CN104700115A - Detection method for meteor crater in soft landing process of martian probe based on sparsely promoted and integrated classifier - Google Patents

Abstract

The invention provides a detection method for a meteor crater in the soft landing process of a martian probe based on a sparsely promoted and integrated classifier. The detection method comprises four steps: step 1, determining a candidate meteor crater; step 2, carrying out feature extraction on the candidate meteor crater; step 3, carrying out feature selection on extracted textural features; step 4, designing the sparsely promoted and integrated classifier by combining a Boost algorithm and a sparse kernel density estimation algorithm RSDE-WLL (Research Software Design Engineer-Wireless Local Loop), so as to realize quick detection on the meteor crater based on an image. According to the method provided by the invention, the classification on the meteor crater and a non-meteor crater can be realized by using the advantages of the sparsely promoted and integrated classifier designed including sparse solution and computation complexity reduction after the feature extraction and selection are carried out on the textural features of the meteor crater based on the image, so as to quickly detect the meteor crater; furthermore, the classification accuracy can reach up to 85% or above, and the certain reference value can be realized for the detection of the meteor crater in the soft landing process of the actual martian probe.

Description

Based on the detection method of meteorite crater in the Mars probes soft landing process of sparse lifting integrated classifier

Technical field

The present invention relates to a kind of detection method based on meteorite crater in the Mars probes soft landing process of sparse lifting integrated (SparseBoost) sorter, be specifically related to the design of the pre-service of meteorite crater in martian surface image, Texture Feature Extraction and selection and SparseBoost sorter, belong to image procossing and area of pattern recognition.

Background technology

In Mars probes soft landing process, detection of obstacles research fundamental purpose is detect barrier and determine safe landing locations accordingly.Wherein meteorite crater is the modal geological form in celestial body surface such as Mars, has the features such as distribution is wide, area is large, characteristics of image is obvious, is one of main barrier needing in detector soft landing process to detect.

In the image of Mars probes collection, meteorite crater can regard a pair crescent moon with bright and dark area as.In the candidate's meteorite crater determined, usually after the textural characteristics of candidate's meteorite crater being extracted, the characteristic extracted has higher dimension, must be classified to the meteorite crater in candidate's meteorite crater and non-meteorite crater by suitable image feature selection algorithm, supervised learning sorting algorithm, to determine the particular location of meteorite crater in image.

SparseBoost sorting algorithm is a kind of sparse supervised learning sorting algorithm, advantage compared with existing AdaBoost, Boost algorithm is, in each iterative process, AdaBoost algorithm utilizes whole feature set, and SparseBoost algorithm only selects an optimum feature; Simultaneously in the process building Weak Classifier, the equal trade-off decision tree of AdaBoost and Boost algorithm builds sorter, utilize whole sample set in computation process, and SparseBoost algorithm adopts sparse Density Estimator RSDE-WL1 to build Weak Classifier, only utilizes a small amount of sample to realize.Therefore SparseBoost algorithm has advantage that is openness and reduction computation complexity, in real process, SparseBoost sorting algorithm is applied to the detection of meteorite crater in Mars probes soft landing process, effectively can reduce computation complexity, realize the quick detection of meteorite crater.

Summary of the invention

1, goal of the invention

The object of this invention is to provide a kind of detection method based on meteorite crater in the Mars probes soft landing process of sparse lifting integrated classifier, first the method carries out pre-service to determine candidate's meteorite crater to the image gathered, secondly abstract image textural characteristics from candidate's meteorite crater, and carry out feature selecting; Then the sparse Density Estimator algorithm (RSDE-WL1) of improving is combined with lifting Ensemble Learning Algorithms (Boost), build sparse lifting integrated classifier (SparseBoost sorter), finally SparseBoost sorter is applied to the detection of meteorite crater, realize detecting fast, obtain higher classify accuracy simultaneously.

2, technical scheme

For achieving the above object, the step in the main frame figure (Fig. 1) that the present invention detects automatically according to meteorite crater, specifically introduces the technical scheme of the method.

