CN110163141A - Satellite image preprocess method based on genetic algorithm - Google Patents
Satellite image preprocess method based on genetic algorithm Download PDFInfo
- Publication number
- CN110163141A CN110163141A CN201910407112.1A CN201910407112A CN110163141A CN 110163141 A CN110163141 A CN 110163141A CN 201910407112 A CN201910407112 A CN 201910407112A CN 110163141 A CN110163141 A CN 110163141A
- Authority
- CN
- China
- Prior art keywords
- chromosome
- image
- rgb
- msi
- multispectral
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Genetics & Genomics (AREA)
- Physiology (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Multimedia (AREA)
- Remote Sensing (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Astronomy & Astrophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of satellite image preprocess method based on genetic algorithm mainly solves the problems, such as to cannot achieve classification balance in the prior art, and implementation is: 1) satellite image carries out the pretreatment of label denoising and data enhancing；2) atural object classification balance is carried out to pretreated satellite figure；3) satellite image after fusion classification balance, generates training sample；4) using training sample training semantic segmentation model；5) shadow positions detection is carried out to satellite image test sample collection；6) satellite image test set image is merged, it is detected using semantic segmentation model；7) the semantic segmentation result detected using the pixel value modification of shadow positions.The present invention solves satellite image classification equilibrium problem, and instructs semantic segmentation result by the pixel value of shadow positions, hence it is evident that improves satellite image semantic segmentation precision, can be used in deep learning classifying and dividing the data prediction of task.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to a kind of satellite image pretreatment side based on genetic algorithm
Method can be used in deep learning, the classification of satellite image and the preprocessing part for dividing task.
Background technique
With the growth of satellite image data, deep learning using more and more extensive, is defended in satellite image process field
Star satellite image pre-processes the performance for also drastically influencing deep learning.
Satellite image pretreatment generally includes classification balance, data enhancing etc..Currently, common satellite image classification
Balance method has: 1) the more carry out down-sampling of categorical measure, a part of satellite photo of random drop, keeps categorical measure difference most
Amount balance；2) what categorical measure was less up-sample, and is overturn, rotated, scaled, cut, translates, increased to satellite photo
The operation such as this noise, is added in data set, balances categorical measure difference as far as possible；3) the training weight in model, classification are modified
Biggish weight is arranged in the data of negligible amounts, and lesser trained weight is arranged in the biggish data of categorical measure, makes to train
Model accuracy integrally reaches higher accuracy rate.
But these above-mentioned traditional satellite image data balance methods, cannot balance the ground species in satellite image well
Not.When the distribution of the atural object classification in satellite photo is not according still further to a certain specific rule, using picture overturning, scaling, cut,
When the simple datas Enhancement Methods such as translation, it will while increasing or decreasing a certain classification atural object quantity, generate another ground
The case where other quantity of species can also change therewith cannot reach the evenly other effect of species.
The patent of Publication No. CN102495901B, which proposes, a kind of to be kept realizing class data balancing by local mean value
Method can balance the atural object classification in satellite image by k-means algorithm.But this method is in the satellite mapping of processing large scene
When picture, cannot the farther away two kinds of atural object of equilibrium distance well, preferable counterbalance effect is not achieved.
Summary of the invention
It is an object of the invention in view of the above shortcomings of the prior art, provide a kind of satellite image based on genetic algorithm
Preprocess method, to improve the counterbalance effect of the farther away two kinds of atural object of distance in large scene satellite image, and will be containing different logical
The image in road carries out channel fusion, obtains the image containing more information.
