CN110298271A - Seawater method for detecting area based on critical point detection network and space constraint mixed model - Google Patents
Seawater method for detecting area based on critical point detection network and space constraint mixed model Download PDFInfo
- Publication number
- CN110298271A CN110298271A CN201910519569.1A CN201910519569A CN110298271A CN 110298271 A CN110298271 A CN 110298271A CN 201910519569 A CN201910519569 A CN 201910519569A CN 110298271 A CN110298271 A CN 110298271A
- Authority
- CN
- China
- Prior art keywords
- sea
- sea horizon
- image
- horizon
- pixel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 46
- 239000013535 sea water Substances 0.000 title claims abstract description 41
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000009826 distribution Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 14
- 230000011218 segmentation Effects 0.000 claims abstract description 13
- 238000013528 artificial neural network Methods 0.000 claims abstract description 5
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 11
- 230000004888 barrier function Effects 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000000694 effects Effects 0.000 claims description 4
- 239000004744 fabric Substances 0.000 claims description 3
- 238000007500 overflow downdraw method Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 230000002411 adverse Effects 0.000 claims description 2
- 238000005315 distribution function Methods 0.000 claims description 2
- 238000003709 image segmentation Methods 0.000 claims description 2
- 238000012360 testing method Methods 0.000 claims description 2
- 238000000354 decomposition reaction Methods 0.000 claims 1
- 230000004927 fusion Effects 0.000 claims 1
- 230000006870 function Effects 0.000 description 8
- 238000004364 calculation method Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 238000005036 potential barrier Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000004438 eyesight Effects 0.000 description 1
- 230000004377 improving vision Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000000465 moulding Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000004310 photopic vision Effects 0.000 description 1
- 238000005381 potential energy Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 230000016776 visual perception Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of seawater method for detecting area based on critical point detection network and space constraint mixed model, comprising the following steps: (1) acquire sea image pattern, construct sea horizon key point training sample；(2) sea horizon key point training sample training critical point detection network is utilized；(3) using sea image to be detected as input, pass through the critical point detection neural network forecast sea horizon key point trained；(4) sea horizon key point is connected, corresponding sea horizon is obtained；(5) mixed model for establishing space constraint is initialized by reference pair model parameter and prior distribution of sea horizon；(6) semantic segmentation is carried out to sea image by expectation-maximization algorithm, to detect seawater region.The seawater region under complex background can be effectively detected out in method of the invention, and has the characteristics that speed is fast, robustness is good.
Description
Technical field
The present invention relates to technical field of image processing, and in particular to one kind is mixed based on critical point detection network and space constraint
The seawater method for detecting area of molding type.
Background technique
Seawater region detection technique can provide a travelable region for unmanned water surface ship, to provide road for independent navigation
Diameter planning auxiliary, is the key components of its environment sensing.With the rise of ocean development and the hair of unmanned technology
Exhibition, the seawater region detection core technology indispensable as unmanned water surface ship also need to have breakthrough.Under sail, unmanned water
Face ship helps to improve the safety of driving to the enhancing of scene perception ability around.Therefore, seawater region detection technique is studied
There is critically important realistic meaning.
