the information of each priori frame markup information opposite with each Ship Target is matched, for
Mix priori frame, ask the priori frame in the feature extraction layer multi-angle rectangle frame position of corresponding points and classification confidence level with
The loss of the opposite markup information of the corresponding Ship Target of point；Based on the loss, using under backpropagation and stochastic gradient
Drop method is adjusted until the convergence of deep learning network model deep learning network model parameter, obtains final naval vessel mesh
Mark detection and identification model.
Further, in the step 33, the opposite mark of the information of each priori frame and each Ship Target is believed
Breath is matched, including：
The multi-angle rectangle frame gt in the opposite markup information of priori frame p and Ship Target is found out using Monte carlo algorithm
IOU (Intersection Over Union), the wherein definition of IOU is：
Wherein SpIndicate the area of priori frame p, SgtIndicate the multi-angle rectangle frame in the opposite markup information of Ship Target
The area of gt, Sp∩gtIt is polygonal in the opposite markup information of the priori frame p and Ship Target that use Monte carlo algorithm to obtain
The area for spending rectangle frame gt intersecting areas, then defines the multi-angle square in the opposite markup information of priori frame p and Ship Target
The similarity of shape frame gt is：
fs(p, gt)=IOU (p, gt) | cos (θp-θgt)|
Wherein θpFor the angle of the long side and x-axis of priori frame p；θgtFor the multi-angle in the opposite markup information of Ship Target
The long side of rectangle frame gt and the angle of x-axis；
If the similarity f of the multi-angle rectangle frame gt in the opposite markup information of priori frame p and Ship TargetsMore than setting
Numerical value, then it is assumed that the multi-angle rectangle frame gt in priori frame p markup informations opposite with Ship Target is matched.
Further, in the step 33, the multi-angle square of priori frame corresponding points in the feature extraction layer is sought
The loss of the opposite markup information of shape frame position and classification confidence level Ship Target corresponding with the point, including：
If the priori frame in the feature extraction layer multi-angle rectangle frame position of corresponding points and classification confidence level with should
The loss of the opposite markup information of the corresponding Ship Target of point meets following loss function：
Wherein, N indicates the priori frame number of the multi-angle rectangle frame successful match in the opposite markup information with Ship Target
Mesh, z indicate that the set of the image after input cutting, c indicate the corresponding class number's set of sample, and α is the constant of setting, l tables
Show that offset of the multi-angle rectangle frame of prediction relative to priori frame, g indicate polygonal in the opposite markup information of Ship Target
Spend rectangle frame, LconfIndicate the position loss of multi-angle rectangle frame, LlocIt is the loss of classification confidence level；
The loss of classification confidence level is soffmax loss functions；
The position loss of multi-angle rectangle frame is defined as follows：
Wherein, smoothL1For a known function,Indicate each priori
The length and width in horizontal and vertical direction of frame,For prediction multi-angle rectangle frame m values relative to i-th of priori frame
M values offset,For the m values in the opposite markup information of j-th of the Ship Target to match with i-th of priori frame；
Then have：
WhereinIndicate the cx values of i-th of priori frame；
For the cx values in the opposite markup information of j-th of Ship Target；
Indicate the cy values of i-th of priori frame；
For the cy values in the opposite markup information of j-th of Ship Target；
For the w values in the opposite markup information of j-th of Ship Target；
For the h values in the opposite markup information of j-th of Ship Target；
Indicate the θ values of i-th of priori frame；
For the θ values in the opposite markup information of j-th of Ship Target；
It is indicator function, the multi-angle rectangle in the opposite markup information of i-th of priori frame and j-th of Ship Target
Its value is 1 when frame matches and corresponds to the classification of kth class, is otherwise 0；Pos indicates all and is formed with the matched priori frame of actual value
Set；xmax, xmin, ymax, yminThe maximum abscissa of priori four apex coordinates of frame, minimum value and ordinate are indicated respectively
Maximum, minimum value；Cx, cy, w, h, θ indicate respectively the corresponding central point rectangular co-ordinate of multi-angle rectangle frame, length and width and
The angle of long side and x-axis.
Further, the step 3 further includes：
Step 34, the described image of cutting is input to the Ship Target Detection and identification model obtains the image not
With the predicted value of the multi-angle rectangle frame position and classification confidence level each put in feature extraction layer, calculated using non-maxima suppression
Method carries out the predicted value of the multi-angle rectangle frame position and classification confidence level each put in the different characteristic extract layer of the image
Duplicate removal obtains the multi-angle rectangle frame position and the classification confidence level that finally detect the Ship Target in the image identified.
Further, in the step 34, using non-maxima suppression algorithm to the different characteristic extract layer of the image
In the predicted value of multi-angle rectangle frame position and classification confidence level each put carry out duplicate removal, obtain finally detecting identify should
The multi-angle rectangle frame position of Ship Target in image and classification confidence level, including：
By the prediction of the multi-angle rectangle frame position and classification confidence level each put in the different characteristic extract layer of the image
Value arranges from high to low according to classification confidence level, and each point of traversal executes following operation successively：By the multi-angle rectangle of current point
The multi-angle rectangle frame of frame other points low with the confidence level than current point is matched, the multi-angle rectangle of removal and current point
The IOU of frame is more than setting numerical value and belongs to other same category of points with current point, and the warship in the image is obtained after the completion of traversal
The multi-angle rectangle frame position of ship target and classification confidence level.
Further, in the step 4, by the corresponding multi-angle of identical Ship Target in all images cut
Rectangle frame position and classification confidence level merge to obtain final testing result, including：
Using non-maxima suppression algorithm to the corresponding multi-angle square of identical Ship Target in all images cut
Shape frame position carries out duplicate removal, the multi-angle rectangle frame position of the Ship Target finally detected；
The corresponding classification confidence level of identical Ship Target in all images cut is arranged from high to low, output is set
The classification confidence level result of fixed number amount.
Using above-mentioned technical proposal, the present invention at least has following advantages：
Mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image of the present invention, design and Implement multiple
Under miscellaneous scene, in large scale environment multi-direction Ship Target Detection with identification network structure, based on convolutional neural networks it is automatic
The various dimensions features such as shallow-layer and the deep layer of Ship Target are excavated, are realized more in the high-resolution remote sensing image under small sample environment
The quick accurate detection identification in direction, different scale Ship Target.The present invention is compared to the speed that traditional object detection method detects
Faster, testing result is more accurate for degree.
Description of the drawings
Fig. 1 be first embodiment of the invention high-resolution remote sensing image in mostly towards Ship Target Detection and recognition methods
Flow chart；
Fig. 2 be second embodiment of the invention high-resolution remote sensing image in mostly towards Ship Target Detection and recognition methods
Flow chart；
Fig. 3 is the high-resolution remote sensing image of second embodiment of the invention marked；
Fig. 4 is the schematic diagram of the Monte carlo algorithm of second embodiment of the invention.