To achieve the above object, technical solution of the present invention includes the following:
(1) training dataset in satellite image, including RGB training dataset and the multispectral MSI training of eight wave bands are read
Data set, and the preliminary treatment of identical label denoising and data enhancing is carried out to the image in both data sets；
(2) atural object classification balance is carried out to the RGB training dataset after preliminary treatment based on genetic algorithm, obtains classification point
RGB data collection after cloth balance, and concentrate to choose from the multispectral MSI training data of eight wave bands and be concentrated with the RGB data after balance
The multispectral MSI data set of eight wave bands after the same image Compositional balance of title；
(3) RGB image concentrated the RGB data after balance and phase in the multispectral MSI data set of eight wave bands after balance
The multispectral MSI image of eight wave bands answered carries out channel fusion, generates the training sample of new tunnel；
(4) training sample is sent into existing image cascade network ICNet to be trained, obtains trained semantic segmentation
Model；
(5) test data set in satellite image, including RGB test data set and the multispectral MSI test of eight wave bands are read
Data set, and shadow positions detection is carried out to the multispectral MSI test data set of eight wave band therein, take out the position of dash area
It sets；
(6) it is new that RGB test data set is merged to generation according to the method for (3) with the correspondence image that MSI test data is concentrated
Test sample, and be sent to the semantic segmentation model obtained by (4), obtain semantic segmentation result；
(7) pixel value for the shadow positions for using (5) to obtain is modified the semantic segmentation result that (6) obtain, and obtains
Semantic segmentation after satellite image optimizes is as a result, complete the pretreatment to satellite image.
In conclusion advantage of the invention is as follows:
First, present invention employs the satellite image preprocess methods based on genetic algorithm, carry out data to training sample
Denoising and enhancing carry out atural object classification balance using genetic algorithm, can accurately and quickly evenly species are other, it is flat to obtain classification
Weighing apparatus and the stronger satellite image data collection of generalization ability.
Second, the present invention utilizes the characteristic of satellite image, including RGB image and the multispectral MSI image of eight wave bands, by two kinds
Various forms of satellite images carry out channel fusion, obtain a kind of information new images more abundant, help to improve semantic point
The precision cut.
Third, the present invention utilize the characteristic of the multispectral MSI image of eight wave bands, extract its near infrared band, red edge wave
Section and 3 channels of yellow band, normalization forms visual image respectively, is extracted using the pixel value of the visual image
Shadow positions, and with the pixel value modification semantic segmentation of shadow positions as a result, improving semantic segmentation precision.
Detailed description of the invention:
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is the sub-process figure that genetic algorithm equilibrium data classification is utilized in the present invention；
Fig. 3 is that shadow positions extract sub-process figure in the present invention；
Fig. 4 is that shadow positions pixel correction semantic segmentation result sub-process figure is used in the present invention.
Specific implementation step
It describes in detail below in conjunction with attached drawing to the present invention:
Referring to Fig.1, steps are as follows for realization of the invention:
Step 1: satellite image preliminary treatment.
1.1) training dataset in satellite image, including RGB training dataset and the multispectral MSI training of eight wave bands are read
Data set；
1.2) it deletes RGB training data and concentrates the image containing a large amount of unmarked labels, the RGB training number after being denoised
According to collection, and leave out the image that the multispectral MSI training data of eight wave bands concentrates same names, eight wave bands after being denoised are multispectral
MSI training dataset；
1.3) the RGB training data after statistics denoising concentrates the pixel quantity c [m] of each classification, calculates RGB training data
Collect the summation C of all categories pixel quantity:
Wherein, m is category label, and m ∈ M, M are the classification sum of the satellite image；
1.4) the summation C that RGB training data concentrates each classification pixel quantity c [m] and all categories pixel quantity is calculated
Ratio:
1.5) to satisfactionThe transformation that is first horizontally or vertically overturn of classification, then carry out brightness or saturation degree
Transformation, the RGB training dataset after obtaining preliminary treatment；
1.6) the multispectral MSI training dataset of eight wave bands after denoising is equally carried out 1.5) operating, obtains preliminary treatment
The multispectral MSI training dataset of eight wave bands afterwards.
Step 2: atural object classification balance being carried out to satellite image using genetic algorithm.
