CN109409285A - Remote sensing video object detection method based on overlapping slice - Google Patents
Remote sensing video object detection method based on overlapping slice Download PDFInfo
- Publication number
- CN109409285A CN109409285A CN201811246304.0A CN201811246304A CN109409285A CN 109409285 A CN109409285 A CN 109409285A CN 201811246304 A CN201811246304 A CN 201811246304A CN 109409285 A CN109409285 A CN 109409285A
- Authority
- CN
- China
- Prior art keywords
- test sample
- remote sensing
- target
- overlapping
- slice
- 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
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
- G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
-
- 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
Abstract
The invention discloses a kind of remote sensing video object detection method based on overlapping slice, mainly solve the problems, such as that target omission factor, target false detection rate are higher in the prior art and target detection process very complicated, time-consuming are long, low efficiency.The specific steps of the present invention are as follows: (1) choosing source data；(2) training sample set and test sample collection are generated；(3) training sample is pre-processed；(4) overlapping slice pretreatment is carried out to test sample；(5) training objective detection model；(6) target in detection overlapping slice；(7) target in test sample is detected；(8) test sample picture is switched into video.The present invention can reduce target omission factor and target false detection rate, improve the accuracy in detection to remote sensing video object, shortens the time-consuming of target detection process, realizes the efficient detection to aircraft, Ship Target in remote sensing video.
Description
Technical field
The invention belongs to technical field of image processing, further relate to one of image object detection technique field base
In the remote sensing video object detection method of overlapping slice.In the remote sensing video image that the present invention can be used for shooting video satellite
The targets such as static and movement aircraft, naval vessel are detected.
Background technique
Video satellite is a kind of satellite of video imaging, its appearance is that real-time continuous be ground-to-ground observed provides skill
Art is supported.The increase resolution of video satellite is to second grade, and compared with conventional satellite, video satellite can provide the dynamic letter of target
Breath, has expanded the data source of remote sensing information, so that the analysis of remote sensing information and interpretation are promoted to dynamic image from still image, is
Remote sensing fields multidate information, which understands, provides new opportunity with processing.Remote sensing video satellite by the dynamic event detected into
Row analysis provides important evidence for disaster monitoring, weather prognosis, war monitoring, business assessment etc..Aircraft, warship in remote sensing video
No matter the detection of ship target in production and living, Transportation Planning or monitoring and the military operations such as information acquisition has important meaning
Justice, it is therefore proposed that a kind of effective remote sensing video object detection method has important society and military value.
Patent document " remote sensing video image motion target real-time intelligence of the Hunan Hang Sheng satellite Science and Technology Ltd. in its application
Can cognitive method and its device " one kind is disclosed in (number of patent application: 201810111223.3, publication number: 108389220A)
Remote sensing video image motion object detection method.This method uses any of traditional images processing method processing remote sensing video first
Frame image obtains being possible to the candidate region image T comprising moving target in the frame image；Then pass through depth convolutional Neural net
Network realization classifies to the candidate region image T, finally realizes small scaled target detection and utilizes a small amount of sample data
Training depth convolutional neural networks, improve the accuracy of moving object detection.But the shortcoming that this method still has
It is to obtain remote sensing video image target candidate region due to directlying adopt traditional images processing method, have ignored image border mesh
Target particularity does not solve the problems, such as to detect low resolution Small object in the biggish remote sensing video image of frame width, leads to mesh
It marks omission factor and target false detection rate is higher.
Paper " remote sensing satellite video image vehicle multidate information Study on Extraction Method " (Chinese section that Yuan Yiqin is delivered at it
University, institute master thesis 2017) in propose the method for remote sensing satellite video frequency vehicle target detection.This method is first
By four image preprocessing, building area mask, threshold method segmentation object and morphology processing steps, to single-frame images into
Row target detection；Then the algorithm of target detection blended using a kind of background subtraction and frame differential method, to multiframe satellite
The moving vehicle target of video is detected.Although method proposed in this paper can inhibit mobile background edge and residual noise
Interference, improves the correctness and quality of detection, and still, the shortcoming that this method still has is, due to using a variety of methods
It blends, this method step is various, and time-consuming for calculating, reduces the efficiency of target detection.
Summary of the invention
It is an object of the invention in view of the above shortcomings of the prior art, propose a kind of remote sensing view based on overlapping slice
Frequency object detection method.
Realizing thinking of the invention is, concentrates every remote sensing images to carry out overlapping slicing treatment test sample, will locate in advance
Each test sample after reason is input in trained YOLO v3 network, obtains the object detection results of each overlapping slice,
The target detection knot of test sample is obtained using bounding box position integration method in overlapping slice and non-maxima suppression NMS algorithm
All test sample pictures with object detection results are switched to video by fruit.