Specific implementation mode
Further to illustrate the present invention to reach the technological means and effect that predetermined purpose is taken, below in conjunction with attached drawing
And preferred embodiment, the present invention is described in detail as after.
The purpose of the embodiment of the present invention is to solve the problems, such as to carry mostly towards Ship Target Detection and identification in remote sensing images
Go out in a kind of remote sensing images based on deep learning mostly towards Ship Target Detection and recognition methods.This method is based on deep learning
Thought provides large-scale Target detection and identification training sample construction method first, compensates for current naval vessel sample size not
The defect of foot fully meets the sample needs for deep learning target detection, the training of identification model and test, is depth
Learning method is applied to detection identification and lays a good foundation；Secondly, it is realized based on small sample and Ship Target Detection is identified, significantly subtracted
Complexity and training amount are trained less, and the strong requirement of Ship Recognition timeliness is met while improving working efficiency；The
Three, it proposes mostly to propose Ship Target pair in high-precision remote sensing images towards the deep learning network structure of Target detection and identification
Should mostly towards the position of boundary rectangle frame and target category prediction technique, solve current naval vessel large scale, aspect than it is more
The detection of direction brought and identification problem；4th, it is based on NMS (non-maxima suppression) algorithm, proposes the same target of more remote sensing
Size provides technical support towards the naval vessel detection recognition method of variation for Ship Target real-time tracking.
First embodiment of the invention, mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image,
As shown in Figure 1, including step in detail below：
Step 1, multi-angle rectangle frame is marked to the Ship Target in remote sensing images, and preserves the original mark of Ship Target
Information, the original markup information include：Center position, length and width of the multi-angle rectangular shaped rim in the remote sensing images
With towards angle；
Specifically, the step 1, including：
To any Ship Target, multi-angle of the line segment as mark is drawn first along a long side of the Ship Target
Then the long side of rectangle frame takes up an official post in another long side of ship and takes some width as mark rectangle frame, and according to the length marked
Multi-angle rectangle frame completion is obtained into original annotation results with width；
The original markup information of each Ship Target includes the following attribute of multi-angle rectangle frame：Multi-angle rectangle frame
The rectangular co-ordinate of center position, the length of multi-angle rectangle frame and wide and multi-angle rectangle frame long side and rectangular coordinate system x-axis
Angle i.e. towards angle, optionally, the original markup information of each target also includes four vertex of multi-angle rectangle frame
Rectangular co-ordinate.The markup information of all targets is stored in local in the form of an xml-file in remote sensing images.
Step 2, to carrying out data augmentation processing by the remote sensing images of mark, institute is inputted using deep learning network model
Sliding of the sliding window of size on the remote sensing images by the processing of data augmentation overlappingly is needed, will include effective Ship Target
The image cropping of sliding window get off, and based on the original markup information to the Ship Target in the described image of cutting from
Center position of the dynamic multi-angle rectangle frame for marking the Ship Target in the described image of cutting, length and width and towards angle
Degree, obtains the opposite markup information of the Ship Target；Multi-angle rectangle frame in the described image of cutting towards angle with
The multi-angle rectangular shaped rim is identical towards angle in the remote sensing images.
Specifically, in the step 2, data augmentation processing is carried out to the remote sensing images by mark, including：
To carrying out one or more of operation by the remote sensing images of mark：Rotation, scaling, affine transformation, color tune
Whole, setting contrast, saturation degree adjustment；
The deep learning network model, including：VGG16, GoogLeNet or ResNet.
In the step 2, the judgment mode of effective Ship Target whether is included in the image of sliding window, including：
Judge the area double ratio R=S of any Ship Target j in sliding windowjWhether/S is more than the ratio of setting, if so,
Then judge to include effective Ship Target in the image of sliding window, effective naval vessel mesh is not otherwise included in the image of sliding window
Mark；
Wherein, SjIndicate that region area of the multi-angle rectangle frame of any Ship Target j in sliding window, S indicate to appoint
The area of the multi-angle rectangle frame of one Ship Target j.
Step 3, the opposite markup information and depth of the Ship Target in described image and described image based on cutting
Degree learning network model training goes out Ship Target Detection and identification model；
Specifically, the step 3, including：
Step 31：The full articulamentum of deep learning network model is changed to convolutional layer, deep learning network model is to input
The described image of cutting repeat convolution sum down-sampling to extract characteristics of image, obtain the characteristic pattern of various dimensions；According to institute
Feature extraction layer of corresponding size is arranged in the characteristic pattern for stating various dimensions, and the characteristic pattern of each dimension corresponds to a feature extraction layer；
Step 32：According to the size generation pair for each putting the receptive field in the described image of cutting in the feature extraction layer
The priori frame answered, the information of priori frame include the center position of priori frame, length and width and towards angle, different feature extraction layers
In point receptive field in the described image of cutting of different sizes, the setting scale of the priori frame to match with the receptive field
It differs；By the way that the characteristics of image in the characteristic pattern of the corresponding various dimensions of the feature extraction layer is input to deep learning network
In the full articulamentum of model, with Ship Target included in the receptive field answered for each pair of point in the feature extraction layer into
Row prediction, predicts multi-angle rectangle frame position and the classification confidence of the Ship Target that each pair of point is answered in the feature extraction layer
Degree；
Step 33, the information of each priori frame markup information opposite with each Ship Target is matched, for
Mix priori frame, ask the priori frame in the feature extraction layer multi-angle rectangle frame position of corresponding points and classification confidence level with
The loss of the opposite markup information of the corresponding Ship Target of point；Based on the loss, using under backpropagation and stochastic gradient
Drop method is adjusted until the convergence of deep learning network model deep learning network model parameter, obtains final naval vessel mesh
Mark detection and identification model；
Step 34, the described image of cutting is input to the Ship Target Detection and identification model obtains the image not
With the predicted value of the multi-angle rectangle frame position and classification confidence level each put in feature extraction layer, calculated using non-maxima suppression
Method carries out the predicted value of the multi-angle rectangle frame position and classification confidence level each put in the different characteristic extract layer of the image
Duplicate removal obtains the multi-angle rectangle frame position and the classification confidence level that finally detect the Ship Target in the image identified.
Further, in the step 33, the opposite mark of the information of each priori frame and each Ship Target is believed
Breath is matched, including：
The multi-angle rectangle frame gt in the opposite markup information of priori frame p and Ship Target is found out using Monte carlo algorithm
IOU (Intersection Over Union), the wherein definition of IOU is：
Wherein SpIndicate the area of priori frame p, SgtIndicate the multi-angle rectangle frame in the opposite markup information of Ship Target
The area of gt, Sp∩gtIt is polygonal in the opposite markup information of the priori frame p and Ship Target that use Monte carlo algorithm to obtain
The area for spending rectangle frame gt intersecting areas, then defines the multi-angle square in the opposite markup information of priori frame p and Ship Target
The similarity of shape frame gt is：
fs(p, gt)=IOU (p, gt) | cos (θp-θgt)|
Wherein θpFor the angle of the long side and x-axis of priori frame p；θgtFor the multi-angle in the opposite markup information of Ship Target
The long side of rectangle frame gt and the angle of x-axis；
If the similarity f of the multi-angle rectangle frame gt in the opposite markup information of priori frame p and Ship TargetsMore than setting
Numerical value, then it is assumed that the multi-angle rectangle frame gt in priori frame p markup informations opposite with Ship Target is matched.