Referring to Fig. 2, this step is implemented as follows:
2.1) random initializtion item chromosome:
It is 1~n that the image that RGB training data after preliminary treatment is concentrated is numbered in sequence, and number represents the figure
As being generated at random with chromosome i-th bit " 0 " or " 1 " in the position of every chromosome, i ∈ [1, n], n is after preliminary treatments
The quantity of RGB training data concentration image；
2.2) circulation executes 2.1) f times, generates the population that a chromosome quantitative is f, and f is set as 10~50；
2.3) it randomly chooses two chromosomes from population of new generation to be intersected and corrected, the population after being intersected:
2.3.1 p) is generated at randomc∈ [0,1], pcFor crossover probability, if pc> 0.6 then randomly chooses a wherein dyeing
One section of chromosome between certain two position of body is switching part, is handed over the chromosome of another item chromosome same position
It changes, the chromosome after obtaining two intersections；
2.3.2 the chromosome after an intersection) is selected, the number a of wherein character " 1 " is counted1, and by the RGB after balance
Amount of images in data set is denoted as a, and according to a1The value on the chromosome is modified with the difference of a and replaces population
In original chromosome:
If a1- a=0 then retains this chromosome；
If a1- a > 0, then randomly select a1- a characters are the position of " 1 ", are revised as " 0 "；
If a1- a < 0, then randomly select a-a1A character is the position of " 0 ", is revised as " 1 "；
2.3.3 the chromosome after) intersecting to other one equally executes step 2.3.2)；
2.3.4 prechiasmal two chromosome, the population after being intersected) are replaced using two revised chromosomes.
2.4) it randomly chooses item chromosome from the population after intersection to be made a variation and corrected, the kind after being made a variation
Group:
2.4.1 p) is generated at randomm∈ [0,1], pmFor mutation probability, if pm> 0.6, then randomly choose on item chromosome
Any position, if the position be " 0 ", become " 1 ", if " 1 ", then become " 0 ", the chromosome after being made a variation；
2.4.2) in the chromosome after statistical variation or dispersion character " 1 " number a3, and the figure that the RGB data after balance is concentrated
As quantity is denoted as a；
2.4.3) according to a3The chromosome after variation is modified with the difference of a:
If a3- a=1 then randomly selects the position that 1 character is " 1 ", is revised as " 0 "；
If a1- a=-1 then randomly selects the position that 1 character is " 0 ", is revised as " 1 "；
2.4.4) using the chromosome before the replacement variation of revised chromosome, the population after being made a variation；
2.5) f individual is selected to be left child chromosome from the population after variation in a manner of roulette:
2.5.1 the quantity c' of all kinds of atural object classifications in every chromosome) is countedj[m], and calculate in the chromosome and include
The mean value of all categories:
Wherein,For the mean value of all categories in j-th strip chromosome, m is category label, and M is the classification of the satellite image
Sum；
2.5.2 the variance yields var [j] of all categories in every chromosome) is calculated:
2.5.3 the fitness) being mapped as variance in genetic algorithm:
Wherein, var [j] is the variance yields of all categories in every chromosome, and fitness [j] indicates j-th strip chromosome
Fitness, max (var) indicate the maximum value of chromosome variance in every generation, and f is population quantity；
2.5.4 the accumulated probability q [s] of the every chromosome needed in the selection strategy of roulette) is calculated:
Wherein, p [s] is the probability that every chromosome is selected, and is expressed as follows:
2.5.5 the number of a k ∈ [0,1]) is generated at random, ifThen the s bars chromosome quilt
Choose reservation；
2.5.6 2.5.5) is repeated) it is f times total, f genome is selected into population of new generation；
2.6) the smallest chromosome h' of variance in population of new generation is recorded, optimal chromosome h is updated:
Wherein, h' represents current best chromosome, and h represents the smallest chromosome of variance yields that roulette retains later, n generation
Table the number of iterations, var [h] represent the variance yields of all categories in chromosome h, and var [h'] represents all categories in chromosome h'
Variance yields；
2.7) step 2.3) -2.6 is repeated), stop until the number of iterations reaches 10000 times；
2.8) optimal chromosome is exported, and after balance is added in the corresponding image in position that character in chromosome is " 1 "
RGB data is concentrated.