The specific steps that the present invention realizes include the following:
(1) source data is chosen:
(1a) extracts at least 110 frames from the video that one section of remote sensing satellite obtains and contains the distant of Aircraft Targets and other scenery
Feel image；
(1b) extracts at least 110 frames in the video that another section of remote sensing satellite obtains and contains the distant of Ship Target and other scenery
Feel image；
(1c) is extracted from the remote sensing video data source except step (1a), step (1b) contains Aircraft Targets, naval vessel mesh
Each at least 90 remote sensing images of mark；
(2) training sample set and test sample collection are generated:
(2a) intercepts every width and contains Aircraft Targets from containing first three of Aircraft Targets and other scenery frame remote sensing images
All remote sensing images therefrom randomly select at least 110；By at least 90 remote sensing images containing Aircraft Targets and selected
Remote sensing images form training sample set 1；
(2b) intercepts every width and contains Ship Target from containing first three of Ship Target and other scenery frame remote sensing images
Remote sensing images therefrom randomly select at least 110；By at least 90 remote sensing images and selected remote sensing containing Ship Target
Image forms training sample set 2；
(2c) from the remote sensing images that every width contains Aircraft Targets and other scenery, interception size is 3200 × 1800 pixels
Remote sensing images, at least 110 remote sensing images after screenshot are formed into test sample collections 1；
(2d) from the remote sensing images that every width contains Ship Target and other scenery, interception size is 3200 × 1800 pixels
Remote sensing images, at least 110 remote sensing images after screenshot are formed into test sample collections 2；
(3) training sample is pre-processed:
All target categories in each training sample of two training samples concentration, target position information are marked
Note, obtains the mark file of 200 xml formats；The target sizes and target position information that mark in file are normalized
Target category and normalization information are switched to the mark file of text formatting by processing；
(4) overlapping slice pretreatment is carried out to test sample:
Every remote sensing images are concentrated to carry out overlapping slice pretreatment two test samples, obtained each slice is overlapping
Slice, the top left corner apex for saving each overlapping slice are in the coordinate (x in corresponding test samplea,ya) and lower right corner apex
Coordinate (x in corresponding test sampleb,yb), obtained all overlappings slice forms pretreated test sample collection；
(5) training objective detection model:
Pretreated all training samples are input in YOLO v3 model and are iterated training, until output the
Weights after 1500 iteration obtain trained YOLO v3 network；
(6) target in detection overlapping slice:
Pretreated each test sample is input in trained YOLO v3 network, the output of YOLO v3 network is every
The object detection results of a overlapping slice；
(7) target in test sample is detected:
(7a) utilizes bounding box position integration method in overlapping slice, by side in the object detection results of each overlapping slice
Boundary's frame top left corner apex is in the coordinate of coordinate, bounding box lower right corner apex in plane coordinate system in plane coordinate system point
The position coordinates in corresponding test sample are not converted to；
(7b) utilizes non-maxima suppression NMS algorithm, removes the redundancy bounding box of target in test sample, retains each mesh
Target maximum confidence bounding box, obtains the object detection results of test sample；
(8) test sample picture is switched into video:
By all test sample pictures with object detection results, switch to remote sensing video according to picture transmission frame number per second
File.
The present invention has the advantage that compared with prior art
First, since the present invention carries out overlapping slice pretreatment to test sample, so that the target at slicing edge is in phase
It can more completely show in neighbour's slice, overcome the particularity for ignoring image border target in the prior art, lead to target missing inspection
The higher problem of rate can detect distant more fully hereinafter so that the present invention improves the detected probability of target in overlapping region
Feel all aircrafts, the Ship Target in video, reduces the omission factor of target detection.
Second, since the present invention carries out overlapping slice pretreatment to every remote sensing images, obtained each slice is overlapping
Slice, converts the biggish Remote Sensing Target test problems of frame width to the target detection problems of multiple small images, simplifies low
The background of resolution ratio Small object overcomes in the prior art without solving to detect low point in the biggish remote sensing video image of frame width
The small target of resolution leads to the higher problem of target false detection rate, so that the present invention is to low in the biggish remote sensing video image of frame width
Resolution ratio Small object preferably detects, and improves the accuracy rate of remote sensing video object detection, reduces target false detection rate.
Third makees target detection since the present invention has used YOLO v3 network in the target that detection is overlapped in slice
For regression problem, end-to-end training and detection are realized, it is cumbersome to overcome detecting step complexity in the prior art, calculates time-consuming
It is long, the problem of target detection low efficiency so that the present invention improves the detection performance of remote sensing video object, can faster, it is more quasi-
True carry out target detection improves the efficiency of remote sensing video object detection.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is the testing result figure of 1 aircraft of emulation experiment of the present invention；
Fig. 3 is the testing result figure on 2 naval vessel of emulation experiment of the present invention.
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawing.
Referring to attached drawing 1, step of the invention is described in further detail.
Step 1. chooses source data.
(1.1) at least 110 frames are extracted from the video that one section of remote sensing satellite obtains contain Aircraft Targets and other scenery
Remote sensing images；
(1.2) at least 110 frames in the video that another section of remote sensing satellite obtains are extracted and contain Ship Target and other scenery
Remote sensing images；
(1.3) it is extracted from the remote sensing video data source except step (1.1), step (1.2) and contains Aircraft Targets, naval vessel
Each at least 90 remote sensing images of target.