In the step 33, ask the priori frame in the feature extraction layer multi-angle rectangle frame position of corresponding points and
The loss of the opposite markup information of classification confidence level Ship Target corresponding with the point, including：
If the priori frame in the feature extraction layer multi-angle rectangle frame position of corresponding points and classification confidence level with should
The loss of the opposite markup information of the corresponding Ship Target of point meets following loss function：
Wherein, N indicates the priori frame number of the multi-angle rectangle frame successful match in the opposite markup information with Ship Target
Mesh, z indicate that the set of the image after input cutting, c indicate the corresponding class number's set of sample, and α is the constant of setting, l tables
Show that offset of the multi-angle rectangle frame of prediction relative to priori frame, g indicate polygonal in the opposite markup information of Ship Target
Spend rectangle frame, LconfIndicate the position loss of multi-angle rectangle frame, LlocIt is the loss of classification confidence level；
The loss of classification confidence level is softmax loss functions；
The position loss of multi-angle rectangle frame is defined as follows：
Wherein, smoothL1For a known function,Indicate each priori
The length and width in horizontal and vertical direction of frame,For prediction multi-angle rectangle frame m values relative to i-th of priori frame
M values offset,For the m values in the opposite markup information of j-th of the Ship Target to match with i-th of priori frame；
Then have：
WhereinIndicate the cx values of i-th of priori frame；
For the cx values in the opposite markup information of j-th of Ship Target；
Indicate the cy values of i-th of priori frame；
For the cy values in the opposite markup information of j-th of Ship Target；
For the w values in the opposite markup information of j-th of Ship Target；
For the h values in the opposite markup information of j-th of Ship Target；
Indicate the θ values of i-th of priori frame；
For the θ values in the opposite markup information of j-th of Ship Target；
It is indicator function, the multi-angle rectangle in the opposite markup information of i-th of priori frame and j-th of Ship Target
Its value is 1 when frame matches and corresponds to the classification of kth class, is otherwise 0；Pos indicates all and is formed with the matched priori frame of actual value
Set；xmax, xmin, ymax, yminThe maximum abscissa of priori four apex coordinates of frame, minimum value and ordinate are indicated respectively
Maximum, minimum value；Cx, cy, w, h, θ indicate respectively the corresponding central point rectangular co-ordinate of multi-angle rectangle frame, length and width and
The angle of long side and x-axis.In the step 34, using non-maxima suppression algorithm to every in the different characteristic extract layer of the image
The multi-angle rectangle frame position of a point and the predicted value of classification confidence level carry out duplicate removal, obtain finally detecting the image identified
In Ship Target multi-angle rectangle frame position and classification confidence level, including：
By the prediction of the multi-angle rectangle frame position and classification confidence level each put in the different characteristic extract layer of the image
Value arranges from high to low according to classification confidence level, and each point of traversal executes following operation successively：By the multi-angle rectangle of current point
The multi-angle rectangle frame of frame other points low with the confidence level than current point is matched, the multi-angle rectangle of removal and current point
The IOU of frame is more than setting numerical value and belongs to other same category of points with current point, and the warship in the image is obtained after the completion of traversal
The multi-angle rectangle frame position of ship target and classification confidence level.
Step 4, the remote sensing images of the sliding window in step 1 of required size are inputted using deep learning network model
On sliding overlappingly, the image cropping of the sliding window comprising effective Ship Target is got off, the image cut is defeated
Enter to predict in the Ship Target Detection and identification model in the image multi-angle rectangle frame position of each Ship Target and
Classification confidence level, according to position of the image on the remote sensing images of step 1 by the image each Ship Target it is more
Angle rectangle frame position is mapped in the remote sensing images of step 1；By identical Ship Target in all images cut
Corresponding multi-angle rectangle frame position and classification confidence level merge to obtain final testing result.
Specifically, in the step 4, by the corresponding multi-angle square of identical Ship Target in all images cut
Shape frame position and classification confidence level merge to obtain final testing result, including：
Using non-maxima suppression algorithm to the corresponding multi-angle square of identical Ship Target in all images cut
Shape frame position carries out duplicate removal, the multi-angle rectangle frame position of the Ship Target finally detected；
The corresponding classification confidence level of identical Ship Target in all images cut is arranged from high to low, output is set
The classification confidence level result of fixed number amount.
The embodiment of the present invention solves the technical issues of the following aspects：
1) the few problem of naval vessel remote sensing images training sample.In the case where the quantity for having marked remote sensing images is few, this hair
Bright can provide meets the required sample size of deep learning model training, to which support trains reliable deep learning mesh
Mark detection and identification model.
2) mostly towards the inaccurate problem of naval vessel testing result.The present invention can effectively solve the Ship Target description of mostly direction
In the interference that generates of background information and noise information, realize mostly accurately identifying towards naval vessel in complex scene.
3) for the bad problem of small scale and large scale Ship Target Detection effect.Small scaled target corresponds to nobby
Equal Ship Targets, and large scale target corresponds to the Ship Targets such as aircraft carrier, it is accurate that conventional method detects nobby or aircraft carrier
Rate is low.
4) the undetectable problem of Ship Target in high-resolution large scene remote sensing images.High score large scene remote sensing images by
In it includes scene it is larger and resolution ratio is higher, target information therein it is smaller to compression of images it will cause target disappearance,
If directly zooming in and out and being input in neural network to image, the target in original image can not be detected and identified.
Second embodiment of the invention, the present embodiment are on the basis of the above embodiments, one to be introduced in conjunction with attached drawing 2~4
The application example of the present invention.
A kind of high-resolution remote sensing image based on deep learning of the embodiment of the present invention mostly towards Ship Target Detection and
Recognition methods, main flow is as shown in Fig. 2, specific steps include：
Step S1, based on deep learning mostly towards the structure of Ship Target Detection and identification model, including：
(1) mark of large scene high-definition remote sensing data, for each warship in the high-resolution remote sensing image of input
Ship target takes the mode of " line-of-sight course " to carry out mostly marking towards rectangle frame to it.So-called " line-of-sight course " is exactly for a warship
Ship target draws long side of the line segment as mark rectangle frame, then in ship first along a long side of the Ship Target
Another long side, which is taken up an official post, takes some height as mark rectangle frame, and is obtained rectangle frame completion finally according to the length and height that have marked
Annotation results.Remote sensing images after mark are as shown in Figure 3.
The markup information of each target includes it mostly towards the following attribute of boundary rectangle frame：Rectangle frame center point
The x set, y-coordinate, the length of rectangle frame and wide and rectangle frame long side and x-axis angle, the x on four vertex of rectangle frame, y-coordinate.