Currently, unmanned water surface ship mainly has laser radar, visible light wave range and infrared waves to the cognition technology of scene around
The Vision imaging system etc. of section.Wherein, it due to including target area detailed information abundant in optical imagery, is regarded with visible light
Feel that the intelligence system based on imaging is easier to perceive complicated Ocean Scenes, and position is made accurately
Judgement.In addition, it also executes the various activities including analytical judgment and monitoring and reconnaissance activities with important to the mankind
Booster action.In recent years, the seawater region detection research based on photopic vision is not also very much.2016, Kristan
Et al. in " Fast Image-Based Obstacle Detection From Unmanned Surface Vehicles "
Propose a kind of semantic segmentation method based on probability graph model.This method mainly passes through three Gaussian Profiles and one is uniformly divided
Cloth describes sky, haze/seashore Mixed Zone, seawater region and potential barrier region (singular value region) respectively,
And pass through the semantic segmentation of expectation-maximization algorithm (EM algorithm) sea Lai Shixian image.This method detection performance preferably, speed
Comparatively fast, but the initial method of its model parameter is too simple, only simply on the image along vertical direction in proportion
Mark off three regions indicate sky, intermediate Mixed Zone, seawater region observation data and corresponding height is calculated with this
This distribution initial parameter.When flating is larger, the seawater region marked off in proportion may includes more centre
The pixel of Mixed Zone, and it is not accurate enough with the Gaussian Distribution Parameters of this calculated seawater classification, eventually lead to sea image
Semantic segmentation result there are large errors.2017, in order to improve the sea image semantic segmentation side of Kristan et al. proposition
The deficiency of method, Bovcon et al. is in " Improving vision-based obstacle detection on USV using
Inertial sensor " in propose a kind of semantic segmentation method of combination inertial navigation sensor.This method is first with inertial navigation number
According to calculating rough sea horizon, and using this sea horizon as reference pair sky, intermediate Mixed Zone, seawater area classification Gauss
Distribution parameter is initialized.The method improve the performances of Kristan et al. semantic segmentation method proposed still to work as nothing
When people's water surface ship is closer apart from seashore, this method can be located in intermediate Mixed Zone using the sea horizon of inertial guidance data estimation, and
It is not located at the intersection of intermediate Mixed Zone and seawater region, to cause initial model parameter not accurate enough.In addition, should
Method needs the support of inertial navigation sensor, increases the cost of unmanned water surface ship visual perception system.
Summary of the invention
In view of the deficiencies of the prior art, the present invention provides one kind to be based on critical point detection network and space constraint mixed model
Seawater method for detecting area, this method can be effectively detected out the seawater region under complex background, and have speed it is fast,
The good feature of robustness.
In order to achieve the above objectives, the present invention adopts the following technical scheme:
A kind of seawater method for detecting area based on critical point detection network and space constraint mixed model, including following step
It is rapid:
(1) sea image pattern is acquired, sea horizon key point training sample is constructed；
(2) sea horizon key point training sample training critical point detection network is utilized；
(3) it using sea image to be detected as input, is closed by the critical point detection neural network forecast sea horizon trained
Key point；
(4) sea horizon key point is connected, corresponding sea horizon is obtained；
(5) mixed model for establishing space constraint carries out just by reference pair model parameter and prior distribution of sea horizon
Beginningization；
(6) semantic segmentation is carried out to sea image by expectation-maximization algorithm (EM), to detect seawater region.
Further, in the step (1), the N sea color images of camera acquisition carried using unmanned water surface ship are used
Annotation tool manually marks acquired image, forms sea horizon key point training set T={ t1, t2,…,tN}.Its
In, each training sample tiBy sea image IiO is constituted with corresponding sea horizon key pointi, i.e. ti={ Ii,Oi}.For sea
Antenna key point OiMark, detailed process are as follows: equidistantly mark n vertical line along the vertical direction of image, then utilize mark
Note tool manually marks out the point intersected on this n vertical line with sea horizon, it is hereby achieved that n sea horizon key point coordinate,
I.e.
Further, in the step (2), the structure of critical point detection network is mainly by 3 continuous 3 × 3 convolutional layers
(block1), 3 continuous are decomposed convolution module (block2), 3 continuous expansions convolution module (block3), 13 × 3
Convolutional layer (S1), 18 × 8 convolutional layer (S2) and 1 full articulamentum constitute (S3).Wherein, convolution module and expansion volume are decomposed
Volume module be all based on residual error module, the difference is that: decompose convolution by residual error module 3 × 3 convolutional layers replace
At 3 × 1 convolutional layers and 1 × 3 convolutional layer, and expanding convolution module then is by 3 × 3 convolution kernels in residual error module
It has been substituted for one 3 × 3 expansion convolution kernel.Also with Multiscale Fusion method, by block3 module, S1 layers, S2 layers
Output characteristic pattern inputs to full articulamentum after being merged.The input of critical point detection network is sea image, and what is exported is
Sea horizon key point coordinate.