The present invention devises a kind of detection method based on meteorite crater in the Mars probes soft landing process of sparse lifting integrated classifier, and the method comprises the following steps:

Step 1. determines candidate's meteorite crater

Determine that the key of candidate's meteorite crater is that the meteorite crater in image can regard a pair crescent moon with bright and dark area as, as shown in Figure 2.The shape of often pair of crescent moon can be determined from image by the shape detecting method based on mathematical morphology, the crescent moon that can mate meteorite crater alternatively.As shown in Figure 3, input is a full-colour image to the building process of candidate's meteorite crater, it comprises a lot of bright and dark features region.Bright and dark shape parallel processing, by using original image process to become clear shape, and uses inverted image to process dark shape.The target of the method eliminates all noise characteristics that cannot be designated as meteorite crater, and only retain bright and dark features.Remaining bright and dark features region matches each other, and these area marking are good, as the candidate region of meteorite crater.

Step 2. candidate meteorite crater Texture Feature Extraction

In order to represent single candidate's meteorite crater from rectangular characteristic aspect, first we extract the square image block around candidate's meteorite crater.In an experiment, in order to comprise the edge of meteorite crater around, we use the twice of candidate's meteorite crater size as masked areas.The textural characteristics of each candidate's meteorite crater the unknown uses the square covering of 9 kinds of different sizes to encode, as shown in Figure 4.Therefore, the attribute of candidate's meteorite crater that image comprises can be described by thousands of textural characteristics.These features are not separate, and these complete features compensate for by the single square restriction hiding the texture information obtained.

Step 3. carries out feature selecting to the textural characteristics extracted

According to the preliminary candidate's meteorite crater textural characteristics extracted, due to the higher-dimension of characteristic, therefore by before training sample and test sample book input sorter, feature selecting must be carried out.In the present invention, the SparseBoost algorithm that we utilize step 4 to design carries out feature selecting, and the maximum difference of itself and AdaBoost algorithm is, the former only chooses an optimum feature in iterative process each time, and the latter utilizes whole feature set usually.Greatly reduce the intrinsic dimensionality of training sample like this, effectively reduce the computation complexity of sorter training.

Boost algorithm is combined with sparse Density Estimator algorithm RSDE-WL1 by step 4., devises sparse lifting integrated classifier, to realize the quick detection to the meteorite crater based on image.

According to selected candidate's meteorite crater textural characteristics, in order to distinguish wherein meteorite crater and non-meteorite crater, the present invention devises a kind of supervised learning sorting algorithm---SparseBoost algorithm.The method is in conjunction with the sparse Density Estimator algorithm (RSDE-WL1) of Boost algorithm and a kind of improvement, while selection character subset, construct the design of some sparse Density Estimator devices for corresponding base sorter, by the weighted array of base sorter, finally realize integrated classifier.

Given n candidate's meteorite crater (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n), wherein y _i=0,1, i=1,2 ..., n correspond to non-meteorite crater (c respectively ₀) and meteorite crater (c ₁) example, n ₀and n ₁correspond to the number of non-meteorite crater and meteorite crater example respectively, n ₀+ n ₁=n.Each candidate's meteorite crater can be expressed as a proper vector x=(f ₁, f ₂..., f _m) ^t, wherein each feature f _i, i=1 ..., m is produced by the square covering of ad-hoc location a certain on candidate's meteorite crater, and m is the feature sum extracted.SparseBoost algorithm (detailed process is as algorithm 1) is utilized to build a series of Weak Classifier h _t(x), and by weighting integrated approach Weak Classifier carried out combining and set up final strong classifier H (x):

$H\left(x\right)=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t{h}_t\left(x\right)---\left(1\right)$

Wherein T is iterations (T < n), α _tthe Weak Classifier h of study _tthe weight of (x).

In each iterative process, need the step realizing following three cores: Weak Classifier study, optimal feature selection and next iteration process sample weights upgrade.Wherein, in Weak Classifier learning process, penalty term is added, the sparse Density Estimator algorithm RSDE-WL1 be improved to compression collection density Estimation algorithm (RSDE).Utilize RSDE-WL1 algorithm to estimate its probability density function to each category attribute, classify according to the sample of Bayes decision rule to input, obtain Weak Classifier.