Step 3: two kinds of image channel fusions generate the training sample of new tunnel.
Since the display effect of vegetation can be enhanced in the near infrared band (V-NIR) of eight wave band multispectral images (MSI), because
RGB image is carried out channel with the multispectral MSI image of eight wave bands and merged by this, generates new tunnel training sample, the training sample packet
The information contained is more abundant.
This step is implemented as follows:
3.1) will be added with the near infrared band (V-NIR) of the channel G of RGB image and eight wave band multispectral image MSI
Power fusion, obtains new tunnel CHnew:
CHnew=CH (G) × P+CH (V-NIR) × (1-P),
Wherein, P is weight, takes 0.8；
3.2) the original channel RGB image G is replaced with into CHnew, retain the channel R and channel B, generate new image.
Step 4: using training sample training image cascade network ICNet.
In order to which whether verifying satellites image preprocessing process have castering action to semantic segmentation precision, by training sample
It is sent into existing image cascade network ICNet to be trained, obtains trained semantic segmentation model, be accomplished by
4.1) the inside weight of random initializtion image cascade network ICNet；
4.2) training sample is sent into image cascade network ICNet, after all training samples pass fully through network, image
Cascade network ICNet modifies to its inside weight automatically according to the penalty values exported at the end of this；
4.3) it repeats 4.2), until the floating range of the loss function value of 20 image cascade network ICNet output
No more than ± 0.5, then save using last time weight as the semantic segmentation model of result.
Step 5: detecting and record shadow positions.
Since the influence of dash area semantic segmentation result is more serious, so to eight wave bands in satellite image test set
Multispectral MSI image carries out shadow positions detection, for instructing the semantic segmentation result of part of detecting.
Referring to Fig. 3, this step is implemented as follows:
5.1) near infrared band in the extraction multispectral MSI image of eight wave bands, red edge wave band and yellow band 3 are logical
Road；
5.2) numberical range of near infrared band is normalized to [0,255], and be rounded downwards, be set to the first of new images
A channel；
5.3) numberical range of red edge wave band is normalized to [0,255], and be rounded downwards, be set to the of new images
Two channels；
5.4) numberical range of yellow band is normalized to [0,255], and be rounded downwards, be set to the third of new images
Channel；
5.5) it finds and records in the new images of above three channel generation pixel value in (0,0,0) to (10,21,16) model
Enclose interior position, as shadow positions.
Step 6: the semantic segmentation model obtained using step 4 tests test sample.
According to the method for step 3 by pair in the RGB image and the multispectral MSI image of eight wave bands in satellite image test set
It answers image co-registration to generate new test sample, and is sent to the semantic segmentation model obtained by step 4, obtain semantic segmentation knot
Fruit.
Step 7: semantic segmentation result being modified using the pixel value of shadow positions.
The pixel value of the shadow positions obtained using step 5 is modified the semantic segmentation result that step 6 obtains, and obtains
Semantic segmentation after satellite image optimizes is as a result, complete the pretreatment to satellite image.
Referring to Fig. 4, this step is implemented as follows:
7.1) pixel value for the shadow positions that step 5 obtains is taken out；
7.2) the semantic segmentation result obtained according to the pixel value modification step 6 of shadow positions:
If the pixel value of shadow positions is (0,255,64~100), then the corresponding position of semantic segmentation result is corrected
For building；
If the pixel value of shadow positions is (0,255,127), then the corresponding position of semantic segmentation result is modified to ground
Face；
If the pixel value of shadow positions is (0,0,127), then the corresponding position of semantic segmentation result is modified to Gao Zhi
Quilt.
The above is only a specific embodiment of the invention, but scope of protection of the present invention is not limited thereto, any to be familiar with
Those skilled in the art in the technical scope disclosed by the present invention, can readily occur in various equivalent modifications or substitutions,
These modifications or substitutions should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be wanted with right
Subject to the protection scope asked.
