CN103714557A - Automatic ground-based cloud detection method - Google Patents

Abstract

The invention discloses an automatic ground-based cloud detection method. The method includes the steps that visible light could pictures are collected; grayscale adjustment at a nonlinearity dynamic range is conducted on blue wave band images and red wave band images of the cloud pictures to obtain enhanced images of two wave bands, and then difference value processing is carried out on the enhanced images of the two wave bands to obtain a feature map; Shearlet conversion is performed on the feature map to acquire multiscale sub-band coefficients in different directions; afterwards, the texture features of modulus value images of all the sub-band coefficients are extracted; at last, a clustering algorithm is used for classifying extracted feature vectors, so that automatic detection of the cloud pictures is achieved. By means of the automatic ground-based cloud detection method, the limitation problem of manual visual judgment can be solved, automatic detection of the visible light cloud pictures is achieved, and good robustness and precision are achieved.

Description

A kind of ground Automatic cloud detection method

Technical field

The invention belongs to technical field of image processing, especially a kind of ground Automatic cloud detection method.

Background technology

Cloud is as an important step of hydrologic cycle on the earth, and itself and terrestrial radiation interaction joint effect the energy equilibrium of local and global range.Therefore, obtain exactly the information of cloud, the climate change in the accuracy of weather forecast, global range and flight support etc. are all of great significance.The at present detection of cloud is mainly by manually judging, and this method is time-consuming, require great effort and with very strong subjective limitation.Simultaneously due to affected by artificial subjective factor and night illumination condition restriction, artificial clouds discharge observation has larger subjective error, and manpower consumption is very large, make troubles to the quantification application of the observational data of cloud, the automatic detection that therefore realizes cloud amount be current in the urgent need to.

Current ground cloud detection algorithm is mainly to take threshold value as basis, with the gray-scale value (or radiance) of red blue wave band to being compared to the basis for estimation of cloud and clear sky.Long etc. have proposed to carry out with fixed threshold the detection of cloud, and this method is better to spissatus detection effect under clear sky, but consider the complicacy of sky cloud atlas, the testing result that fixed threshold can not make all images obtain.The application maximum between-cluster variance threshold method that the people such as Yang Jun proposes afterwards can calculate threshold value to different cloud atlas self-adaptations, but because the form of cloud is ever-changing and the impact of illumination, to view picture cloud atlas, adopts a global threshold can not obtain good accuracy of detection.

Summary of the invention

In order to solve the problem of ground Automatic cloud detection, the object of the invention is to make full use of the textural characteristics of ground visible cloud image, realize the automatic detection of cloud, there is very strong robustness and correctness.

The technical matters that patent of the present invention solves can adopt following technical solution to realize:

A ground Automatic cloud detection method, is characterized in that comprising the following steps:

Step 1: utilize imaging device to visible cloud image collection;

Step 2: the cloud atlas that input is collected carries out the separation of each Color Channel, the gray scale adjustment of again blue red two band images of cloud atlas being carried out to Nonlinear Dynamic scope obtain the blue wave band of cloud atlas strengthen image B ' with red wave band enhancing image R ', then the pixel that two wave bands is strengthened to the differential chart (B '-R ') of image normalizes to [0255], obtains characteristic image;

Step 3: characteristic image carries out multiple dimensioned multidirectional Shearlet and decomposes;

Step 4: on the basis of step 3, each yardstick different directions sub-band coefficients is carried out to non-local mean filtering processing, reduce the conversion of identical texture region feature, increase the difference in different texture region simultaneously.Then to filtered sub-band coefficients delivery value, calculate the d direction patrix value image I of l layer ^ldlocal energy under (2n+1) * (2n+1) size windows

with local energy variance

$e_l^d(x,y)=\frac{1}{{(2n+1)}^2}\underset{x=p-n}{\overset{p+n}{\mathrm{\Σ}}}\underset{y=q-n}{\overset{q+n}{\mathrm{\Σ}}}{\left(I^{\mathrm{ld}}(x,y)\right)}^{2}l=\mathrm{1,2},...,\mathrm{Ld}=\mathrm{1,2},...,D$

${\mathrm{\δ}}_l^d(p,q)=\frac{1}{{(2n+1)}^2}\underset{x=p-n}{\overset{p+n}{\mathrm{\Σ}}}\underset{y=q-n}{\overset{q+n}{\mathrm{\Σ}}}{[e_l^d(x,y)-u_l]}^{2}l=\mathrm{1,2},...,\mathrm{Ld}=\mathrm{1,2},...,D$

Wherein x, y represent the position in each subband mould value image, u _lin moving window

average,

for the variance of window self-energy is as proper vector to be sorted, in cloud atlas F, the final proper vector of pixel (x, y) is expressed as: $({\mathrm{\δ}}_1^1(p,q),{\mathrm{\δ}}_1^2(p,q),...,{\mathrm{\δ}}_1^D(p,q),...,{\mathrm{\δ}}_L^1(p,q),{\mathrm{\δ}}_L^2(p,q),...,{\mathrm{\δ}}_L^D(p,q)).$

Step 5: the proper vector of utilizing Fuzzy C-Means Cluster Algorithm to extract step 4 is classified, and finally realizes the automatic detection of cloud.