(1) Weak Classifier study

In the t time iterative process, for the single optimal characteristics f ∈ { f chosen ₁, f ₂..., f _mthe Weak Classifier h that builds _tx (), realizes by building Bayes classifier.Before the two classification problem of classifying about meteorite crater and non-meteorite crater is discussed, first introduce Bayes classifier.General for Bayes's classification problem, other posterior probability density of expectation estimation given input amendment x lower class.In order to obtain one about the probability classification of density Estimation, first train a Multilayer networks device for each category attribute c wherein x is the proper vector for representing single candidate's meteorite crater, and β is core weight vector, and c is the category attribute of candidate's meteorite crater, c ∈ { c ₀, c ₁, c ₀represent non-meteorite crater classification, c ₁represent meteorite crater classification.Then use Bayes rule (2) to calculate posterior probability, final test sample is assigned to the category attribute with maximum a posteriori probability.

$\hat{p}\left(c\right|x;\mathrm{\&beta;})=\frac{\hat{p}(x;\mathrm{\&beta;}|c)p\left(c\right)}{{\mathrm{\&Sigma;}}_{i=0}^1\hat{p}(x;\mathrm{\&beta;}|c_i)p\left(c_i\right)}---\left(2\right)$

For two classification problem in the present invention, first estimate two conditional probability densities under given classification with these two density can be obtained by follow-up sparse Density Estimator RSDE-WL1 (calculating according to formula (5) and (6)).Then according to formula (2), corresponding posterior probability is calculated respectively with (directly calculating according to formula (2)).According to the sample size of each category attribute, calculate the prior probability p (c of two kinds ₀) and p (c ₁): p (c ₀)=n ₀/ n, p (c ₁)=n ₁/ n, p (c ₀)+p (c ₁)=1.The sample of Bayes decision rule (3) to input is finally utilized to classify

$\left.\begin{array}{ccc}\mathrm{if}& \hat{p}(x;{\mathrm{\&beta;}}_{n_1},h_1|c_1)p\left(c_1\right)\&GreaterEqual;\hat{p}(x;{\mathrm{\&beta;}}_{n_0},h_0|c_0)p\left(c_0\right),& x\&Element;c_1\\ \mathrm{else}& & x\&Element;c_0\end{array}\right\}---\left(3\right)$

Therefore, Weak Classifier h _tx the expression formula of () is

$h_t\left(x\right)=\left\{\begin{array}{ccc}1& \mathrm{if}& \hat{p}(x;{\mathrm{\&beta;}}_{n_1},h_1|c_1)p\left(c_1\right)\&GreaterEqual;\hat{p}(x;{\mathrm{\&beta;}}_{n_0},h_0|c_0)p\left(c_0\right)\\ 0& \mathrm{otherwise}& \end{array}\right.---\left(4\right)$

Wherein two kind attribute lower probability density with sparse estimation expression be respectively:

$\hat{p}(x;{\mathrm{\&beta;}}_{n_0},h_0|c_0)={\mathrm{\&Sigma;}}_{k=1}^{m_0}{\mathrm{\&beta;}}_k{K}_{h_0}(x,x_k)---\left(5\right)$

$\hat{p}(x;{\mathrm{\&beta;}}_{n_1},h_1|c_1)={\mathrm{\&Sigma;}}_{k=1}^{m_1}{\mathrm{\&beta;}}_k{K}_{h_1}(x,x_k)---\left(6\right)$

M ₀and m ₁the number of non-zero core weights in sparse Density Estimator RSDE-WL1 expression formula under correspond to two kind attributes respectively, n ₀and n ₁correspond to the number of non-meteorite crater and meteorite crater example respectively, usual m ₀< n ₀and m ₁< n ₁. with for core weight vector, β _kfor core weights coefficient (0≤β _k≤ 1), h ₀and h ₁for the wide (h of nucleus band ₀> 0, h ₁> 0), with for kernel function.