Step 2. generates training sample set and test sample collection.
From containing first three of Aircraft Targets and other scenery frame remote sensing images, intercepts every width and contain all of Aircraft Targets
Remote sensing images therefrom randomly select at least 110；By at least 90 remote sensing images and selected remote sensing containing Aircraft Targets
Image forms training sample set 1.
From containing first three of Ship Target and other scenery frame remote sensing images, the remote sensing that every width contains Ship Target is intercepted
Image therefrom randomly selects at least 110；By at least 90 remote sensing images and selected remote sensing figure containing Ship Target
Picture forms training sample set 2.
From the remote sensing images that every width contains Aircraft Targets and other scenery, interception size is the distant of 3200 × 1800 pixels
Feel image, at least 110 remote sensing images after screenshot are formed into test sample collection 1.
From the remote sensing images that every width contains Ship Target and other scenery, interception size is the distant of 3200 × 1800 pixels
Feel image, at least 110 remote sensing images after screenshot are formed into test sample collection 2.
Step 3. pre-processes training sample.
All target categories in each training sample of two training samples concentration, target position information are marked
Note, obtains the mark file of 200 xml formats；The target sizes and target position information that mark in file are normalized
Target category and normalization information are switched to the mark file of text formatting by processing.
Steps are as follows for the target position information normalized:
The first step calculates the normalized value of target's center's point abscissa using following formula:
Wherein, x indicates each target's center's point abscissa, x0Indicate the minimum abscissa of each target position, x1Indicate every
The maximum abscissa of a target position, * indicate multiplication operations, and w' indicates the width of the sample of each training.
Second step calculates the normalized value of target's center's point ordinate using following formula:
Wherein, y indicates each target's center's point ordinate, y0Indicate the minimum ordinate of each target position, y1Indicate every
The maximum ordinate of a target position, * indicate multiplication operations, and h' indicates the height of the sample of each training.
Third step calculates the normalized value of target width using following formula:
Wherein, w indicates the normalization width of each target, x1Indicate the maximum abscissa of each target position, x0Indicate every
The minimum abscissa of a target position, w' indicate the width of the sample of each training.
4th step calculates the normalized value of object height using following formula:
Wherein, h indicates the normalization width of each target, y1Indicate the maximum ordinate of each target position, y0Indicate every
The minimum ordinate of a target position, h' indicate the width of the sample of each training.
Step 4. carries out overlapping slice pretreatment to test sample.
Every remote sensing images are concentrated to carry out overlapping slice pretreatment two test samples, obtained each slice is overlapping
Slice, the top left corner apex for saving each overlapping slice are in the coordinate (x in corresponding test samplea,ya) and lower right corner apex
Coordinate (x in corresponding test sampleb,yb), obtained all overlappings slice forms pretreated test sample collection.
The overlapping slice pretreatment refers to: being the sliding window of 608 × 608 pixels with a size, from test sample collection
In every remote sensing images top left corner apex rise, along X direction slide, 500 pixels are divided between sliding, obtain X direction
Overlapping slice；From test sample concentrates the top left corner apex of every remote sensing images, along y direction, using 500 pixels as interval,
As every row sliding window starting point, sliding window is slided in X direction, 500 pixels are divided between sliding, the overlapping for obtaining X direction is cut
Piece；Every row intercepts 6 overlapping slices, and each column intercepts 3 overlapping slices；The overlapping region size of each X direction contiguous slices
For 108 × 608 pixels, the overlapping region size of each y direction contiguous slices is 608 × 108 pixels；It each of obtains big
Small is that the slice of 608 × 608 pixels is known as being overlapped slice.
Step 5. training objective detection model.
Pretreated all training samples are input in YOLO v3 model and are iterated training, until output the
Weights after 1500 iteration obtain trained YOLO v3 network.
The YOLO v3 model uses the darknet53.conv.74 convolution weight of the pre-training on Imagenet.
Target in step 6. detection overlapping slice.
Pretreated each test sample is input in trained YOLO v3 network, the output of YOLO v3 network is every
The object detection results of a overlapping slice.
The object detection results of each overlapping slice include following three parts, are each heavy for marking respectively
All coordinate (x that may be in containing the top left corner apex of the bounding box of mesh target area in plane coordinate system in folded slicec,
yc), the coordinate (x of the lower right corner apex of each bounding box in plane coordinate systemd,yd) and each bounding box confidence level.
Step 7. detects the target in test sample.
Using bounding box position integration method in overlapping slice, by bounding box in the object detection results of each overlapping slice
Coordinate of the coordinate, bounding box lower right corner apex that top left corner apex is in plane coordinate system in plane coordinate system turns respectively
The position coordinates being changed in corresponding test sample.