The invention has the beneficial effects as follows: the present invention uses the Shearlet conversion extraction different scale of cloud atlas, the textural characteristics of direction, and utilizes Fuzzy C-Means Cluster Algorithm to textural characteristics Classification of Matrix, finally reaches the automatic detection of ground cloud.The present invention has weakened the impact of solar irradiation on ground cloud, makes full use of color characteristic and the textural characteristics of cloud atlas simultaneously, has very strong robustness and higher accuracy; The present invention is simple in structure, utilizes existing graph capture device and common computer to realize, and has improved practicality and applicability.

Accompanying drawing explanation

Fig. 1 is a kind of overview flow chart of ground Automatic cloud detection method

Fig. 2 divides detection design sketch according to embodiments of the invention on cloud atlas data set top

Embodiment

As shown in Figure 1, specific embodiment of the invention method comprises following concrete steps:

1) utilize imaging device to gather sky cloud atlas picture.

2) the RGB cloud atlas of input is decomposed into the image of R, G, tri-Color Channels of B, the gray scale adjustment of more respectively image of R Color Channel and B Color Channel being carried out to Nonlinear Dynamic scope obtain respectively the enhancing image B of enhancing image R ' He the B Color Channel (blue wave band) of R Color Channel (red wave band) '.

3) the blue red band image of the cloud atlas strengthening is carried out to difference processing, obtain (B '-R ') image of single channel.Again the pixel of single channel (B '-R ') image is normalized to [0255], obtain characteristic image.

4) utilize Shearlet transfer pair characteristic image to carry out multiple dimensioned decomposition, then utilize shearing matrix to carry out shearing manipulation to the multi-scale image obtaining, travel direction decomposes, and obtains the sub-band coefficients of different scale, different directions.

5) each sub-band coefficients obtaining is adopted to non-local mean filtering, reduce the conversion of feature in identical texture region, increase the difference of zones of different.Each yardstick different directions sub-band coefficients I={D (x) | the filtered estimated value of x ∈ I} is NL[D (x)]:

$\mathrm{NL}\left[D\left(x\right)\right]=\underset{y\∈l}{\mathrm{\Σ}}w(x,y)D\left(y\right)$

W (x wherein, y) represent the similarity degree of pixel x and pixel y, G (D (x)) and G (D (y)) represent respectively the neighborhood window centered by x and y two pixels, the similarity of G (D (x)) and G (D (y)) determines by the Euclidean distance between them, and weights are defined as:

$w(x,y)=\frac{1}{Z\left(x\right)}\exp \left(\frac{{\left|\right|G\left(D\left(x\right)\right)-G\left(D\left(y\right)\right)\left|\right|}^2}{h^2}\right)$

The rate of decay of parameter h control characteristic function, weights satisfy condition 0≤w (x, y)≤1 and ∑ w (x, y)=1, Z (x) is weights normalized factor, its computing formula is:

$Z\left(x\right)=\underset{y\∈I}{\mathrm{\Σ}}\exp (-\frac{{\left|\right|G(D\left(x\right)-G\left(D\left(y\right)\right))\left|\right|}^2}{h^2})$

6) the sub-band coefficients delivery value after after filtering to each yardstick different directions, then calculates the d direction patrix value image I of l layer ^ldlocal energy under (2n+1) * (2n+1) size windows with local energy variance

$e_l^d(x,y)=\frac{1}{{(2n+1)}^2}\underset{x=p-n}{\overset{p+n}{\mathrm{\Σ}}}\underset{y=q-n}{\overset{q+n}{\mathrm{\Σ}}}{\left(I^{\mathrm{ld}}(x,y)\right)}^{2}l=\mathrm{1,2},...,\mathrm{Ld}=\mathrm{1,2},...,D$

${\mathrm{\δ}}_l^d(p,q)=\frac{1}{{(2n+1)}^2}\underset{x=p-n}{\overset{p+n}{\mathrm{\Σ}}}\underset{y=q-n}{\overset{q+n}{\mathrm{\Σ}}}{[e_l^d(x,y)-u_l]}^{2}l=\mathrm{1,2},...,\mathrm{Ld}=\mathrm{1,2},...,D$

Wherein x, y represent the position in each subband mould value image, u _lin moving window

average,

for the variance of window self-energy is as proper vector to be sorted, in cloud atlas F, the final proper vector of pixel (x, y) is expressed as: $({\mathrm{\δ}}_1^1(p,q),{\mathrm{\δ}}_1^2(p,q),...,{\mathrm{\δ}}_1^D(p,q),...,{\mathrm{\δ}}_L^1(p,q),{\mathrm{\δ}}_L^2(p,q),...,{\mathrm{\δ}}_L^D(p,q)).$

7) utilize Fuzzy C-Means Cluster Algorithm, the proper vector that the energy variance of each pixel in cloud atlas region is formed, classifies, and detailed process is as follows:

Proper vector in A definition in step is X={x ₁, x ₂..., x _n, the given classification center of proper vector initialization is counted C (2≤C≤n), exponential factor m, and iteration is by error ε, algorithm maximum iteration time T _max, u _tjrepresent that j sample belongs to the degree of membership at i center, degree of membership matrix U=[u _tj], x _jrepresent j sample, definition V={v ₁, v ₂..., v _cthe set on vector X, represent C cluster centre vector.