Further, in the step (5), it is assumed that mixed model is uniformly distributed by 3 Gaussian Profiles and 1 and is formed, wherein 3
A Gaussian Profile is respectively used to description description sky, haze/seashore Mixed Zone, seawater region, and is uniformly distributed for describing
Potential barrier region (singular value region).Then, in image ith pixel feature vector yiProbability can indicate
Are as follows:
In above formula, N (| m, C) indicates that mean value is m and gauss of distribution function that covariance is C, and U ()=ε indicates uniform
Distribution function (wherein, ε is a very small positive value hyper parameter)；yiIndicate the feature vector of ith pixel in image (also referred to as
Observe data), mainly it is made of the color characteristic of pixel (c1, c2, c3) and coordinate (r, c)；θ indicates all Gausses point in model
Parameter (the i.e. θ={ m of clothk,Ck}K=1,2,3)；π indicates that the category prior of all pixels in image is distributed (i.e. π={ πi}I=1:M,
In, M is the number of pixel in image), πiIndicate that the category prior of ith pixel is distributed (i.e. πi=[πi1..., πik..., πi4],
Wherein, πik=p (xi=k) indicate ith pixel classification xiProbability when for k, it is assumed that k=1 indicates sky classification, k=2 table
Show intermediate seashore/haze hybrid category, k=3 indicates seawater classification, and k=4 indicates barrier classification)；H expression is examined by key point
The sea horizon that neural network forecast obtains is surveyed, andWhen indicating known to sea horizon, the condition prior distribution of ith pixel classification
It, can be by introducing Ma Erke in order to overcome local noise adverse effect caused by image segmentation in the image of sea
Husband's random field (Markov Random Field, MRF) to carry out space constraint to mixed model, i.e. all pictures in hypothesis image
The category prior of element is distributed π={ πi}I=1:MAnd Posterior distrbutionp P={ pi}I=1:MIt is a MRF about neighborhood system.According to
The joint probability distribution of Besag method, prior distribution π can be approximated to be:
In above formula, NiFor the neighborhood of pixel i,For neighborhood NiCategory prior distribution:
Wherein, λijFor fixed positive value weight, when pixel j and range pixel i are smaller, λijIt is bigger, and ∑jλij=1.
In addition, the potential-energy function in MRF is (i.e.) can be with is defined as:
In above formula,For KL divergence item, H (πi) it is entropy item.
And Posterior distrbutionp P={ pi}I=1:MJoint probability distribution are as follows:
Wherein, the Posterior distrbutionp p of pixel ii={ pik}K=1:4Calculation formula it is as follows:
Simultaneous formula (1), (2), (4) and (5), the available semantic segmentation model based on the uniform mixed distribution of gaussian sum
Joint probability density function:
In above formula, due toWithIn there are coupled relations, therefore, it is difficult to directly carry out model parameter to it
Estimation.In order to solve this problem, secondary category prior distribution collection s={ s can be introducedi}I=1:MWith auxiliary Posterior distrbutionp collection q=
{qi}I=1:MInto above formula, and peer-to-peer both sides are derived from right logarithm operation simultaneously, to obtain punishing for space constraint mixed model
Penalize log-likelihood function:
In above formula, ο indicates Hadamard product operation；And work as si≡πiAnd qi≡piWhen, formula can be reduced to
(9).In addition, according to maximum a posteriori criterion above-mentioned formula can be maximized by EM algorithm, to realize to the excellent of mixed model
Change.
The initialization of mixed model parameter:
In fact, when known to sea horizon h, the pixel of sea horizon h following region can only be seawater or barrier classification, because
This is when pixel i is located at sea horizon h following region, it is assumed that condition prior probability p (xi=k | h) are as follows:
And when pixel i is located at sea horizon h area above, it is assumed that condition prior probability p (xi=k | h) are as follows:
In addition, the initial Gaussian parameter θ in mixed model is determined see also the position of sea horizon h.Specifically, for sea
Initial Gaussian parameter { the m of water classification3,C3, can by count sea horizon h following region all pixels characteristic mean and
Covariance matrix obtains；For the initial Gaussian parameter { m of sky classification and intermediate seashore/haze hybrid categoryk,Ck}K=1,2, can
Sea horizon h or more is first divided into two regions up and down, thus pixel characteristic mean value and association side using the two regions
Initial Gaussian parameter of the poor matrix respectively as sky classification and intermediate seashore/haze hybrid category.For initial category priori
It is distributed πi, initialize formula are as follows:
Wherein, ε is a very small positive value hyper parameter.