Wherein, the simple realization process of the sparse Density Estimator RSDE-WL1 of improvement is as follows:

First compression collection density Estimation (RSDE) algorithm is introduced.RSDE based on experience points square error (ISE) criterion, with total regression matrix based on, make core weights as much as possible be tending towards 0, thus obtain the sparse expression formula of density p (x), wherein K _i,k=K _h(x _i, x _k) be Φ _nthe i-th row k column element.In particular, the RSDE with gaussian kernel estimates, its core weight vector β can obtain by minimizing integrated square error, as follows: wherein, represent N × N and tie up space of matrices;

In formula (7), parameter beta is consistent with the meaning of parameters in formula (2), and dx represents differential term, E _{p (x)}represent expectation value;

Wherein item has nothing to do due to itself and parameter beta, can not consider, E _{p (x)}{ } represents the expectation value about density p (x).By Density Estimator expression formula substitution formula (7), through series of transformations, obtains the non-negative double optimization problem of belt restraining

$\mathrm{\&beta;}=\underset{\mathrm{\&beta;}}{\arg \min }\{\frac{1}{2}{\mathrm{\&beta;}}_N^T{C}_N{\mathrm{\&beta;}}_N-{\hat{P}}_N^T{\mathrm{\&beta;}}_N\}---\left(8\right)$

Constraint condition β _k>=0,1≤k≤N and wherein, matrix element definition is g _h() is gaussian kernel function, and h is that nucleus band is wide, ${\hat{p}}_N={[p_1{p}_2...p_N]}^T={[\hat{p}(x_1;h)\hat{p}(x_2;h)...\hat{p}(x_N;h)]}^T$ The Parzen window estimated value vector of each sample point, β _n=[β ₁, β ₂..., β _n] ^t.

In order to reduce the degree of rarefication of the aggregation extent of weights coefficient in some region and raising density Estimation, we introduce the weighting l of weights coefficient ₁norm as penalty term, the sparse Density Estimator algorithm RSDE-WL1 be improved. also regularization term is called, wherein for diagonal matrix.Definition w=[w ₁, w ₂..., w _n] ^t, β _n=[β ₁, β ₂..., β _n] ^t, adding the new double optimization problem after penalty term is

$\begin{array}{c}\mathrm{\&beta;}=\underset{\mathrm{\&beta;}}{\arg \min }\{\frac{1}{2}{\mathrm{\&beta;}}_N^T{C}_N{\mathrm{\&beta;}}_N-{\hat{P}}_N^T{\mathrm{\&beta;}}_N+\mathrm{\&lambda;}{w}^T{\mathrm{\&beta;}}_N\}\\ =\underset{\mathrm{\&beta;}}{\arg \min }\{\frac{1}{2}{\mathrm{\&beta;}}_N^T{C}_N{\mathrm{\&beta;}}_N-{({\hat{P}}_N-\mathrm{\&lambda;w})}^T{\mathrm{\&beta;}}_N\}\end{array}---\left(9\right)$

Notice that problem (9) is non-convex, utilize corresponding iterative algorithm to solve sparse solution that the problems referred to above can obtain weights coefficient.

(2) optimal feature selection

Calculate Weak Classifier h _tx the weighted error summation of (), selects the single optimal characteristics f meeting least error ^tfor building the optimum Weak Classifier of current iteration

$f^t=\underset{f}{\arg \min }{\mathrm{\&Sigma;}}_{i=1}^n{w}_i^t|h(x_i,f)-y_i|---\left(10\right)$

h _t(x)＝h(x,f ^t) (11)

(3) next iteration process sample weights upgrades

SparseBoost algorithm combines classification results this information relevant of current sample weights in AdaBoost algorithm and selected feature in the past, and this information contributes to selecting current optimal characteristics.In implementation procedure, strengthen by the sample weights of mis-classification, and reduce by the sample weights of correctly classifying.When calculating weighted error summation, more may be selected in next iteration process by the sample of mis-classification.It is as follows that weight upgrades expression formula