The markup information of all targets is stored in local in the form of an xml-file in each image.
(2) data augmentation and data are cut
1. carrying out multi-angle rotary to the large scene remote sensing images marked, generally one is rotated for ladder with 30 degree or 90 degree
Week, and use the white space of the postrotational extraneous rectangle frame of image correspondence of filled black.
2. using the input size of deep learning target detection model to be trained as the size of sliding window, using containing 1/3 weight
Folded mode is scanned postrotational image.If including effective target in current sliding window mouth, by current sliding window
Mouth and its actual value of corresponding target preserve.What the actual value of the target of the embodiment of the present invention referred to is exactly the phase in first embodiment
To markup information.
The area double ratio of Ship Target is in definition sliding window：R=Sj/ S, wherein SjIt indicates mostly towards outside Ship Target
Region area of the rectangle frame in sliding window is connect, S indicates the area of Ship Target boundary rectangle frame.If a Ship Target
Area double ratio be more than 0.5, it is considered that the Ship Target is that have a effective target.
During the cutting process, it needs the target to area double ratio less than 1 to carry out stipulations and is working as front slide to obtain the target
Boundary rectangle frame in window.When to external rectangle frame stipulations, the method for taking " protecting short cut out to grow " is preferentially protected when cutting
The short side in original boundary rectangle frame is stayed, and long side is cut to obtain new boundary rectangle frame.
3. the target image after cutting is overturn, luminance contrast adjustment etc. operations, to training and test data into
One step augmentation.
(3) design and realization of the Ship Target Detection based on deep learning and identification model, network structure such as Fig. 4 institutes
Show, is implemented as follows：
1. being obtained by convolution sum down-sampling as basic network using the VGG16 networks of 300x300 or 512x512
The characteristic pattern of input picture different latitude.
2. choose VGG16 basic networks in Conv3_3, Conv4_3, FC7, Conv6_2, Conv7_2, Conv8_2,
As feature extraction layer, each different feature extraction layer generates different scale, different length-width ratios, no by Conv9_2, Conv10_2
With the priori frame of angle.Two different scales of wherein same feature extraction layer indicate the length of long side, different characteristic extraction
The specifying information that layer generates priori frame is as shown in table 1 below：
The priori box properties information that 1 different characteristic figure of table generates
3. for each training sample, the IOU of priori frame and actual value is calculated using Monte carlo algorithm.Then institute is calculated
There is priori frame and marks the similarity f of actual valuesIf fs＞ 0.5 then thinks that priori frame is matched with actual value.And according to matching
On priori frame and true value information, calculate network output position loss and classification loss, and using backpropagation to network
Parameter is adjusted.
The wherein calculation of Monte carlo algorithm, steps are as follows：
1) actual value S is calculated according to the length and wide information of actual value and priori framegtWith the area S of priori framep。
2) if the area of the corresponding boundary rectangle frame of priori frame is less than 10000 pixels, the boundary rectangle frame is traversed successively
Interior all points simultaneously find out Sp∩gt.Otherwise random 10000 points of uniform sampling inside the corresponding boundary rectangle frame of priori frame, so
The number n at intersecting area midpoint is counted afterwardsp∩gt, obtainWherein SpbIndicate that priori frame corresponds to external square
The area of shape frame.
4. being constantly modified until network convergence to network parameter or reaching specified iterations.
5. after the completion of training, different feature extraction layers is predicted using NMS (non-maxima suppression) algorithms polygonal
Degree rectangle frame is refined, and the prediction result of the final prediction result and target category information of target location is obtained.
This completes based on deep learning mostly towards the structure of Ship Target Detection and identification model.
Step S2, the object detection and recognition of large scene high-resolution remote sensing image.
(1) input large scene remote sensing images have been sliced with having carried out overlapping region, slice size is target detection and knowledge
The size of other network inputs image, to ensure that in the small figure after Ship Target to be detected is at least sliced at one be complete.
(2) each slice is input in target detection neural network and predicts each corresponding position of target in current slice
It sets, and the location information of each target is mapped in former remote sensing images by the location information according to slice in former remote sensing images.
(3) testing result of all slices is merged, and using NMS algorithms to final testing result duplicate removal, to obtain
Final testing result.
Since then, it is achieved that the high-resolution remote sensing image based on deep learning mostly towards the method for Ship Target Detection.
By the explanation of specific implementation mode, should can to the present invention for reach technological means that predetermined purpose is taken and
Effect is able to more go deep into and specifically understand, however appended diagram is only to provide reference and description and is used, and is not used for originally
Invention limits.
Claims (10)
1. mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image, which is characterized in that including：
Step 1, multi-angle rectangle frame is marked to the Ship Target in remote sensing images, and preserves the original mark letter of Ship Target
Breath, the original markup information include：Center position of the multi-angle rectangular shaped rim in the remote sensing images, length and width and
Towards angle；
Step 2, to carrying out data augmentation processing by the remote sensing images of mark, using big needed for the input of deep learning network model
Sliding of the small sliding window on the remote sensing images by the processing of data augmentation overlappingly, will include the cunning of effective Ship Target
The image cropping of dynamic window gets off, and is marked automatically to the Ship Target in the described image of cutting based on the original markup information
It notes center position of the multi-angle rectangle frame of the Ship Target in the described image of cutting, length and width and towards angle, obtains
To the opposite markup information of the Ship Target；
Step 3, the opposite markup information and depth of the Ship Target in described image and described image based on cutting
It practises network model and trains Ship Target Detection and identification model；
Step 4, the sliding window that required size is inputted using deep learning network model is heavy on the remote sensing images of step 1
The sliding on folded ground, the image cropping of the sliding window comprising effective Ship Target is got off, and the image cut is inputted institute
It states and predicts the multi-angle rectangle frame position of each Ship Target and classification in the image in Ship Target Detection and identification model
Confidence level, according to position of the image on the remote sensing images of step 1 by the multi-angle of each Ship Target in the image
Rectangle frame position is mapped in the remote sensing images of step 1；Identical Ship Target in all images cut is corresponded to
Multi-angle rectangle frame position and classification confidence level merge to obtain final testing result.
2. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 1
Sign is, the step 1, including：
To any Ship Target, multi-angle rectangle of the line segment as mark is drawn first along a long side of the Ship Target
Then the long side of frame takes up an official post in another long side of ship and takes some width as mark rectangle frame, and according to the length and width marked
Multi-angle rectangle frame completion is obtained into original annotation results；
The original markup information of each Ship Target includes the following attribute of multi-angle rectangle frame：Multi-angle rectangle frame center
The rectangular co-ordinate of point position, the length of multi-angle rectangle frame and wide and multi-angle rectangle frame long side and rectangular coordinate system x-axis folder
Angle is i.e. towards angle, and optionally, the original markup information of each target also includes the right angle on four vertex of multi-angle rectangle frame
Coordinate.
3. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 1
Sign is, in the step 2, data augmentation processing is carried out to the remote sensing images by mark, including：
To carrying out one or more of operation by the remote sensing images of mark：Rotation, scaling, affine transformation, color adjust, are right
Than degree adjustment, saturation degree adjustment；
The deep learning network model, including：VGG16, GoogLeNet or ResNet.
4. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 1
Sign is, in the step 2, the judgment mode of effective Ship Target whether is included in the image of sliding window, including：
Judge the area double ratio R=S of any Ship Target j in sliding windowjWhether/S is more than the ratio of setting, if so, judgement
Include effective Ship Target in the image of sliding window, effective Ship Target is not otherwise included in the image of sliding window；
Wherein, SjIndicate that region area of the multi-angle rectangle frame of any Ship Target j in sliding window, S indicate any naval vessel
The area of the multi-angle rectangle frame of target j.
5. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 1
Sign is, the step 3, including：
Step 31：The full articulamentum of deep learning network model is changed to convolutional layer, sanction of the deep learning network model to input
The described image cut repeats convolution sum down-sampling to extract characteristics of image, obtains the characteristic pattern of various dimensions；According to described more
Feature extraction layer of corresponding size is arranged in the characteristic pattern of dimension, and the characteristic pattern of each dimension corresponds to a feature extraction layer；
Step 32：It is generated according to the size for each putting the receptive field in the described image of cutting in the feature extraction layer corresponding
Priori frame, the information of priori frame include the center position of priori frame, length and width and towards angle；By by the feature extraction layer
Characteristics of image in the characteristic pattern of corresponding various dimensions is input in the full articulamentum of deep learning network model, described to be directed to
Ship Target is predicted included in the receptive field that each pair of point is answered in feature extraction layer, predicts the feature extraction layer
The multi-angle rectangle frame position for the Ship Target that middle each pair of point is answered and classification confidence level；
Step 33, the information of each priori frame markup information opposite with each Ship Target is matched, for matching
Priori frame asks the priori frame multi-angle rectangle frame position of corresponding points and classification confidence level and the point in the feature extraction layer
The loss of the opposite markup information of corresponding Ship Target；Based on the loss, backpropagation and stochastic gradient descent side are used
Method, which is adjusted deep learning network model parameter, obtains final Ship Target inspection until the convergence of deep learning network model
Survey and identification model.
6. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 5
Sign is, in the step 33, the information of each priori frame markup information opposite with each Ship Target is matched,
Including：
Find out the multi-angle rectangle frame gt's in the opposite markup information of priori frame p and Ship Target using Monte carlo algorithm
IOU (Intersection Over Union), the wherein definition of IOU is：
Wherein SpIndicate the area of priori frame p, SgtIndicate the multi-angle rectangle frame gt's in the opposite markup information of Ship Target
Area, Sp∩gtFor the multi-angle square in the opposite markup information of the priori frame p and Ship Target that use Monte carlo algorithm to obtain
The area of shape frame gt intersecting areas, then defines the multi-angle rectangle frame in the opposite markup information of priori frame p and Ship Target
The similarity of gt is：
fs(p, gt)=IOU (p, gt) | cos (θp-θgt)|
Wherein θpFor the angle of the long side and x-axis of priori frame p；θgtFor the multi-angle rectangle in the opposite markup information of Ship Target
The long side of frame gt and the angle of x-axis；
If the similarity f of the multi-angle rectangle frame gt in the opposite markup information of priori frame p and Ship TargetsMore than the number of setting
Value, then it is assumed that the multi-angle rectangle frame gt in priori frame p markup informations opposite with Ship Target is matched.
7. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 5
Sign is, in the step 33, ask the priori frame in the feature extraction layer multi-angle rectangle frame position of corresponding points and
The loss of the opposite markup information of classification confidence level Ship Target corresponding with the point, including：
If the priori frame multi-angle rectangle frame position of corresponding points and classification confidence level and the point pair in the feature extraction layer
The loss of the opposite markup information for the Ship Target answered meets following loss function：
Wherein, N indicates the priori frame number of the multi-angle rectangle frame successful match in the opposite markup information with Ship Target, z
Indicate that the set of the image after input is cut, c indicate that the corresponding class number's set of sample, α are the constant of setting, l indicates pre-
Offset of the multi-angle rectangle frame of survey relative to priori frame, g indicate the multi-angle square in the opposite markup information of Ship Target
Shape frame, LconfIndicate the position loss of multi-angle rectangle frame, LlocIt is the loss of classification confidence level；
The loss of classification confidence level is softmax loss functions；
The position loss of multi-angle rectangle frame is defined as follows：
Wherein, smoothL1For a known function,Indicate each priori frame
The length and width in horizontal and vertical direction,For m value of the m values relative to i-th of priori frame of the multi-angle rectangle frame of prediction
Offset,For the m values in the opposite markup information of j-th of the Ship Target to match with i-th of priori frame；
Then have：
WhereinIndicate the cx values of i-th of priori frame；
For the cx values in the opposite markup information of j-th of Ship Target；
Indicate the cy values of i-th of priori frame；
For the cy values in the opposite markup information of j-th of Ship Target；
For the w values in the opposite markup information of j-th of Ship Target；
For the h values in the opposite markup information of j-th of Ship Target；
Indicate the θ values of i-th of priori frame；
For the θ values in the opposite markup information of j-th of Ship Target；
It is indicator function, the multi-angle rectangle frame in the opposite markup information of i-th of priori frame and j-th of Ship Target
Its value is 1 when matching and corresponding to the classification of kth class, is otherwise 0；Pos indicates all collection with the matched priori frame composition of actual value
It closes；xmax, xmin, ymax, yminThe maximum abscissa of priori four apex coordinates of frame, minimum value and ordinate are indicated respectively most
Greatly, minimum value；Cx, cy, w, h, θ indicate the corresponding central point rectangular co-ordinate of multi-angle rectangle frame, length and width and long side respectively
With the angle of x-axis.
8. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 5
Sign is that the step 3 further includes：
Step 34, the described image of cutting is input to the Ship Target Detection and identification model and obtains the Bu Tong special of the image
The predicted value of the multi-angle rectangle frame position and classification confidence level each put in sign extract layer, uses non-maxima suppression algorithm pair
The predicted value of the multi-angle rectangle frame position and classification confidence level each put in the different characteristic extract layer of the image carries out duplicate removal,
Obtain finally detecting multi-angle rectangle frame position and the classification confidence level of the Ship Target in the image identified.
9. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 8
Sign is, in the step 34, using non-maxima suppression algorithm to each putting in the different characteristic extract layer of the image
The predicted value of multi-angle rectangle frame position and classification confidence level carries out duplicate removal, obtains finally detecting the warship in the image identified
The multi-angle rectangle frame position of ship target and classification confidence level, including：
The predicted value of the multi-angle rectangle frame position each put in the different characteristic extract layer of the image and classification confidence level is pressed
It is arranged from high to low according to classification confidence level, each point of traversal executes following operation successively：By the multi-angle rectangle frame of current point with
The multi-angle rectangle frame of other points lower than the confidence level of current point is matched, removal and the multi-angle rectangle frame of current point
IOU is more than setting numerical value and belongs to other same category of points with current point, and the naval vessel mesh in the image is obtained after the completion of traversal
Target multi-angle rectangle frame position and classification confidence level.