Further, in the step (6), the specific steps of expectation-maximization algorithm (EM) are as follows:
In E-step:
1. by θ, π and p (xi=k | formula (6) h) are substituted into, calculate the Posterior distrbutionp P={ p of all super-pixeli}I=1:M。
2. (12) according to the following formula calculate secondary category prior distribution collection s={ si}I=1:M；
In above formula, ο indicates Hadamard product operation, and * indicates convolution algorithm,For normaliztion constant.
3. calculating auxiliary Posterior distrbutionp collection q={ qi}I=1:M, calculation formula is as follows:
In above formula,For normaliztion constant.
In M-step:
4. update condition category prior distribution collection, calculation formula are as follows:
5. updating Parameters of Normal Distribution collection, calculation formula is as follows:
6. judging whether EM algorithm reaches stopping criterion for iteration；If reached, stop iteration, otherwise, then continue 1.~
⑥.Wherein, stopping criterion for iteration is as follows:
According to the Posterior distrbutionp collection q={ q after EM algorithm optimizationi}I=1:M, and seawater region detection is obtained using formula (18)
As a result.
Compared with prior art, the invention has the benefit that
Firstly, the present invention predicts sea horizon key point using critical point detection network, and as reference pair
The parameter of the mixed model of space constraint is initialized, and is largely avoided mixed model and is fallen into EM algorithm optimization
Locally optimal solution, to improve the accuracy of seawater region detection.Secondly, seawater method for detecting area designed by the present invention
It is simple with structure, the characteristics of fast speed, it is easy to Practical Project deployment.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention.
Fig. 2 is critical point detection network structure of the present invention.
Fig. 3 is critical point detection network configuration figure of the present invention.
Fig. 4 is that convolution module (Factorized Block) signal is decomposed used in critical point detection network of the present invention
Figure.
Fig. 5 is that convolution module (Dilated Block) schematic diagram is expanded used in critical point detection network of the present invention.
Fig. 6 is the schematic diagram of the method for the present invention one embodiment, wherein (a) is embodiment mapping to be checked；It (b) is sea horizon
Critical point detection result figure；It (c) is embodiment semantic segmentation result figure；It (d) is seawater region testing result figure.
Specific embodiment
It is clear to be more clear the object, technical solutions and advantages of the present invention, with reference to the accompanying drawing, to tool of the invention
Body embodiment elaborates.Method or step involved in following embodiment is then unless otherwise instructed this technology neck
The conventional method or step in domain, those skilled in the art can make conventional selection or adaptability tune according to concrete application scene
It is whole.Following embodiment is realized using python programming language, pytorch frame.
As shown in Figure 1, the seawater method for detecting area based on critical point detection network and space constraint mixed model, specifically
Realize that steps are as follows:
(1) sea image pattern is acquired, sea horizon key point training sample is constructed；
3000 sea color images are acquired using the camera that unmanned water surface ship carries, using annotation tool to collected
Image is manually marked, and is formed sea horizon key point training set (totally 2000 samples), verifying collection (totally 500 samples) and is surveyed
Examination collection (totally 500 samples).Wherein, each sample is by sea image IiO is constituted with corresponding sea horizon key pointi, i.e. ti=
{Ii,Oi}.For sea horizon key point OiMark, detailed process are as follows: zoom to the resolution ratio of collected sea image
512 × 512, equidistantly mark 17 vertical lines along the vertical direction of image, then manually marked out using annotation tool this 17
The point intersected on vertical line with sea horizon, it is hereby achieved that 17 sea horizon key point coordinates.In order to reach algorithm in real time,
The resolution ratio of every image in training sample is zoomed to 64 × 64, at the same by sea horizon key point coordinate by adjust in proportion to
In 64 × 64 ranges.