$w_i^{t+1}=w_i^t{\mathrm{\&gamma;}}_t^{1-|h_t\left(x_i\right)-y_i|}---\left(12\right)$

3, advantage and effect

A kind of detection method based on meteorite crater in the Mars probes soft landing process of sparse lifting integrated classifier of the present invention, it compared with the conventional method, its major advantage is: the time complexity of (1) classifier algorithm is O (Tm (m ₀+ m ₁)), compared with Boost algorithm, the time complexity of Boost algorithm is O (Tmn), due to m ₀+ m ₁< n ₀+ n ₁=n, therefore has significant advantage on time complexity.(2) when each structure Weak Classifier, the feature that only selection one is optimum as sample vector, and does not utilize whole feature, reduces computation complexity, accelerates classification speed.(3) classify to based on the meteorite crater of image and non-meteorite crater, classification accuracy can be similar to reach 85% and more than, have certain reference value to the detection of meteorite crater in the Mars probes soft landing process of reality.

Accompanying drawing explanation

Fig. 1 meteorite crater detects main frame figure automatically.

The physical principle that a meteorite crater is made up of the bright and dark area as crescent moon is explained in Fig. 2 (A) meteorite crater crescent moon region.

Bright and the dark area of the real 1km size meteorite crater of Fig. 2 (B).

Fig. 3 builds candidate's meteorite crater process flow diagram.

Fig. 4 (A) 2 kinds of 2-rectangles hide.

Fig. 4 (B) 2 kinds of 3-rectangles hide.

Fig. 4 (C) 5 kinds of 4-rectangles hide.

The example that Fig. 4 (D) hides meteorite crater 2-rectangle.

Fig. 5 (A) territory, West meteorite crater real image.

Fig. 5 (B) zone line meteorite crater real image.

Fig. 5 (C) territory, East meteorite crater real image.

Embodiment

A kind of detection method based on meteorite crater in the Mars probes soft landing process of sparse lifting integrated classifier of the present invention, the step that the method comprises is shown in Fig. 1.Its main thought makes full use of the advantage that designed sparse lifting integrated classifier has sparse solution and reduction computation complexity, after feature extraction and selection are carried out to the meteorite crater textural characteristics based on image, realize the classification to meteorite crater and non-meteorite crater, to reach the quick detection of meteorite crater.The physical principle that Fig. 2 (A)-(B) is bright in meteorite crater crescent moon region and dark area and real example.Fig. 3 is for building candidate's meteorite crater process flow diagram.

The present invention selects a part of High Resolution Stereo Camera (HRSC) full-colour image minimum point h0905_0000 as test set, this image is by the quick airship shooting of Mars, as shown in Fig. 5 (A)-(C), selected image resolution ratio is 12.5 meters/pixel, and size is 3000 × 4500 pixel (37500 × 56250m ²).Domain expert is manual on this pictures have been marked about 3500 meteorite craters and has compared as the truth on earth's surface and automatic testing result.This pictures is a great challenge for automatically detecting meteorite crater algorithm, because it contains the form with spatial variations, and the contrast of image is on duty mutually.This pictures is divided into three parts, be designated as territory, West, zone line and territory, East, there are similar landforms in territory, West and territory, East, but there is more meteorite crater in territory, West than East territory, and zone line has visibly different surperficial geographic entity compared with other two regions.

Step 1,2: determine candidate's meteorite crater and candidate's meteorite crater Texture Feature Extraction

According to the defining method of candidate's meteorite crater, we tentatively determine 13075 candidate's meteorite craters from full-colour image 5.By the method that Fig. 4 (A)-(D) 9 kinds side row hides, from candidate's meteorite crater image, 1089 image texture characteristics are extracted.Training set Stochastic choice 204 true meteorite craters and 292 non-meteorite crater examples compositions from the half of candidate's meteorite crater in north, territory, East in experiment, are respectively 2935,1181 and 1223 from test set corresponding candidate's meteorite crater number in territory, West, middle region and territory, East in experiment.