B is to t=1, and 2 ..., T _maxcarry out iterative computation, calculate V _t=[v _{1, t}v _{2, t}, v _{c, t}], wherein

$v_{1,t}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\left(u_{\mathrm{tj},t-1}\right)}^m{x}_j/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\left(u_{\mathrm{tj},t-1}\right)}^m,1\≤i\≤C$

Calculate U _t=[u _{tj, t}] _{c * n}, v wherein _{tj, t}deterministic process be: make d _{tj, t}=|| x _j-v _{1, t}|| ²if, d _{tj, t}=0, u _{tj, t}=1, and to k ≠ i, u _{kj, t}=0; If d _{tj, t}> 0, $u_{\mathrm{tj},t}=\frac{1}{\underset{k=1}{\overset{C}{\mathrm{\Σ}}}{(d_{\mathrm{tj},t}/d_{\mathrm{kj},t})}^{1/(m-1)}}1\≤i\≤C,1\≤j\≤n$

If C || U ^t-U ^t-1|| < ε, termination of iterations is exported picture, otherwise next t is calculated.Fig. 2 is that the part on cloud atlas data set detects design sketch according to embodiments of the invention.

Claims (3)

1. a ground Automatic cloud detection method, is characterized in that, the method comprises the following steps:

Step 1: utilize imaging device to visible cloud image collection;

Step 2: the cloud atlas that input is collected carries out the separation of each Color Channel, the blue wave band that cloud atlas is obtained in the gray scale adjustment that the redder two wave band cloud atlas of indigo plant is looked like to carry out Nonlinear Dynamic scope strengthen image B ' and red wave band strengthen image R ', then the differential chart (B '-R ') that two wave bands is strengthened to image normalizes to [0255], obtains characteristic image;

Step 3: characteristic image carries out multiple dimensioned multidirectional Shearlet and decomposes;

Step 4: to each yardstick different directions sub-band coefficients delivery value, and the mould value image of each subband is carried out to feature extraction on the basis of step 3;

Step 5: the proper vector of utilizing clustering algorithm to extract step 4 is classified, and finally realizes the automatic detection of visible Shekinah.

2. a kind of ground Automatic cloud detection method according to claim 1, is characterized in that, the concrete computation process described in step 4 is as follows:

Step 4.1 is usingd the differential chart that blue red two wave bands of cloud atlas strengthen images and is carried out Shearlet conversion as characteristic pattern, obtains a series of subbands;

The sub-band coefficients of each yardstick different directions of step 4.2 is carried out non-local mean filtering, reduces the conversion of the feature of identical texture region, increases the difference of zones of different simultaneously;

Filtered each sub-band coefficients delivery value of step 4.3, the mould value image I in the d direction of calculating l layer _ldlocal energy under (2n+1) * (2n+1) size windows

with local energy variance

$e_l^d(x,y)=\frac{1}{{(2n+1)}^2}\underset{x=p-n}{\overset{p+n}{\mathrm{\&Sigma;}}}\underset{y=q-n}{\overset{q+n}{\mathrm{\&Sigma;}}}{\left(I^{\mathrm{ld}}(x,y)\right)}^{2}l=\mathrm{1,2},...,\mathrm{Ld}=\mathrm{1,2},...,D$

${\mathrm{\&delta;}}_l^d(p,q)=\frac{1}{{(2n+1)}^2}\underset{x=p-n}{\overset{p+n}{\mathrm{\&Sigma;}}}\underset{y=q-n}{\overset{q+n}{\mathrm{\&Sigma;}}}{[e_l^d(x,y)-u_l]}^{2}l=\mathrm{1,2},...,\mathrm{Ld}=\mathrm{1,2},...,D$

Wherein x, y represent the position in each subband mould value image, u _lin moving window

average,

for the variance of window self-energy is as proper vector to be sorted, in cloud atlas F, the final proper vector of pixel (x, y) is expressed as: $({\mathrm{\&delta;}}_1^1(p,q),{\mathrm{\&delta;}}_1^2(p,q),...,{\mathrm{\&delta;}}_1^D(p,q),...,{\mathrm{\&delta;}}_L^1(p,q),{\mathrm{\&delta;}}_L^2(p,q),...,{\mathrm{\&delta;}}_L^D(p,q)).$

3. a kind of ground Automatic cloud detection method according to claim 1, is characterized in that, the sorting algorithm described in step 4 is: Fuzzy C-Means Cluster Algorithm.

Images