Claims (10)
1. a kind of satellite image preprocess method based on genetic algorithm, which is characterized in that include the following:
(1) training dataset in satellite image, including RGB training dataset and the multispectral MSI training data of eight wave bands are read
Collection, and the preliminary treatment of identical label denoising and data enhancing is carried out to the image in both data sets；
(2) atural object classification balance is carried out to the RGB training dataset after preliminary treatment based on genetic algorithm, it is flat obtains category distribution
RGB data collection after weighing apparatus, and concentrate to choose from the multispectral MSI training data of eight wave bands and concentrate title with the RGB data after balance
The multispectral MSI data set of eight wave bands after the same image Compositional balance；
(3) RGB image concentrated the RGB data after balance is corresponding with the multispectral MSI data set of eight wave bands after balance
The multispectral MSI image of eight wave bands carries out channel fusion, generates the training sample of new tunnel；
(4) training sample is sent into existing image cascade network ICNet to be trained, obtains trained semantic segmentation mould
Type；
(5) test data set in satellite image, including RGB test data set and the multispectral MSI test data of eight wave bands are read
Collection, and shadow positions detection is carried out to the multispectral MSI test data set of eight wave band therein, take out the position of dash area；
(6) RGB test data set is merged according to the method for (3) with the correspondence image that MSI test data is concentrated and generates new survey
Sample sheet, and it is sent to the semantic segmentation model obtained by (4), obtain semantic segmentation result；
(7) pixel value for the shadow positions for using (5) to obtain is modified the semantic segmentation result that (6) obtain, and obtains through defending
Semantic segmentation after star image optimization is as a result, complete the pretreatment to satellite image.
2. the method according to claim 1, wherein multispectral to RGB training dataset and eight wave bands in (1)
The image of MSI training dataset carries out preliminary treatment, is accomplished by
(1a) deletes RGB training dataset and the multispectral MSI training data of eight wave bands concentrates the number containing a large amount of unmarked labels
According to carrying out denoising operation to data；
(1b) statistics RGB training data concentrates the pixel quantity c [m] of each classification, calculates RGB training dataset all categories picture
The summation C of prime number amount:
Wherein, m is category label, and m ∈ M, M are the classification sum of the satellite image；
(1c) calculates the ratio that RGB training data concentrates the summation C of each classification pixel quantity c [m] and all categories pixel quantity
Value:
(1d) is to satisfactionThe transformation that is first horizontally or vertically overturn of classification, then carry out brightness or the change of saturation degree
It changes, data is enhanced with this.
(1e) equally carries out (1d) operation to the multispectral MSI training dataset of eight wave bands after denoising, after obtaining preliminary treatment
The multispectral MSI training dataset of eight wave bands.
3. the method according to claim 1, wherein being instructed based on genetic algorithm to the RGB after preliminary treatment in (2)
Practice data set and carry out atural object classification balance, is accomplished by
(2a) random initializtion item chromosome:
It is 1~n that the image that RGB training data after preliminary treatment is concentrated is numbered in sequence, and number represents the image and exists
The position of every chromosome is generated " 0 " or " 1 " at random with chromosome i-th bit, and i ∈ [1, n], n are the RGB instruction after preliminary treatment
Practice the quantity of image in data set；
(2b) circulation executes (2a) f times, generates the population that a chromosome quantitative is f, and f is set as 10~50；
(2c) is intersected and is corrected, the population after being intersected from population of new generation two chromosomes of random selection；
(2d) randomly chooses item chromosome from the population after intersection and is made a variation and corrected, the population after being made a variation；
(2e) selects f individual to be left population of new generation in a manner of roulette from the population after variation；
(2f) records the smallest chromosome h' of variance in population of new generation, updates optimal chromosome h:
Wherein, h' represents current best chromosome, and h represents the smallest chromosome of variance yields that roulette retains later, and n representative changes
Generation number, var [h] represent the variance yields of all categories in chromosome h, and var [h'] represents the side of all categories in chromosome h'
Difference；
(2g) repeats step (2c)-(2f), stops until the number of iterations reaches 10000 times.