Step 3: feature selecting is carried out to the textural characteristics extracted

The least possible in order to realize selected characteristic number, and classify accuracy is large as far as possible, the iterations T of the present invention to SpraseBoost algorithm is set to 2,5,10,15,20,25,30,50,100,150,200 respectively, then in territory, West, middle region and territory, East test the classification results choosing individual features subset respectively.Simultaneously due to the unbalancedness that meteorite crater in candidate's meteorite crater data and non-meteorite crater distribute, successfully detect that true meteorite crater is more important than non-meteorite crater.Therefore, the present invention uses accuracy rate (Accuracy=(TP+TN)/(TP+TN+FP+FN)), recall ratio (Recall=TP/ (TP+FN)), precision ratio (Precision=TP/ (TP+FP)) and F measured value (F-measure=2/ (1/ (Recall)+1/ (Precision))) as evaluation index.Wherein TP represents by the true meteorite crater number of correctly classifying, and TN represents by the non-meteorite crater number of correctly classifying, and FP represents by mis-classification to be the non-meteorite crater number of meteorite crater, and FN represents by mis-classification to be the true meteorite crater number of non-meteorite crater.

In the arranging of iterations T, select 2 features mainly because candidate's meteorite crater is made up of bright and dark area, therefore these two features that representative is bright and dark are most important, maximum to choose 200 features be in order to compared with the experimental result of other documents, and middle characteristic number is random selecting.In feature selection process, utilize the thinking of Boost algorithm, in T iterative process, choose the single optimal characteristics f meeting least error each time ^t, obtain the structure of T feature for sample set, wherein T < n.

Step 4: be combined with sparse Density Estimator algorithm RSDE-WL1 by Boost algorithm, devise sparse lifting integrated classifier, to realize the quick detection to the meteorite crater based on image.

In the design of sorter, mainly comprise the step of three cores: Weak Classifier study, optimal feature selection and next iteration process sample weights upgrade.First initialization is carried out, for the initial weight w of input amendment _iif, y _i=0, then w _i=1/2n ₀; If y _i=1, then w _i=1/2n ₁, wherein n ₀and n ₁correspond to the number of non-meteorite crater and meteorite crater example respectively, n ₀+ n ₁=n.In Weak Classifier study, parameters λ=0.001, l _max=8, ε=1/ (0.3*N), the density Estimation expression formula of the sparse Density Estimator device RSDE-WL1 be improved, then according to Bayesian decision criterion, obtain a Weak Classifier h in each iterative process _t(x).In optimal feature selection, calculate the error in classification of Weak Classifier choose and meet the minimum feature of error in classification as there being most feature, γ is set _t=ε _t/ 1-ε _t.It is finally the weight of next iteration process more new samples thus obtain T Weak Classifier.Be weighted by this T Weak Classifier and be combined into final strong classifier, the weight of corresponding Weak Classifier is set to α _t=ln (1/ γ _t).

Table 1-3 show respectively on territory, West, zone line and territory, East, characteristic number selected by different iterations and classification accuracy, recall ratio, precision ratio and F measured value.As can be seen from table 1-3, on territory, West, zone line and region, three, territory, East, when selected characteristic number is respectively 10,20, when 20, corresponding classification accuracy reaches the highest, be respectively 0.790,0.854 and 0.874, F measured value also reaches maximum simultaneously, be respectively 0.790,0.796 and 0.818.Therefore, when training set is determined, the optimal characteristics number chosen when classifying to trizonal test data is respectively 10,20 and 20.