(2) sea horizon key point training sample training critical point detection network is utilized；
As shown in Figures 2 and 3, the structure of critical point detection network mainly by 3 continuous 3 × 3 convolutional layers (block1),
3 continuous decompose convolution module (block2), 3 continuous expansion convolution modules (block3, expansion factor is respectively 2,3,
5), 13 × 3 convolutional layer (S1), 18 × 8 convolutional layer (S2) and 1 full articulamentum constitute (S3).In addition, using multiple dimensioned
Fusion method inputs to full articulamentum after being merged block3 module, S1 layers, S2 layers of output characteristic pattern.Wherein, it decomposes
Convolution module and expansion convolution module be all based on residual error module, the difference is that: decompose convolution will be in residual error module
3 × 3 convolutional layers be substituted for 3 × 1 convolutional layers and 1 × 3 convolutional layer, as shown in Figure 4；And expand convolution module then
It is the expansion convolution kernel that 3 × 3 convolution kernels in residual error module have been substituted for one 3 × 3, as shown in Figure 5.Critical point detection net
The input of network is sea image, and what is exported is sea horizon key point coordinate.Loss function uses mean square error function (Mean-
Square Error, MSE).
(3) it using sea image to be detected as input, is closed by the critical point detection neural network forecast sea horizon trained
Key point；
It is as shown in Figure 6 a the present embodiment sea image to be detected, picture used mainly includes seashore, obstruction buoy, sea
Unrestrained clutter etc..Fig. 6 b is the prediction result of the critical point detection network of the present embodiment.
(4) sea horizon key point is connected, corresponding sea horizon is obtained；
According to the coordinate of sea horizon key point, these key points are connected from beginning to end from left to right using straight line, to obtain
Corresponding sea horizon h.
(5) mixed model for establishing space constraint carries out just by reference pair model parameter and prior distribution of sea horizon
Beginningization；
It is formed assuming that mixed model is uniformly distributed by 3 Gaussian Profiles and 1, wherein 3 Gaussian Profiles are respectively used to retouch
Description sky, haze/seashore Mixed Zone, seawater region are stated, and is uniformly distributed (unusual for describing potential barrier region
It is worth region).In addition, by introducing Markov random field (Markov Random Field, MRF) to carry out mixed model
Space constraint, i.e., the category prior of all pixels is distributed π={ π in hypothesis imagei}I=1:MAnd Posterior distrbutionp P={ pi}I=1:M
It is a MRF about neighborhood system.By derivation, the penalized log-likelihood function of space constraint mixed model is finally obtained:
The initialization of mixed model:
Sea horizon position h can be determined by step (3) (4), and the pixel of sea horizon h following region can only be seawater
Or barrier classification, therefore when pixel i is located at sea horizon h following region, it is assumed that condition prior probability p (xi=k | h) are as follows:
And when pixel i is located at sea horizon h area above, it is assumed that condition prior probability p (xi=k | h) are as follows:
In addition, the initial Gaussian parameter θ in mixed model is determined see also the position of sea horizon h.Specifically, for sea
Initial Gaussian parameter { the m of water classification3, C3, can by count sea horizon h following region all pixels characteristic mean and
Covariance matrix obtains；For the initial Gaussian parameter { m of sky classification and intermediate seashore/haze hybrid categoryk,Ck}K=1,2, can
Sea horizon h or more is first divided into two regions up and down, thus pixel characteristic mean value and association side using the two regions
Initial Gaussian parameter of the poor matrix respectively as sky classification and intermediate seashore/haze hybrid category.For initial category priori
It is distributed πi, initialize formula are as follows:
Wherein, ε is a very small positive value hyper parameter, in the present embodiment, ε=1 × 10-15。
In the present embodiment, uniformly distributed function U ()=ε=1 × 10-15。
(6) semantic segmentation is carried out to sea image by expectation-maximization algorithm (EM), to detect seawater region.
Space constraint mixed model is after EM algorithm optimization, available Posterior distrbutionp collection q={ qi}I=1:M, utilize formulaObtain embodiment semantic segmentation result as fig. 6 c.Fig. 6 d is the inspection of the present embodiment seawater region
Survey result.