(2h) exports optimal chromosome, and the corresponding image in position that character in chromosome is " 1 " is added to the RGB number after balance
According to concentration.
4. according to the method described in claim 3, it is characterized in that, two dyeing are randomly choosed in (2c) from population of new generation
Body is intersected and is corrected, and is accomplished by
(2c1) generates p at randomc∈ [0,1], pcFor crossover probability, if pc> 0.6 then randomly chooses wherein item chromosome certain
One section of chromosome between two positions is switching part, swaps, obtains with the chromosome of another item chromosome same position
Chromosome after intersecting to two；
(2c2) selects the chromosome after an intersection, counts the number a of wherein character " 1 "1, and by the RGB data collection after balance
In amount of images be denoted as a, and according to a1The value on the chromosome is modified with the difference of a and is replaced original in population
Chromosome:
If a1- a=0 then retains this chromosome；
If a1- a > 0, then randomly select a1- a characters are the position of " 1 ", are revised as " 0 "；
If a1- a < 0, then randomly select a-a1A character is the position of " 0 ", is revised as " 1 "；
Chromosome after (2c3) intersects other one equally executes step (2d2)；
(2c4) replaces prechiasmal two chromosome, the population after being intersected using two revised chromosomes.
5. according to the method described in claim 3, it is characterized in that, a dye is randomly choosed in (2d) from the population after intersection
Colour solid is made a variation and is corrected, and is accomplished by
(2d1) generates p at randomm∈ [0,1], pmFor mutation probability, if pm> 0.6 then randomly chooses any on item chromosome
Position becomes " 1 ", if " 1 ", then becomes " 0 ", the chromosome after being made a variation if the position is " 0 "；
The number a of character " 1 " in chromosome after (2d2) statistical variation or dispersion3, and the amount of images that the RGB data after balance is concentrated
It is denoted as a；
(2d3) is according to a3The chromosome after variation is modified with the difference of a:
If a3- a=1 then randomly selects the position that 1 character is " 1 ", is revised as " 0 "；
If a1- a=-1 then randomly selects the position that 1 character is " 0 ", is revised as " 1 "；
(2d4) replaces the chromosome before variation, the population after being made a variation using revised chromosome.
6. according to the method described in claim 3, it is characterized in that, (2e) is selected from the population after variation in a manner of roulette
It selects f individual and is left child chromosome, be accomplished by
(2e1) counts the quantity c' of all kinds of atural object classifications in every chromosomej[m], and calculate all classes for including in the chromosome
Other mean value:
Wherein,For the mean value of all categories in j-th strip chromosome, m is category label, and M is the classification sum of the satellite image；
(2e2) calculates the variance yields var [j] of all categories in every chromosome:
Variance is mapped as the fitness in genetic algorithm by (2e3):
Wherein, var [j] is the variance yields of all categories in every chromosome, and fitness [j] indicates the adaptation of j-th strip chromosome
Degree, max (var) indicate the maximum value of chromosome variance in every generation, and f is population quantity；
The accumulated probability of the every chromosome needed in the selection strategy of (2e4) calculating roulette:
Wherein, p [s] is the probability that every chromosome is selected, and is expressed as follows:
(2e5) generates the number of a k ∈ [0,1] at random, ifThen the s articles chromosome is selected
Retain；
(2e6) repeat (2e5) it is f times total, select f genome into population of new generation.
7. the method according to claim 1, wherein the RGB image for concentrating the RGB data after balance in (3)
Channel fusion is carried out with the multispectral MSI image of eight wave band corresponding in the multispectral MSI data set of eight wave bands after balance, is realized such as
Under:
(3a) will be weighted with the near infrared band (V-NIR) in the channel G of RGB image and eight wave band multispectral image MSI to be melted
It closes, obtains new tunnel CHnew:
CHnew=CH (G) × P+CH (V-NIR) × (1-P)
Wherein, P is weight, takes 0.8；
The original channel RGB image G is replaced with CH by (3b)new, retain the channel R and channel B, generate new image.