Bounding box position integration method is as follows in the overlapping slice:
Step 1, the top left corner apex for reading each object boundary frame in each overlapping slice are in plane coordinate system
Coordinate (xc,yc), coordinate (x of the lower right corner apex in plane coordinate systemd,yd)；Each overlapping slice is read to test corresponding
Top left corner apex in sample is in the coordinate (x in plane coordinate systema,ya), seat of the lower right corner apex in plane coordinate system
Mark (xb,yb)。
The top left corner apex of each object boundary frame in each overlapping slice is in plane using following formula by step 2
Coordinate (x in coordinate systemc,yc) be converted in corresponding test sample at bounding box top left corner apex in plane coordinate system
Coordinate (xe,ye):
xe=xc*(xb-xa)+xa
ye=yc*(yb-ya)+ya
Wherein, xeIndicate that the top left corner apex of object boundary frame in each test sample is in the horizontal seat in plane coordinate system
Mark, xcIndicate that the top left corner apex of each object boundary frame in each overlapping slice is in the abscissa in plane coordinate system, * table
Show multiplication operations, xbIndicate an abscissa of the overlapping slice lower right corner apex in corresponding test sample, xaIndicate each overlapping
Slice top left corner apex is in the abscissa in corresponding test sample；yeIndicate the upper left of object boundary frame in each test sample
Angular vertex is in the ordinate in plane coordinate system, ycIndicate the top left corner apex of each object boundary frame in each overlapping slice
The ordinate being in plane coordinate system, ybIndicate a vertical seat of the overlapping slice lower right corner apex in corresponding test sample
Mark, yaIndicate that each overlapping slice top left corner apex is in the ordinate in corresponding test sample.
Third step, using following formula by the lower right corner apex of each object boundary frame in each overlapping slice in plane
Coordinate (x in coordinate systemd,yd) be converted to coordinate of the lower right corner apex in plane coordinate system in corresponding test sample
(xf,yf):
xf=xd*(xb-xa)+xa
yf=yd*(yb-ya)+ya
Wherein, xfIndicate abscissa of the lower right corner apex of each object boundary frame in corresponding test sample, xdIt indicates
Abscissa of the lower right corner apex of each object boundary frame in plane coordinate system, * indicate the behaviour that is multiplied in each overlapping slice
Make, xbIndicate an abscissa of the overlapping slice lower right corner apex in corresponding test sample, xaIndicate each overlapping slice upper left
Angular vertex is in the abscissa in corresponding test sample；yfIndicate that the lower right corner apex of each object boundary frame is tested corresponding
Ordinate in sample, ydIndicate the lower right corner apex of each object boundary frame in each overlapping slice in plane coordinate system
Ordinate, * indicate multiplication operations, ybIndicate an abscissa of the overlapping slice lower right corner apex in corresponding test sample, ya
Indicate that each overlapping slice top left corner apex is in the abscissa in corresponding test sample.
Using non-maxima suppression NMS algorithm, the redundancy bounding box of target in test sample is removed, each target is retained
Maximum confidence bounding box obtains the object detection results of test sample.
Test sample picture is switched to video by step 8..
By all test sample pictures with object detection results, switch to remote sensing video according to picture transmission frame number per second
File.
Effect of the invention is described further below with reference to emulation experiment:
1. simulated conditions:
Emulation experiment of the invention is in Intel (R) Core (TM) i5-7500CPU of dominant frequency 3.4GHz, core frequency
It is carried out under the software environment of the GTX1060-6GD5 of 1569-1784MHz, the interior hardware environment for saving as 8GB and Opencv.
2. emulation content and interpretation of result:
Using method of the invention, under above-mentioned simulated conditions, step according to the invention has carried out emulation experiment twice,
Emulation experiment 1 is tested the Aircraft Targets in remote sensing video, emulation experiment 2 be to the Ship Target in remote sensing video into
Row test.
Emulation experiment 1 of the invention is to extract 297 from Santiago remote sensing video image that one section of remote sensing satellite obtains
Frame contains Aircraft Targets and the remote sensing images of other scenery, and 90 containing Aircraft Targets are extracted from other remote sensing video data sources
Open remote sensing images.All images containing Aircraft Targets, image size system are intercepted from first three frame of Santiago remote sensing images
One is 672 × 672 pixels, and therefrom randomly selects 90 that 110 images and other remote sensing video datas obtain and contain aircraft
The picture of target collectively constitutes the training sample set of Aircraft Targets detection.To 297 frame Santiago remote sensing video images of acquisition
It is intercepted from from high 1800 pixel, wide 3200 pixel of every frame, interception image size is 3200 × 1800 pixels, is obtained
297 Zhang Fei's machine target detection test samples.To training sample be labeled with normalized pretreatment, to test sample carry out weight
Folded slice pretreatment.YOLO v3 model is trained with pretreated training sample, with trained model to pretreated survey
Sample this progress target detection.Using bounding box position integration method and non-maxima suppression NMS algorithm in overlapping slice, to survey
Sample this progress Aircraft Targets detection, finally switchs to video for the test sample image containing Aircraft Targets testing result.Contain
Video the second frame image of Aircraft Targets testing result is as shown in Figure 2.