Claims (4)
1. a kind of seawater method for detecting area based on critical point detection network and space constraint mixed model, which is characterized in that
The following steps are included:
(1) sea image pattern is acquired, sea horizon key point training sample is constructed；
(2) sea horizon key point training sample training critical point detection network is utilized；
(3) using sea image to be detected as input, pass through the critical point detection neural network forecast sea horizon key point trained；
(4) sea horizon key point is connected, corresponding sea horizon is obtained；
(5) mixed model for establishing space constraint is initialized by reference pair model parameter and prior distribution of sea horizon；
(6) semantic segmentation is carried out to sea image by expectation-maximization algorithm, to detect seawater region.
2. the seawater region detection side according to claim 1 based on critical point detection network and space constraint mixed model
Method, which is characterized in that in the step (1), the N sea color images of camera acquisition carried using unmanned water surface ship are used
Annotation tool manually marks acquired image, forms sea horizon key point training set T={ t1,t2,…,tN}；Its
In, each training sample tiBy sea image IiO is constituted with corresponding sea horizon key pointi, i.e. ti={ Ii,Oi}；For sea
Antenna key point OiMark, detailed process are as follows: equidistantly mark n vertical line along the vertical direction of image, then utilize mark
Note tool manually marks out the point intersected on this n vertical line with sea horizon, to obtain n sea horizon key point coordinate, i.e.,
3. the seawater region detection side according to claim 1 based on critical point detection network and space constraint mixed model
Method, which is characterized in that in the step (2), the structure of critical point detection network is mainly by 3 continuous 3 × 3 convolutional layers
Block1,3 continuous decomposition convolution module block2,3 continuous expansion convolution module block3,13 × 3 convolutional layer
S1,18 × 8 convolutional layer S2 and 1 full articulamentum S3 are constituted；Wherein, decompose convolution module and expansion convolution module be all with
Based on residual error module, the difference is that: convolution, which is decomposed, by 3 × 3 convolutional layers in residual error module has been substituted for one 3 × 1
Convolutional layer and 1 × 3 convolutional layer, and expanding convolution module then is that 3 × 3 convolution kernels in residual error module have been substituted for one
3 × 3 expansion convolution kernel；Also with Multiscale Fusion method, by block3 module, S1 layers, S2 layers of output characteristic pattern into
Full articulamentum is inputed to after row fusion, the input of critical point detection network is sea image, and what is exported is sea horizon key point
Coordinate.
4. the seawater region detection side according to claim 1 based on critical point detection network and space constraint mixed model
Method, which is characterized in that in the step (5), it is assumed that mixed model is uniformly distributed by 3 Gaussian Profiles and 1 and is formed, wherein 3
A Gaussian Profile is respectively used to description sky, haze/seashore Mixed Zone, seawater region, and is uniformly distributed potential for describing
Barrier region, i.e. singular value region；Then, in image ith pixel feature vector yiProbability be expressed as:
In above formula, N (| m, C) indicates that mean value is m and gauss of distribution function that covariance is C, and U ()=ε expression is uniformly distributed
Function, wherein ε is a very small positive value hyper parameter；yiThe feature vector for indicating ith pixel in image, is also referred to as observed
Data are mainly made of the color characteristic of pixel (c1, c2, c3) and coordinate (r, c)；θ indicates all Gaussian Profiles in model
Parameter, i.e. θ={ mk,Ck}K=1,2,3；π indicates the category prior distribution of all pixels in image, i.e. π={ πi}I=1:M, wherein M
For the number of pixel in image, πiIndicate the category prior distribution of ith pixel, i.e. πi=[πi1..., πik,…,πi4], wherein
πik=p (xi=k) indicate ith pixel classification xiProbability when for k, it is assumed that k=1 indicates sky classification, in k=2 expression
Between seashore/haze hybrid category, k=3 indicate seawater classification, k=4 indicate barrier classification；H is indicated by critical point detection network