10. special mostly towards Ship Target Detection and recognition methods in high-resolution remote sensing image according to claim 1
Sign is, in the step 4, by identical Ship Target corresponding multi-angle rectangle frame position in all images cut
Merge to obtain final testing result with classification confidence level, including：
Using non-maxima suppression algorithm to the corresponding multi-angle rectangle frame of identical Ship Target in all images cut
Position carries out duplicate removal, the multi-angle rectangle frame position of the Ship Target finally detected；
The corresponding classification confidence level of identical Ship Target in all images cut is arranged from high to low, output setting number
The classification confidence level result of amount.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810075285.3A CN108319949A (en) | 2018-01-26 | 2018-01-26 | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810075285.3A CN108319949A (en) | 2018-01-26 | 2018-01-26 | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108319949A true CN108319949A (en) | 2018-07-24 |
Family
ID=62888527
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810075285.3A Pending CN108319949A (en) | 2018-01-26 | 2018-01-26 | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108319949A (en) |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109166106A (en) * | 2018-08-02 | 2019-01-08 | 山东大学 | A kind of target detection aligning method and apparatus based on sliding window |
| CN109165603A (en) * | 2018-08-28 | 2019-01-08 | 中国科学院遥感与数字地球研究所 | A kind of Ship Detection and device |
| CN109255317A (en) * | 2018-08-31 | 2019-01-22 | 西北工业大学 | A kind of Aerial Images difference detecting method based on dual network |
| CN109344717A (en) * | 2018-09-01 | 2019-02-15 | 哈尔滨工程大学 | A kind of deep-sea target on-line checking recognition methods of multi-threshold dynamic statistics |
| CN109523015A (en) * | 2018-11-09 | 2019-03-26 | 上海海事大学 | Image processing method in a kind of neural network |
| CN109711427A (en) * | 2018-11-19 | 2019-05-03 | 深圳市华尊科技股份有限公司 | Object detection method and Related product |
| CN109800755A (en) * | 2018-12-14 | 2019-05-24 | 中国科学院深圳先进技术研究院 | A kind of remote sensing image small target detecting method based on Analysis On Multi-scale Features |
| CN109800637A (en) * | 2018-12-14 | 2019-05-24 | 中国科学院深圳先进技术研究院 | A kind of remote sensing image small target detecting method |
| CN109800735A (en) * | 2019-01-31 | 2019-05-24 | 中国人民解放军国防科技大学 | Accurate detection and segmentation method for ship target |
| CN109919183A (en) * | 2019-01-24 | 2019-06-21 | 北京大学 | A kind of image-recognizing method based on small sample, device, equipment and storage medium |
| CN109948415A (en) * | 2018-12-30 | 2019-06-28 | 中国科学院软件研究所 | Remote sensing image object detection method based on filtering background and scale prediction |
| CN110008863A (en) * | 2019-03-22 | 2019-07-12 | 哈尔滨工业大学 | Efficient multi-scale sampling method based on high-resolution bridge area ship image |
| CN110147837A (en) * | 2019-05-14 | 2019-08-20 | 中国电子科技集团公司第二十八研究所 | The intensive object detection method of any direction, system and the equipment focused based on feature |
| CN110223302A (en) * | 2019-05-08 | 2019-09-10 | 华中科技大学 | A kind of naval vessel multi-target detection method extracted based on rotary area |
| CN110335238A (en) * | 2019-05-08 | 2019-10-15 | 吉林大学 | A kind of Automotive Paint Film Properties defect recognition system constituting method of deep learning |
| CN110443259A (en) * | 2019-07-29 | 2019-11-12 | 中科光启空间信息技术有限公司 | A method of extracting sugarcane from intermediate resolution remote sensing image |
| CN110689021A (en) * | 2019-10-17 | 2020-01-14 | 哈尔滨理工大学 | Real-time target detection method in low-visibility environment based on deep learning |
| CN110728307A (en) * | 2019-09-20 | 2020-01-24 | 天津大学 | Method for realizing small sample character recognition of X-ray image by self-generating data set and label |
| CN110796185A (en) * | 2019-10-17 | 2020-02-14 | 北京爱数智慧科技有限公司 | Method and device for detecting image annotation result |
| CN110853295A (en) * | 2019-11-12 | 2020-02-28 | 江西赣鄱云新型智慧城市技术研究有限公司 | High-altitude parabolic early warning method and device |
| CN110942455A (en) * | 2019-11-26 | 2020-03-31 | 南方电网数字电网研究院有限公司 | Method and device for detecting missing of cotter pin of power transmission line and computer equipment |
| CN111079688A (en) * | 2019-12-27 | 2020-04-28 | 中国电子科技集团公司第十五研究所 | Living body detection method based on infrared image in face recognition |
| CN111191566A (en) * | 2019-12-26 | 2020-05-22 | 西北工业大学 | Optical remote sensing image multi-target detection method based on pixel classification |
| CN111354378A (en) * | 2020-02-12 | 2020-06-30 | 北京声智科技有限公司 | Voice endpoint detection method, device, equipment and computer storage medium |
| CN111368625A (en) * | 2019-11-08 | 2020-07-03 | 深圳北航新兴产业技术研究院 | Pedestrian target detection method based on cascade optimization |
| CN111382762A (en) * | 2018-12-28 | 2020-07-07 | 同方威视技术股份有限公司 | Empty box identification method and system |
| CN111401302A (en) * | 2020-04-07 | 2020-07-10 | 中国人民解放军海军航空大学 | Remote sensing image ship target integrated detection and fine-grained identification method |
| CN111524095A (en) * | 2020-03-24 | 2020-08-11 | 西安交通大学 | Target detection method for rotating object |
| CN111723809A (en) * | 2019-06-10 | 2020-09-29 | 阿里巴巴集团控股有限公司 | Method and system for evaluating target detection model |
| CN111723614A (en) * | 2019-03-20 | 2020-09-29 | 北京四维图新科技股份有限公司 | Traffic signal lamp identification method and device |
| CN111723788A (en) * | 2019-03-19 | 2020-09-29 | 杭州海康威视数字技术股份有限公司 | Character recognition method and device |
| CN111814584A (en) * | 2020-06-18 | 2020-10-23 | 北京交通大学 | Vehicle weight identification method under multi-view-angle environment based on multi-center measurement loss |
| CN111914909A (en) * | 2020-07-15 | 2020-11-10 | 南京审计大学 | Hyperspectral change detection method based on space-spectrum combined three-direction convolution network |
| CN112052787A (en) * | 2020-09-03 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Target detection method and device based on artificial intelligence and electronic equipment |
| CN112307978A (en) * | 2020-10-30 | 2021-02-02 | 腾讯科技（深圳）有限公司 | Target detection method and device, electronic equipment and readable storage medium |