Emulation experiment 2 of the invention is extracted 297 frames from Bogotá remote sensing video image that one section of remote sensing satellite obtains and is contained
There are Ship Target and the remote sensing images of other scenery, is extracted from other remote sensing video data sources 90 distant containing Ship Target
Feel image.All images containing Ship Target are intercepted from first three frame of Bogotá remote sensing images, image size is unified for 672
× 672 pixels, and 90 that 110 images and other remote sensing video datas obtain therefrom are randomly selected containing Ship Target
Picture collectively constitutes the training sample set of Ship Target Detection.To 297 frame Bogotá remote sensing video images of acquisition from every frame
It is intercepted at high 1800 pixel, wide 3200 pixel, interception image size is 3200 × 1800 pixels, obtains 297 warships
Ship detection test sample.To training sample be labeled with normalized pretreatment, overlapping slice is carried out to test sample
Pretreatment.YOLO v3 model is trained with pretreated training sample, with trained model to pretreated test sample
Carry out target detection.Using bounding box position integration method and non-maxima suppression NMS algorithm in overlapping slice, to test sample
Ship Target Detection is carried out, the test sample image containing Ship Target Detection result is finally switched into video.Contain naval vessel mesh
Video the second frame image for marking testing result is as shown in Figure 3.
White rectangle frame in Fig. 2 and Fig. 3 is the rectangle frame for aircraft and Ship Target in labeled test sample.
The target that white rectangle collimation mark is remembered from the testing result figure that emulation experiment twice obtains can be seen that of the invention
Method can accurately detect aircraft, the Ship Target of the motion and standstill in remote sensing video.
In conclusion the present invention is utilized by concentrating every remote sensing images to carry out overlapping slicing treatment test sample
YOLO v3 network carries out target detection, it is contemplated that the particularity of image border target improves the tested of target in overlapping region
Probability is surveyed, the biggish Remote Sensing Target test problems of frame width is converted to the target detection problems of multiple small images, simplifies
The background of low resolution Small object realizes end-to-end training and detection, reduces target omission factor and false detection rate, improves pair
The accuracy in detection of remote sensing video object shortens the time-consuming of target detection process, realizes to aircraft, naval vessel in remote sensing video
The efficient detection of target.
Claims (6)
1. a kind of remote sensing video object detection method based on overlapping slice, which is characterized in that concentrated to test sample every distant
Sense image carries out overlapping slicing treatment, and pretreated each test sample is input in trained YOLO v3 network, is obtained
To the object detection results of each overlapping slice, bounding box position integration method and non-maxima suppression in overlapping slice are utilized
NMS algorithm obtains the object detection results of test sample, and all test sample pictures with object detection results are switched to regard
Frequently；The specific steps of this method include the following:
(1) source data is chosen:
(1a) extracts at least 110 frames from the video that one section of remote sensing satellite obtains and contains Aircraft Targets and the remote sensing figure of other scenery
Picture；
(1b) extracts at least 110 frames in the video that another section of remote sensing satellite obtains and contains Ship Target and the remote sensing figure of other scenery
Picture；
(1c) extracts each containing Aircraft Targets, Ship Target from the remote sensing video data source except step (1a), step (1b)
At least 90 remote sensing images；
(2) training sample set and test sample collection are generated:
(2a) intercepts every width and contains all of Aircraft Targets from containing first three of Aircraft Targets and other scenery frame remote sensing images
Remote sensing images therefrom randomly select at least 110；By at least 90 remote sensing images and selected remote sensing containing Aircraft Targets
Image forms training sample set 1；
(2b) intercepts the remote sensing that every width contains Ship Target from containing first three of Ship Target and other scenery frame remote sensing images
Image therefrom randomly selects at least 110；By at least 90 remote sensing images and selected remote sensing figure containing Ship Target
Picture forms training sample set 2；
(2c) from the remote sensing images that every width contains Aircraft Targets and other scenery, interception size is the distant of 3200 × 1800 pixels
Feel image, at least 110 remote sensing images after screenshot are formed into test sample collection 1；
(2d) from the remote sensing images that every width contains Ship Target and other scenery, interception size is the distant of 3200 × 1800 pixels
Feel image, at least 110 remote sensing images after screenshot are formed into test sample collection 2；
(3) training sample is pre-processed:
All target categories in each training sample of two training samples concentration, target position information are labeled, obtained
To the mark file of 200 xml formats；The target sizes and target position information that mark in file are normalized,
Target category and normalization information are switched to the mark file of text formatting；
(4) overlapping slice pretreatment is carried out to test sample:
Every remote sensing images are concentrated to carry out overlapping slice pretreatment two test samples, obtained each slice is that overlapping is cut
Piece, the top left corner apex for saving each overlapping slice are in the coordinate (x in corresponding test samplea,ya) and lower right corner apex exist
Coordinate (x in corresponding test sampleb,yb), obtained all overlappings slice forms pretreated test sample collection；
(5) training objective detection model:
Pretreated all training samples are input in YOLO v3 model and are iterated training, until output the 1500th time
Weights after iteration obtain trained YOLO v3 network；
(6) target in detection overlapping slice:
Pretreated each test sample is input in trained YOLO v3 network, the output of YOLO v3 network is each heavy
The object detection results of folded slice；
(7) target in test sample is detected:
(7a) utilizes bounding box position integration method in overlapping slice, by bounding box in the object detection results of each overlapping slice
Coordinate of the coordinate, bounding box lower right corner apex that top left corner apex is in plane coordinate system in plane coordinate system turns respectively
The position coordinates being changed in corresponding test sample；
(7b) utilizes non-maxima suppression NMS algorithm, removes the redundancy bounding box of target in test sample, retains each target
Maximum confidence bounding box obtains the object detection results of test sample；
(8) test sample picture is switched into video:
By all test sample pictures with object detection results, switch to remote sensing video text according to picture transmission frame number per second
Part.