8. the method according to claim 1, wherein training sample is sent into existing image cascade network in (4)
Network ICNet is trained, and is accomplished by
The inside weight of (4a) random initializtion image cascade network ICNet；
Training sample is sent into image cascade network ICNet by (4b), after all training samples pass fully through network, image cascade
Network ICNet modifies to its inside weight automatically according to the penalty values exported at the end of this；
(4c) repeats (4b), until the floating range of the loss function value of 20 image cascade network ICNet output does not surpass
± 0.5 is crossed, then is saved using last time weight as the semantic segmentation model of result.
9. the method according to claim 1, wherein MSI test data set multispectral to eight wave bands carries out in (5)
Shadow positions detection, takes out the position of dash area, is accomplished by
3 near infrared band, red edge wave band and yellow band channels in (5a) extraction multispectral MSI image of eight wave bands；
The numberical range of near infrared band is normalized to [0,255] by (5b), and is rounded downwards, and first for being set to new images is logical
Road；
The numberical range of red edge wave band is normalized to [0,255] by (5c), and is rounded downwards, and second of new images is set to
Channel；
The numberical range of yellow band is normalized to [0,255] by (5d), and is rounded downwards, and the third for being set to new images is logical
Road；
(5e) finds and records pixel value in the new images of above three channel generation and arrive in (10,21,16) range in (0,0,0)
Position, as shadow positions.
10. the method according to claim 1, wherein the pixel value for the shadow positions for using (5) to obtain in (7)
The semantic segmentation result that (6) obtain is modified, is accomplished by
(7a) takes out the pixel value for the shadow positions that (5) obtain；
The semantic segmentation result that (7b) is obtained according to the pixel value modification (6) of shadow positions:
If the pixel value of shadow positions is (0,255,64~100), then the corresponding position of semantic segmentation result is modified to and is built
Build object；
If the pixel value of shadow positions is (0,255,127), then the corresponding position of semantic segmentation result is modified to ground；
If the pixel value of shadow positions is (0,0,127), then the corresponding position of semantic segmentation result is modified to high vegetation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910407112.1A CN110163141B (en) | 2019-05-16 | 2019-05-16 | Satellite image preprocessing method based on genetic algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910407112.1A CN110163141B (en) | 2019-05-16 | 2019-05-16 | Satellite image preprocessing method based on genetic algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110163141A true CN110163141A (en) | 2019-08-23 |
| CN110163141B CN110163141B (en) | 2023-04-07 |
Family
ID=67634728
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910407112.1A Active CN110163141B (en) | 2019-05-16 | 2019-05-16 | Satellite image preprocessing method based on genetic algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110163141B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112215815A (en) * | 2020-10-12 | 2021-01-12 | 杭州视在科技有限公司 | Bare soil coverage automatic detection method for construction site |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090175514A1 (en) * | 2004-11-19 | 2009-07-09 | Koninklijke Philips Electronics, N.V. | Stratification method for overcoming unbalanced case numbers in computer-aided lung nodule false positive reduction |
| US20120269430A1 (en) * | 2011-04-22 | 2012-10-25 | Michael Paul Deskevich | System and method for combining color information with spatial information in multispectral images |
| US8761506B1 (en) * | 2011-04-22 | 2014-06-24 | DigitalGlobe, Incorporated | Pan sharpening digital imagery |
| CN107862667A (en) * | 2017-11-23 | 2018-03-30 | 武汉大学 | A kind of city shadow Detection and minimizing technology based on high-resolution remote sensing image |
| CN108632279A (en) * | 2018-05-08 | 2018-10-09 | 北京理工大学 | A kind of multilayer method for detecting abnormality based on network flow |