| CN112347895A (en) * | 2020-11-02 | 2021-02-09 | 北京观微科技有限公司 | Ship remote sensing target detection method based on boundary optimization neural network |
| CN112801070A (en) * | 2021-04-14 | 2021-05-14 | 浙江啄云智能科技有限公司 | Target detection method, device, equipment and storage medium |
| CN113011398A (en) * | 2021-04-28 | 2021-06-22 | 北京邮电大学 | Target change detection method and device for multi-temporal remote sensing image |
| CN113516053A (en) * | 2021-05-28 | 2021-10-19 | 西安空间无线电技术研究所 | Ship target refined detection method with rotation invariance |
| CN114582140A (en) * | 2022-01-17 | 2022-06-03 | 浙江银江智慧交通工程技术研究院有限公司 | Method, system, device and medium for identifying traffic flow of urban road intersection |
| CN115965582A (en) * | 2022-11-22 | 2023-04-14 | 哈尔滨岛田大鹏工业股份有限公司 | Ultrahigh-resolution-based engine cylinder body and cylinder cover surface defect detection method |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563303A (en) * | 2017-08-09 | 2018-01-09 | 中国科学院大学 | A kind of robustness Ship Target Detection method based on deep learning |
-
2018
- 2018-01-26 CN CN201810075285.3A patent/CN108319949A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563303A (en) * | 2017-08-09 | 2018-01-09 | 中国科学院大学 | A kind of robustness Ship Target Detection method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| LIU LEI,ET AL: "《Learning a rotation invariant detector with rotatable bounding box》", 《ARXIV PREPRINT ARXIV:1711.09405 (2017)》 * |
Cited By (61)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109166106B (en) * | 2018-08-02 | 2021-07-30 | 山东大学 | Target detection position correction method and device based on sliding window |
| CN109166106A (en) * | 2018-08-02 | 2019-01-08 | 山东大学 | A kind of target detection aligning method and apparatus based on sliding window |
| CN109165603B (en) * | 2018-08-28 | 2022-03-01 | 中国科学院遥感与数字地球研究所 | Ship detection method and device |
| CN109165603A (en) * | 2018-08-28 | 2019-01-08 | 中国科学院遥感与数字地球研究所 | A kind of Ship Detection and device |
| CN109255317A (en) * | 2018-08-31 | 2019-01-22 | 西北工业大学 | A kind of Aerial Images difference detecting method based on dual network |
| CN109255317B (en) * | 2018-08-31 | 2021-06-11 | 西北工业大学 | Aerial image difference detection method based on double networks |
| CN109344717A (en) * | 2018-09-01 | 2019-02-15 | 哈尔滨工程大学 | A kind of deep-sea target on-line checking recognition methods of multi-threshold dynamic statistics |
| CN109344717B (en) * | 2018-09-01 | 2021-10-19 | 哈尔滨工程大学 | Multi-threshold dynamic statistical deep sea target online detection and identification method |
| CN109523015A (en) * | 2018-11-09 | 2019-03-26 | 上海海事大学 | Image processing method in a kind of neural network |
| CN109711427A (en) * | 2018-11-19 | 2019-05-03 | 深圳市华尊科技股份有限公司 | Object detection method and Related product |
| CN109800755A (en) * | 2018-12-14 | 2019-05-24 | 中国科学院深圳先进技术研究院 | A kind of remote sensing image small target detecting method based on Analysis On Multi-scale Features |
| CN109800637A (en) * | 2018-12-14 | 2019-05-24 | 中国科学院深圳先进技术研究院 | A kind of remote sensing image small target detecting method |
| CN111382762A (en) * | 2018-12-28 | 2020-07-07 | 同方威视技术股份有限公司 | Empty box identification method and system |
| CN109948415A (en) * | 2018-12-30 | 2019-06-28 | 中国科学院软件研究所 | Remote sensing image object detection method based on filtering background and scale prediction |
| CN109919183B (en) * | 2019-01-24 | 2020-12-18 | 北京大学 | Image identification method, device and equipment based on small samples and storage medium |
| CN109919183A (en) * | 2019-01-24 | 2019-06-21 | 北京大学 | A kind of image-recognizing method based on small sample, device, equipment and storage medium |
| CN109800735A (en) * | 2019-01-31 | 2019-05-24 | 中国人民解放军国防科技大学 | Accurate detection and segmentation method for ship target |
| CN111723788A (en) * | 2019-03-19 | 2020-09-29 | 杭州海康威视数字技术股份有限公司 | Character recognition method and device |
| CN111723788B (en) * | 2019-03-19 | 2023-12-29 | 杭州海康威视数字技术股份有限公司 | Character recognition method and device |
| CN111723614A (en) * | 2019-03-20 | 2020-09-29 | 北京四维图新科技股份有限公司 | Traffic signal lamp identification method and device |
| CN110008863A (en) * | 2019-03-22 | 2019-07-12 | 哈尔滨工业大学 | Efficient multi-scale sampling method based on high-resolution bridge area ship image |
| CN110335238A (en) * | 2019-05-08 | 2019-10-15 | 吉林大学 | A kind of Automotive Paint Film Properties defect recognition system constituting method of deep learning |
| CN110223302B (en) * | 2019-05-08 | 2021-11-19 | 华中科技大学 | Ship multi-target detection method based on rotation region extraction |
| CN110223302A (en) * | 2019-05-08 | 2019-09-10 | 华中科技大学 | A kind of naval vessel multi-target detection method extracted based on rotary area |
| CN110147837A (en) * | 2019-05-14 | 2019-08-20 | 中国电子科技集团公司第二十八研究所 | The intensive object detection method of any direction, system and the equipment focused based on feature |
| CN111723809B (en) * | 2019-06-10 | 2021-09-07 | 创新先进技术有限公司 | Method and system for evaluating target detection model |
| US10956769B2 (en) | 2019-06-10 | 2021-03-23 | Advanced New Technologies Co., Ltd. | Method and system for evaluating an object detection model |
| CN111723809A (en) * | 2019-06-10 | 2020-09-29 | 阿里巴巴集团控股有限公司 | Method and system for evaluating target detection model |
| US10891510B2 (en) | 2019-06-10 | 2021-01-12 | Advanced New Technologies Co., Ltd. | Method and system for evaluating an object detection model |
| CN110443259A (en) * | 2019-07-29 | 2019-11-12 | 中科光启空间信息技术有限公司 | A method of extracting sugarcane from intermediate resolution remote sensing image |
| CN110443259B (en) * | 2019-07-29 | 2023-04-07 | 中科光启空间信息技术有限公司 | Method for extracting sugarcane from medium-resolution remote sensing image |
| CN110728307A (en) * | 2019-09-20 | 2020-01-24 | 天津大学 | Method for realizing small sample character recognition of X-ray image by self-generating data set and label |
| CN110689021A (en) * | 2019-10-17 | 2020-01-14 | 哈尔滨理工大学 | Real-time target detection method in low-visibility environment based on deep learning |
| CN110796185A (en) * | 2019-10-17 | 2020-02-14 | 北京爱数智慧科技有限公司 | Method and device for detecting image annotation result |
| CN111368625B (en) * | 2019-11-08 | 2024-02-27 | 深圳北航新兴产业技术研究院 | Pedestrian target detection method based on cascade optimization |
| CN111368625A (en) * | 2019-11-08 | 2020-07-03 | 深圳北航新兴产业技术研究院 | Pedestrian target detection method based on cascade optimization |