Predict obtained sea horizon, andWhen indicating known to sea horizon, the condition prior distribution of ith pixel classification p(xi=k | when h) known to sea horizon h, the classification x of pixel ii
The probability of=k；
In order to overcome local noise adverse effect caused by image segmentation in the image of sea, by introducing Markov random field
MRF carries out space constraint to mixed model, i.e., the category prior of all pixels is distributed π={ π in hypothesis imagei}I=1:MAfter and
Test distribution P={ pi}I=1:MIt is a MRF about neighborhood system；By deriving, punishing for space constraint mixed model is finally obtained
Penalize log-likelihood function:
In above formula,Expression Hadamard product operation, D (| |) indicate KL divergence, s={ si}I=1:MFor auxiliary category prior point
Cloth collection, q={ qi}I=1:MTo assist Posterior distrbutionp collection,For the neighborhood N of pixel iiCategory prior distribution,For pixel i's
Neighborhood NiPosterior distrbutionp；
The initialization of mixed model parameter:
When known to sea horizon h, the pixel of sea horizon h following region can only be seawater or barrier classification, therefore work as pixel i
When sea horizon h following region, it is assumed that condition prior probability p (xi=k | h) are as follows:
And when pixel i is located at sea horizon h area above, it is assumed that condition prior probability p (xi=k | h) are as follows:
In addition, the initial Gaussian parameter θ in mixed model is determined referring also to the position of sea horizon h；Specifically, for seawater classification
Initial Gaussian parameter { m3, C3, the characteristic mean and covariance matrix of all pixels by counting sea horizon h following region
It obtains；For the initial Gaussian parameter { m of sky classification and intermediate seashore/haze hybrid categoryk, Ck}K=1,2, first by sea horizon h
It is above to be divided into two regions up and down, thus using the two regions pixel characteristic mean value and covariance matrix respectively as
The initial Gaussian parameter of sky classification and intermediate seashore/haze hybrid category；For initial category prior distribution πi, initialization
Formula are as follows:
Wherein, ε is a very small positive value hyper parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910519569.1A CN110298271A (en) | 2019-06-17 | 2019-06-17 | Seawater method for detecting area based on critical point detection network and space constraint mixed model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910519569.1A CN110298271A (en) | 2019-06-17 | 2019-06-17 | Seawater method for detecting area based on critical point detection network and space constraint mixed model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110298271A true CN110298271A (en) | 2019-10-01 |
Family
ID=68028079
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910519569.1A Pending CN110298271A (en) | 2019-06-17 | 2019-06-17 | Seawater method for detecting area based on critical point detection network and space constraint mixed model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110298271A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111583325A (en) * | 2020-05-10 | 2020-08-25 | 上海大学 | Image processing-based method for detecting sea waves by unmanned ship |
| CN112327265A (en) * | 2020-10-23 | 2021-02-05 | 北京理工大学 | Division and treatment detection method based on semantic segmentation network |
| CN112731521A (en) * | 2019-10-14 | 2021-04-30 | 中国石油化工股份有限公司 | Marine seismic wave simulation method and system based on Gaussian random distribution |
| CN113538425A (en) * | 2021-09-16 | 2021-10-22 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Passable water area segmentation equipment, image segmentation model training and image segmentation method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102279973A (en) * | 2010-06-11 | 2011-12-14 | 中国兵器工业第二○五研究所 | Sea-sky-line detection method based on high gradient key points |
| CN107808386A (en) * | 2017-09-26 | 2018-03-16 | 上海大学 | A kind of sea horizon detection method based on image, semantic segmentation |
| CN108764027A (en) * | 2018-04-13 | 2018-11-06 | 上海大学 | A kind of sea-surface target detection method calculated based on improved RBD conspicuousnesses |
| CN109376591A (en) * | 2018-09-10 | 2019-02-22 | 武汉大学 | The ship object detection method of deep learning feature and visual signature joint training |
| CN109886336A (en) * | 2019-02-21 | 2019-06-14 | 山东超越数控电子股份有限公司 | A kind of object detection method and system based on warship basic image |
-
2019
- 2019-06-17 CN CN201910519569.1A patent/CN110298271A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102279973A (en) * | 2010-06-11 | 2011-12-14 | 中国兵器工业第二○五研究所 | Sea-sky-line detection method based on high gradient key points |