| CN109101943A (en) * | 2018-08-27 | 2018-12-28 | 寿带鸟信息科技（苏州）有限公司 | It is a kind of for detecting the machine vision method of Falls Among Old People |
-
2019
- 2019-05-16 CN CN201910407112.1A patent/CN110163141B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090175514A1 (en) * | 2004-11-19 | 2009-07-09 | Koninklijke Philips Electronics, N.V. | Stratification method for overcoming unbalanced case numbers in computer-aided lung nodule false positive reduction |
| US20120269430A1 (en) * | 2011-04-22 | 2012-10-25 | Michael Paul Deskevich | System and method for combining color information with spatial information in multispectral images |
| US8761506B1 (en) * | 2011-04-22 | 2014-06-24 | DigitalGlobe, Incorporated | Pan sharpening digital imagery |
| CN107862667A (en) * | 2017-11-23 | 2018-03-30 | 武汉大学 | A kind of city shadow Detection and minimizing technology based on high-resolution remote sensing image |
| CN108632279A (en) * | 2018-05-08 | 2018-10-09 | 北京理工大学 | A kind of multilayer method for detecting abnormality based on network flow |
| CN109101943A (en) * | 2018-08-27 | 2018-12-28 | 寿带鸟信息科技（苏州）有限公司 | It is a kind of for detecting the machine vision method of Falls Among Old People |
Non-Patent Citations (2)
| Title |
|---|
| 王生生等: "基于轻量和积网络及无人机遥感图像的大豆田杂草识别", 《农业工程学报》 * |
| 韩闪闪等: "高分辨率遥感影像融合研究", 《测绘科学》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112215815A (en) * | 2020-10-12 | 2021-01-12 | 杭州视在科技有限公司 | Bare soil coverage automatic detection method for construction site |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110163141B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114120037B (en) | Germinated potato image recognition method based on improved yolov5 model | |
| CN110443143A (en) | The remote sensing images scene classification method of multiple-limb convolutional neural networks fusion | |
| CN100573523C (en) | A kind of image inquiry method based on marking area | |
| CN105631415A (en) | Video pedestrian recognition method based on convolution neural network | |
| CN105938564A (en) | Rice disease recognition method based on principal component analysis and neural network and rice disease recognition system thereof | |
| CN104463128B (en) | Eyeglass detection method and system for recognition of face | |
| CN110414387A (en) | A kind of lane line multi-task learning detection method based on lane segmentation | |
| CN108009629A (en) | A kind of station symbol dividing method based on full convolution station symbol segmentation network | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN107862239A (en) | A kind of combination text carries out the method and its device of picture recognition with picture | |
| CN106446957A (en) | Haze image classification method based on random forest | |
| CN109145960A (en) | Based on the data characteristics selection method and system for improving particle swarm algorithm | |
| CN109815357A (en) | A kind of remote sensing image retrieval method based on Nonlinear Dimension Reduction and rarefaction representation | |
| CN108681715A (en) | Hainan Tropical wildwood vegetation type group sorting technique | |
| CN110287882A (en) | A kind of big chrysanthemum kind image-recognizing method based on deep learning | |
| CN109272487A (en) | The quantity statistics method of crowd in a kind of public domain based on video | |
| CN106845513A (en) | Staff detector and method based on condition random forest | |
| CN104952067A (en) | Method for segmenting color images on basis of NSGA-II (non-dominated sorting genetic algorithm-II) evolution algorithms | |
| CN113887410A (en) | Deep learning-based multi-category food material identification system and method | |
| CN108734717A (en) | The dark weak signal target extracting method of single frames star chart background based on deep learning | |
| CN109978074A (en) | Image aesthetic feeling and emotion joint classification method and system based on depth multi-task learning | |
| CN111814563A (en) | Method and device for classifying planting structures | |
| CN114758132B (en) | Fruit tree disease and pest identification method and system based on convolutional neural network | |
| CN115631462A (en) | AM-YOLOX-based strawberry disease and pest detection method and system | |
| CN110163141A (en) | Satellite image preprocess method based on genetic algorithm |
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 |