| CN110853295A (en) * | 2019-11-12 | 2020-02-28 | 江西赣鄱云新型智慧城市技术研究有限公司 | High-altitude parabolic early warning method and device |
| CN110942455A (en) * | 2019-11-26 | 2020-03-31 | 南方电网数字电网研究院有限公司 | Method and device for detecting missing of cotter pin of power transmission line and computer equipment |
| CN111191566A (en) * | 2019-12-26 | 2020-05-22 | 西北工业大学 | Optical remote sensing image multi-target detection method based on pixel classification |
| CN111191566B (en) * | 2019-12-26 | 2022-05-17 | 西北工业大学 | Optical remote sensing image multi-target detection method based on pixel classification |
| CN111079688A (en) * | 2019-12-27 | 2020-04-28 | 中国电子科技集团公司第十五研究所 | Living body detection method based on infrared image in face recognition |
| CN111354378A (en) * | 2020-02-12 | 2020-06-30 | 北京声智科技有限公司 | Voice endpoint detection method, device, equipment and computer storage medium |
| CN111524095A (en) * | 2020-03-24 | 2020-08-11 | 西安交通大学 | Target detection method for rotating object |
| CN111401302B (en) * | 2020-04-07 | 2022-08-02 | 中国人民解放军海军航空大学 | Remote sensing image ship target integrated detection and fine-grained identification method |
| CN111401302A (en) * | 2020-04-07 | 2020-07-10 | 中国人民解放军海军航空大学 | Remote sensing image ship target integrated detection and fine-grained identification method |
| CN111814584B (en) * | 2020-06-18 | 2024-03-12 | 北京交通大学 | Vehicle re-identification method based on multi-center measurement loss under multi-view environment |
| CN111814584A (en) * | 2020-06-18 | 2020-10-23 | 北京交通大学 | Vehicle weight identification method under multi-view-angle environment based on multi-center measurement loss |
| CN111914909A (en) * | 2020-07-15 | 2020-11-10 | 南京审计大学 | Hyperspectral change detection method based on space-spectrum combined three-direction convolution network |
| CN111914909B (en) * | 2020-07-15 | 2021-01-12 | 南京审计大学 | Hyperspectral change detection method based on space-spectrum combined three-direction convolution network |
| CN112052787A (en) * | 2020-09-03 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Target detection method and device based on artificial intelligence and electronic equipment |
| CN112052787B (en) * | 2020-09-03 | 2021-07-30 | 腾讯科技（深圳）有限公司 | Target detection method and device based on artificial intelligence and electronic equipment |
| CN112307978A (en) * | 2020-10-30 | 2021-02-02 | 腾讯科技（深圳）有限公司 | Target detection method and device, electronic equipment and readable storage medium |
| CN112347895A (en) * | 2020-11-02 | 2021-02-09 | 北京观微科技有限公司 | Ship remote sensing target detection method based on boundary optimization neural network |
| CN112801070B (en) * | 2021-04-14 | 2021-09-21 | 浙江啄云智能科技有限公司 | Target detection method, device, equipment and storage medium |
| CN112801070A (en) * | 2021-04-14 | 2021-05-14 | 浙江啄云智能科技有限公司 | Target detection method, device, equipment and storage medium |
| CN113011398A (en) * | 2021-04-28 | 2021-06-22 | 北京邮电大学 | Target change detection method and device for multi-temporal remote sensing image |
| CN113516053A (en) * | 2021-05-28 | 2021-10-19 | 西安空间无线电技术研究所 | Ship target refined detection method with rotation invariance |
| CN113516053B (en) * | 2021-05-28 | 2024-05-14 | 西安空间无线电技术研究所 | Ship target refinement detection method with rotation invariance |
| CN114582140A (en) * | 2022-01-17 | 2022-06-03 | 浙江银江智慧交通工程技术研究院有限公司 | Method, system, device and medium for identifying traffic flow of urban road intersection |
| CN115965582A (en) * | 2022-11-22 | 2023-04-14 | 哈尔滨岛田大鹏工业股份有限公司 | Ultrahigh-resolution-based engine cylinder body and cylinder cover surface defect detection method |
| CN115965582B (en) * | 2022-11-22 | 2024-03-08 | 哈尔滨岛田大鹏工业股份有限公司 | Ultrahigh-resolution-based method for detecting surface defects of cylinder body and cylinder cover of engine |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108319949A (en) | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image | |
| CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN108898047B (en) | Pedestrian detection method and system based on blocking and shielding perception | |
| CN109711288A (en) | Remote sensing ship detecting method based on feature pyramid and distance restraint FCN | |
| CN109740665A (en) | Shielded image ship object detection method and system based on expertise constraint | |
| CN102654902B (en) | Contour vector feature-based embedded real-time image matching method | |
| CN109635875A (en) | A kind of end-to-end network interface detection method based on deep learning | |
| CN109886939B (en) | Bridge crack detection method based on tensor voting | |
| CN104574393B (en) | A kind of three-dimensional pavement crack pattern picture generates system and method | |
| CN107248159A (en) | A kind of metal works defect inspection method based on binocular vision | |
| CN108960198A (en) | A kind of road traffic sign detection and recognition methods based on residual error SSD model | |
| CN107886120A (en) | Method and apparatus for target detection tracking | |
| CN105957054B (en) | A kind of image change detection method | |
| CN105740910A (en) | Vehicle object detection method and device | |
| CN108830331A (en) | A kind of Ground Penetrating Radar object detection method based on full convolutional network | |
| CN107451982A (en) | A kind of high canopy density standing forest tree crown area acquisition methods based on unmanned plane image | |
| CN108346144A (en) | Bridge Crack based on computer vision monitoring and recognition methods automatically | |
| CN108564085A (en) | A kind of method of automatic reading pointer type instrument reading | |
| CN115345881B (en) | Pavement disease detection method based on computer vision | |
| CN107392929A (en) | A kind of intelligent target detection and dimension measurement method based on human vision model | |
| CN110189375A (en) | A kind of images steganalysis method based on monocular vision measurement | |
| CN112348758B (en) | Optical remote sensing image data enhancement method and target identification method | |
| CN108648184A (en) | A kind of detection method of remote sensing images high-altitude cirrus | |
| CN108804992A (en) | A kind of Demographics' method based on deep learning | |
| CN109584206B (en) | Method for synthesizing training sample of neural network in part surface flaw detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180724 |