Territory, table 1 West

Iterations T	Selected characteristic number	Accuracy	Recall	Precision	F-measure
						2	2	0.783	0.840	0.737	0.785
5	5	0.788	0.814	0.765	0.789
						10	10	0.790	0.796	0.767	0.790
15	15	0.788	0.779	0.774	0.776
						20	20	0.775	0.724	0.783	0.752
25	25	0.771	0.714	0.781	0.746
						30	30	0.763	0.691	0.782	0.734
50	50	0.740	0.625	0.781	0.694
						100	100	0.688	0.502	0.755	0.603

150	150	0.658	0.423	0.737	0.541
						200	200	0.638	0.378	0.721	0.496

Table 2 zone line

Iterations T	Selected characteristic number	Accuracy	Recall	Precision	F-measure
						2	2	0.850	0.761	0.827	0.751
5	5	0.848	0.743	0.836	0.760
						10	10	0.843	0.700	0.854	0.769
15	15	0.844	0.689	0.869	0.768
						20	20	0.854	0.761	0.834	0.796
25	25	0.841	0.709	0.842	0.770
						30	30	0.837	0.707	0.832	0.764
50	50	0.831	0.661	0.854	0.746
						100	100	0.813	0.587	0.873	0.702
150	150	0.782	0.497	0.866	0.631
						200	200	0.789	0.512	0.873	0.646

Territory, table 3 East

Iterations T	Selected characteristic number	Accuracy	Recall	Precision	F-measure
						2	2	0.861	0.755	0.838	0.801
5	5	0.867	0.746	0.858	0.811
						10	10	0.865	0.738	0.883	0.804
15	15	0.868	0.725	0.905	0.805
						20	20	0.874	0.753	0.894	0.818
25	25	0.868	0.716	0.911	0.802
						30	30	0.867	0.716	0.909	0.801
50	50	0.863	0.699	0.914	0.792
						100	100	0.860	0.666	0.941	0.780
150	150	0.841	0.611	0.946	0.743
						200	200	0.842	0.611	0.950	0.744

Choose four kinds of supervised classification algorithms detected for meteorite crater and the SparseBoost algorithm that proposes of the present invention compares, as Boost, AdaBoost, SVM and J48 algorithm.Boost algorithm adopts decision tree as base sorter, lifting Integrated Algorithm and feature selecting algorithm is merged and classifies.And other three kinds of algorithms can not carry out feature selecting to raw data set in experimentation.Table 4 lists the classification accuracy (Accuracy) of these five kinds of algorithms on territory, West, zone line and territory, East, recall ratio (Recall), precision ratio (Precision) and F measured value (F-measure).

Table 4 meteorite crater detection algorithm of the present invention and other four kinds of meteorite crater detection algorithm Performance comparision

As can be seen from Table 4, on territory, West and territory, East, have the classification accuracy of the sorting algorithm (SparseBoost and Boost) of feature selecting and F measured value all apparently higher than the sorting algorithm (AdaBoost, SVM and J48) without feature selecting, wherein SparseBoost algorithm is better than Boost algorithm.And at zone line, there is the algorithm of feature selecting compared with the algorithm without feature selecting, difference little (as SparseBoost, AdaBoost and J48) on classification accuracy and F measured value, even poorer (as Boost), this may be that training sample owing to taking is from territory, East, and the special geomorphology that middle region is different from territory, East causes losing some important characteristic informations during feature selecting, can find out that from Fig. 5 (B) the geographical form in middle region obviously has larger difference with west, territory, East.Therefore, totally it seems, SparseBoost sorting algorithm proposed by the invention has good classifying quality in meteorite crater context of detection, and computation complexity is minimum.

It should be noted last that: above embodiment is the unrestricted technical scheme of the present invention in order to explanation only, although with reference to above-described embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that: still can modify to the present invention or equivalent replacement, and not departing from any modification or partial replacement of the spirit and scope of the present invention, it all should be encompassed in the middle of right of the present invention.