| CN107808386A (en) * | 2017-09-26 | 2018-03-16 | 上海大学 | A kind of sea horizon detection method based on image, semantic segmentation |
| CN108764027A (en) * | 2018-04-13 | 2018-11-06 | 上海大学 | A kind of sea-surface target detection method calculated based on improved RBD conspicuousnesses |
| CN109376591A (en) * | 2018-09-10 | 2019-02-22 | 武汉大学 | The ship object detection method of deep learning feature and visual signature joint training |
| CN109886336A (en) * | 2019-02-21 | 2019-06-14 | 山东超越数控电子股份有限公司 | A kind of object detection method and system based on warship basic image |
Non-Patent Citations (3)
| Title |
|---|
| MATEJ KRISTAN，ET AL: "Fast Image-Based Obstacle Detection From Unmanned Surface Vehicles", 《IEEE TRANSACTIONS ON CYBERNETICS》 * |
| 涂兵等: "多特征与边缘校正融合的天际线检测算法研究", 《计算机工程与应用》 * |
| 胡耀辉等: "基于海天线的舰船弱小目标检测", 《西北工业大学学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112731521A (en) * | 2019-10-14 | 2021-04-30 | 中国石油化工股份有限公司 | Marine seismic wave simulation method and system based on Gaussian random distribution |
| CN111583325A (en) * | 2020-05-10 | 2020-08-25 | 上海大学 | Image processing-based method for detecting sea waves by unmanned ship |
| CN112327265A (en) * | 2020-10-23 | 2021-02-05 | 北京理工大学 | Division and treatment detection method based on semantic segmentation network |
| CN113538425A (en) * | 2021-09-16 | 2021-10-22 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Passable water area segmentation equipment, image segmentation model training and image segmentation method |
| CN113538425B (en) * | 2021-09-16 | 2021-12-24 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Passable water area segmentation equipment, image segmentation model training and image segmentation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108596101B (en) | Remote sensing image multi-target detection method based on convolutional neural network | |
| CN107818326B (en) | A kind of ship detection method and system based on scene multidimensional characteristic | |
| CN109615611B (en) | Inspection image-based insulator self-explosion defect detection method | |
| CN110298271A (en) | Seawater method for detecting area based on critical point detection network and space constraint mixed model | |
| Xia et al. | A fast edge extraction method for mobile LiDAR point clouds | |
| CN108805906A (en) | A kind of moving obstacle detection and localization method based on depth map | |
| CN109255317B (en) | Aerial image difference detection method based on double networks | |
| Schoenberg et al. | Segmentation of dense range information in complex urban scenes | |
| CN111429514A (en) | Laser radar 3D real-time target detection method fusing multi-frame time sequence point clouds | |
| CN102982555B (en) | Guidance Tracking Method of IR Small Target based on self adaptation manifold particle filter | |
| CN110287837A (en) | Sea obstacle detection method based on prior estimate network and space constraint mixed model | |
| CN106960442A (en) | Based on the infrared night robot vision wide view-field three-D construction method of monocular | |
| CN110009010A (en) | Wide area optical remote sensing target detection method based on the re-detection of interest region | |
| CN109919026A (en) | A kind of unmanned surface vehicle local paths planning method | |
| CN112017243A (en) | Medium visibility identification method | |
| Shi et al. | Obstacle type recognition in visual images via dilated convolutional neural network for unmanned surface vehicles | |
| Li et al. | CSF-Net: Color spectrum fusion network for semantic labeling of airborne laser scanning point cloud | |
| CN103688289A (en) | Method and system for estimating a similarity between two binary images | |
| CN116978009A (en) | Dynamic object filtering method based on 4D millimeter wave radar | |
| CN115187959B (en) | Method and system for landing flying vehicle in mountainous region based on binocular vision | |
| CN113920254B (en) | Monocular RGB (Red Green blue) -based indoor three-dimensional reconstruction method and system thereof | |
| CN103236053B (en) | A kind of MOF method of moving object detection under mobile platform | |
| CN112069997B (en) | Unmanned aerial vehicle autonomous landing target extraction method and device based on DenseHR-Net | |
| CN104851090A (en) | Image change detection method and image change detection device | |
| CN114066795A (en) | DF-SAS high-low frequency sonar image fine registration fusion method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191001 |