2. the remote sensing video object detection method according to claim 1 based on overlapping slice, which is characterized in that step
(3) steps are as follows for the target position information normalized described in:
The first step calculates the normalized value of target's center's point abscissa using following formula:
Wherein, x indicates each target's center's point abscissa, x0Indicate the minimum abscissa of each target position, x1Indicate each mesh
The maximum abscissa of cursor position, * indicate multiplication operations, and w' indicates the width of the sample of each training；
Second step calculates the normalized value of target's center's point ordinate using following formula:
Wherein, y indicates each target's center's point ordinate, y0Indicate the minimum ordinate of each target position, y1Indicate each mesh
The maximum ordinate of cursor position, * indicate multiplication operations, and h' indicates the height of the sample of each training；
Third step calculates the normalized value of target width using following formula:
Wherein, w indicates the normalization width of each target, x1Indicate the maximum abscissa of each target position, x0Indicate each mesh
The minimum abscissa of cursor position, w' indicate the width of the sample of each training；
4th step calculates the normalized value of object height using following formula:
Wherein, h indicates the normalization width of each target, y1Indicate the maximum ordinate of each target position, y0Indicate each mesh
The minimum ordinate of cursor position, h' indicate the width of the sample of each training.
3. the remote sensing video object detection method according to claim 1 based on overlapping slice, which is characterized in that step
(4) the overlapping slice pretreatment described in refers to: being the sliding window of 608 × 608 pixels with a size, concentrates from test sample every
The top left corner apex for opening remote sensing images rises, and slides along X direction, 500 pixels are divided between sliding, obtain the overlapping of X direction
Slice；From test sample concentrates the top left corner apex of every remote sensing images, along y direction, using 500 pixels as interval, as
Every row sliding window starting point slides sliding window in X direction, 500 pixels is divided between sliding, obtains the overlapping slice of X direction；Often
Row intercepts 6 overlapping slices, and each column intercepts 3 overlapping slices；The overlapping region size of each X direction contiguous slices is 108
× 608 pixels, the overlapping region size of each y direction contiguous slices are 608 × 108 pixels；Obtained each size is
The slice of 608 × 608 pixels is known as being overlapped slice.
4. the remote sensing video object detection method according to claim 1 based on overlapping slice, which is characterized in that step
(5) the YOLO v3 model described in uses the darknet53.conv.74 convolution weight of the pre-training on Imagenet.
5. the remote sensing video object detection method according to claim 1 based on overlapping slice, which is characterized in that step
(6) object detection results of each overlapping slice described in include following three parts, are for marking each overlapping respectively
All coordinate (x that may be in containing the top left corner apex of the bounding box of mesh target area in plane coordinate system in slicec,yc)、
Coordinate (x of the lower right corner apex of each bounding box in plane coordinate systemd,yd) and each bounding box confidence level.
6. the remote sensing video object detection method according to claim 1 based on overlapping slice, which is characterized in that step
Bounding box position integration method is as follows in the slice of overlapping described in (7a):
The first step, the top left corner apex for reading each object boundary frame in each overlapping slice are in the coordinate in plane coordinate system
(xc,yc), coordinate (x of the lower right corner apex in plane coordinate systemd,yd)；Each overlapping slice is read in corresponding test sample
In top left corner apex be in the coordinate (x in plane coordinate systema,ya), coordinate of the lower right corner apex in plane coordinate system
(xb,yb)；
The top left corner apex of each object boundary frame in each overlapping slice is in plane coordinates using following formula by second step
Coordinate (x in systemc,yc) be converted to coordinate in corresponding test sample at bounding box top left corner apex in plane coordinate system
(xe,ye):
xe=xc*(xb-xa)+xa
ye=yc*(yb-ya)+ya
Wherein, xeIndicate that the top left corner apex of object boundary frame in each test sample is in the abscissa in plane coordinate system, xc
Indicate that the top left corner apex of each object boundary frame in each overlapping slice is in the abscissa in plane coordinate system, * indicates phase
Multiply operation, xbIndicate an abscissa of the overlapping slice lower right corner apex in corresponding test sample, xaIndicate each overlapping slice
Top left corner apex is in the abscissa in corresponding test sample；yeIndicate the upper left corner top of object boundary frame in each test sample
The ordinate that point is in plane coordinate system, ycIndicate that the top left corner apex of each object boundary frame in each overlapping slice is in
Ordinate in plane coordinate system, ybIndicate an ordinate of the overlapping slice lower right corner apex in corresponding test sample, yaTable
Show that each overlapping slice top left corner apex is in the ordinate in corresponding test sample；
Third step, using following formula by the lower right corner apex of each object boundary frame in each overlapping slice in plane coordinates
Coordinate (x in systemd,yd) be converted to coordinate (x of the lower right corner apex in plane coordinate system in corresponding test samplef,
yf):
xf=xd*(xb-xa)+xa
yf=yd*(yb-ya)+ya
Wherein, xfIndicate abscissa of the lower right corner apex of each object boundary frame in corresponding test sample, xdIndicate each
Abscissa of the lower right corner apex of each object boundary frame in plane coordinate system, * indicate multiplication operations, x in overlapping sliceb
Indicate an abscissa of the overlapping slice lower right corner apex in corresponding test sample, xaIndicate each overlapping slice upper left corner top
The abscissa that point is in corresponding test sample；yfIndicate the lower right corner apex of each object boundary frame in corresponding test sample
In ordinate, ydIndicate that the lower right corner apex of each object boundary frame in each overlapping slice is vertical in plane coordinate system
Coordinate, * indicate multiplication operations, ybIndicate an abscissa of the overlapping slice lower right corner apex in corresponding test sample, yaIt indicates
Each overlapping slice top left corner apex is in the abscissa in corresponding test sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811246304.0A CN109409285B (en) | 2018-10-24 | 2018-10-24 | Remote sensing video target detection method based on overlapped slices |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811246304.0A CN109409285B (en) | 2018-10-24 | 2018-10-24 | Remote sensing video target detection method based on overlapped slices |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109409285A true CN109409285A (en) | 2019-03-01 |
| CN109409285B CN109409285B (en) | 2021-11-09 |
Family
ID=65469674
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811246304.0A Active CN109409285B (en) | 2018-10-24 | 2018-10-24 | Remote sensing video target detection method based on overlapped slices |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109409285B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110765951A (en) * | 2019-10-24 | 2020-02-07 | 西安电子科技大学 | Remote sensing image airplane target detection method based on bounding box correction algorithm |
| CN111027511A (en) * | 2019-12-23 | 2020-04-17 | 西安电子科技大学 | Remote sensing image ship detection method based on region of interest block extraction |
| CN111024072A (en) * | 2019-12-27 | 2020-04-17 | 浙江大学 | Satellite map aided navigation positioning method based on deep learning |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130324830A1 (en) * | 2012-06-01 | 2013-12-05 | Xerox Corporation | Minute ventilation estimation based on depth maps |
| CN103455988A (en) * | 2013-09-17 | 2013-12-18 | 西安电子科技大学 | Super-resolution image reconstruction method based on structure self-similarity and sparse representation |
| CN105741252A (en) * | 2015-11-17 | 2016-07-06 | 西安电子科技大学 | Sparse representation and dictionary learning-based video image layered reconstruction method |
| CN107527352A (en) * | 2017-08-09 | 2017-12-29 | 中国电子科技集团公司第五十四研究所 | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks |
| CN108389220A (en) * | 2018-02-05 | 2018-08-10 | 湖南航升卫星科技有限公司 | Remote sensing video image motion target real-time intelligent cognitive method and its device |
| CN108960190A (en) * | 2018-07-23 | 2018-12-07 | 西安电子科技大学 | SAR video object detection method based on FCN Image Sequence Model |
| CN109102534A (en) * | 2018-08-29 | 2018-12-28 | 长光卫星技术有限公司 | Optical remote sensing image registration method and system under the conditions of haze weather |
| CN109242883A (en) * | 2018-08-14 | 2019-01-18 | 西安电子科技大学 | Optical remote sensing video target tracking method based on depth S R-KCF filtering |
| CN109242884A (en) * | 2018-08-14 | 2019-01-18 | 西安电子科技大学 | Remote sensing video target tracking method based on JCFNet network |
| CN109325407A (en) * | 2018-08-14 | 2019-02-12 | 西安电子科技大学 | Optical remote sensing video object detection method based on F-SSD network filtering |
| CN110390292A (en) * | 2018-10-26 | 2019-10-29 | 西安电子科技大学 | Based on the remote sensing video frequency vehicle object detecting and tracking method for dynamically associating model |
-
2018
- 2018-10-24 CN CN201811246304.0A patent/CN109409285B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130324830A1 (en) * | 2012-06-01 | 2013-12-05 | Xerox Corporation | Minute ventilation estimation based on depth maps |
| CN103455988A (en) * | 2013-09-17 | 2013-12-18 | 西安电子科技大学 | Super-resolution image reconstruction method based on structure self-similarity and sparse representation |
| CN105741252A (en) * | 2015-11-17 | 2016-07-06 | 西安电子科技大学 | Sparse representation and dictionary learning-based video image layered reconstruction method |
| CN107527352A (en) * | 2017-08-09 | 2017-12-29 | 中国电子科技集团公司第五十四研究所 | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks |
| CN108389220A (en) * | 2018-02-05 | 2018-08-10 | 湖南航升卫星科技有限公司 | Remote sensing video image motion target real-time intelligent cognitive method and its device |
| CN108960190A (en) * | 2018-07-23 | 2018-12-07 | 西安电子科技大学 | SAR video object detection method based on FCN Image Sequence Model |
| CN109242883A (en) * | 2018-08-14 | 2019-01-18 | 西安电子科技大学 | Optical remote sensing video target tracking method based on depth S R-KCF filtering |
| CN109242884A (en) * | 2018-08-14 | 2019-01-18 | 西安电子科技大学 | Remote sensing video target tracking method based on JCFNet network |
| CN109325407A (en) * | 2018-08-14 | 2019-02-12 | 西安电子科技大学 | Optical remote sensing video object detection method based on F-SSD network filtering |
| CN109102534A (en) * | 2018-08-29 | 2018-12-28 | 长光卫星技术有限公司 | Optical remote sensing image registration method and system under the conditions of haze weather |
| CN110390292A (en) * | 2018-10-26 | 2019-10-29 | 西安电子科技大学 | Based on the remote sensing video frequency vehicle object detecting and tracking method for dynamically associating model |
Non-Patent Citations (6)
| Title |
|---|
| BIAO HOU等: "Object Detection and Trcacking Based on Convolutional Neural Networks for High-Resolution Optical Remote Sensing Video", 《IGARSS 2019 - 2019 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM》 * |
| RUI LIANG等: "Aviation video moving-target detection with inter-frame difference", 《2010 3RD INTERNATIONAL CONGRESS ON IMAGE AND SIGNAL PROCESSING》 * |
| 周驿枰: "基于生成对抗网络的SAR图像生成与数据扩充研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 李井亮: "基于深度神经网络的光学遥感视频目标检测", 《中国优秀博硕士学位论文全文数据库(硕士)工程科技Ⅱ辑》 * |
| 王珍: "一种交互式视频卫星遥感图像车辆目标检测与跟踪方法", 《信息通信》 * |
| 袁益琴等: "背景差分与帧间差分相融合的遥感卫星视频运动车辆检测方法", 《中国科学院大学学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110765951A (en) * | 2019-10-24 | 2020-02-07 | 西安电子科技大学 | Remote sensing image airplane target detection method based on bounding box correction algorithm |
| CN110765951B (en) * | 2019-10-24 | 2023-03-10 | 西安电子科技大学 | Remote sensing image airplane target detection method based on bounding box correction algorithm |
| CN111027511A (en) * | 2019-12-23 | 2020-04-17 | 西安电子科技大学 | Remote sensing image ship detection method based on region of interest block extraction |
| CN111027511B (en) * | 2019-12-23 | 2022-04-29 | 西安电子科技大学 | Remote sensing image ship detection method based on region of interest block extraction |
| CN111024072A (en) * | 2019-12-27 | 2020-04-17 | 浙江大学 | Satellite map aided navigation positioning method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109409285B (en) | 2021-11-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107025652B (en) | A kind of flame detecting method based on kinetic characteristic and color space time information | |
| CN108416378B (en) | Large-scene SAR target recognition method based on deep neural network | |
| Huang et al. | A new building extraction postprocessing framework for high-spatial-resolution remote-sensing imagery | |
| CN108427912A (en) | Remote sensing image object detection method based on the study of dense target signature | |
| CN104408482B (en) | A kind of High Resolution SAR Images object detection method | |
| CN103810503B (en) | Depth study based method for detecting salient regions in natural image | |
| CN105608454B (en) | Character detecting method and system based on text structure component detection neural network | |
| CN108319949A (en) | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image | |
| CN106446933B (en) | Multi-target detection method based on contextual information | |
| CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
| CN109409256A (en) | A kind of forest rocket detection method based on 3D convolutional neural networks | |
| CN109948415A (en) | Remote sensing image object detection method based on filtering background and scale prediction | |
| CN109670429A (en) | A kind of the monitor video multiple target method for detecting human face and system of Case-based Reasoning segmentation | |
| CN109409190A (en) | Pedestrian detection method based on histogram of gradients and Canny edge detector | |
| CN108830225B (en) | Method, device, equipment and medium for detecting target object in terahertz image | |
| CN103824088B (en) | SAR target variant recognition method based on multi-information joint dynamic sparse representation | |
| CN109409285A (en) | Remote sensing video object detection method based on overlapping slice | |
| CN107480607B (en) | Method for detecting and positioning standing face in intelligent recording and broadcasting system | |
| CN105608691A (en) | High-resolution SAR image individual building extraction method | |
| CN107403433A (en) | A kind of complicated cloud infrared small target in background detection method | |
| CN109636771A (en) | Airbound target detection method and system based on image procossing | |
| CN103279752B (en) | A kind of eye locating method based on improving Adaboost algorithm and Face geometric eigenvector | |
| CN106054189B (en) | Radar target identification method based on dpKMMDP models | |
| CN108764310A (en) | SAR target identification methods based on multiple dimensioned multiple features depth forest | |
| CN109190624A (en) | Kitchen fume concentration detection method based on image procossing |
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 |