Claims (1)

1., based on the detection method of meteorite crater in the Mars probes soft landing process of sparse lifting integrated classifier, it is characterized in that: the method comprises the following steps:

Step 1. determines candidate's meteorite crater

Determine that the key of candidate's meteorite crater is that the meteorite crater in image regards a pair crescent moon with bright and dark area as, the shape of often pair of crescent moon is determined from image by the shape detecting method based on mathematical morphology, the crescent moon meteorite crater alternatively of coupling; The building process of candidate's meteorite crater is input is a full-colour image, it comprises a lot of bright and dark features region; Bright and dark shape parallel processing, by using original image process to become clear shape, and uses inverted image to process dark shape; The target of the method eliminates all noise characteristics that cannot be designated as meteorite crater, and only retain bright and dark features; Remaining bright and dark features region matches each other, and these area marking are good, as the candidate region of meteorite crater;

Step 2. candidate meteorite crater Texture Feature Extraction

In order to represent single candidate's meteorite crater from rectangular characteristic aspect, first extract the square image block around candidate's meteorite crater; In an experiment, in order to comprise the edge of meteorite crater around, use the twice of candidate's meteorite crater size as masked areas; The textural characteristics of each candidate's meteorite crater the unknown uses the square covering of 9 kinds of different sizes to encode, and therefore, the attribute of candidate's meteorite crater that image comprises is described by thousands of textural characteristics; These features are not separate, and these complete features compensate for by the single square restriction hiding the texture information obtained;

Step 3. carries out feature selecting to the textural characteristics extracted

According to the preliminary candidate's meteorite crater textural characteristics extracted, due to the higher-dimension of characteristic, therefore by before training sample and test sample book input sorter, feature selecting must be carried out; The SparseBoost algorithm utilizing step 4 to design carries out feature selecting, and the maximum difference of itself and AdaBoost algorithm is, the former only chooses an optimum feature in iterative process each time, and the latter utilizes whole feature set usually; Which decrease the intrinsic dimensionality of training sample, effectively reduce the computation complexity of sorter training;

Boost algorithm is combined with sparse Density Estimator algorithm RSDE-WL1 by step 4., devises sparse lifting integrated classifier, to realize the quick detection to the meteorite crater based on image;

According to selected candidate's meteorite crater textural characteristics, in order to distinguish wherein meteorite crater and non-meteorite crater, devising SparseBoost algorithm and carrying out supervised learning classification; The method is in conjunction with the sparse Density Estimator algorithm RSDE-WL1 of Boost algorithm and a kind of improvement, while selection character subset, construct some sparse Density Estimator devices for the design of corresponding base sorter, by the weighted array of base sorter, finally realize integrated classifier;

Given n candidate's meteorite crater (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n), wherein y _i=0,1, i=1,2 ..., n correspond to non-meteorite crater (c respectively ₀) and meteorite crater (c ₁) example, n ₀and n ₁correspond to the number of non-meteorite crater and meteorite crater example respectively, n ₀+ n ₁=n; Each candidate's meteorite crater is expressed as a proper vector x=(f ₁, f ₂..., f _m) ^t, wherein each feature f _i, i=1 ..., m is produced by the square covering of ad-hoc location a certain on candidate's meteorite crater, and m is the feature sum extracted; SparseBoost algorithm is utilized to build a series of Weak Classifier h _t(x), and by weighting integrated approach Weak Classifier carried out combining and set up final strong classifier H (x):

$H\left(x\right)=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t{h}_t\left(x\right)---\left(1\right)$

Wherein T is iterations (T < n), α _tthe Weak Classifier h of study _tthe weight of (x);

In each iterative process, need the step realizing following three cores: Weak Classifier study, optimal feature selection and next iteration process sample weights upgrade; Wherein, in Weak Classifier learning process, penalty term is added, the sparse Density Estimator algorithm RSDE-WL1 be improved to compression collection density Estimation algorithm RSDE; Utilize RSDE-WL1 algorithm to estimate its probability density function to each category attribute, classify according to the sample of Bayes decision rule to input, obtain Weak Classifier;

(1) Weak Classifier study

In the t time iterative process, for the single optimal characteristics f ∈ { f chosen ₁, f ₂..., f _mthe Weak Classifier h that builds _tx (), realizes by building Bayes classifier; Before the two classification problem of classifying about meteorite crater and non-meteorite crater is discussed, first introduce Bayes classifier; For Bayes's classification problem, other posterior probability density of expectation estimation given input amendment x lower class; In order to obtain one about the probability classification of density Estimation, first train a Multilayer networks device for each category attribute c wherein x is the proper vector for representing single candidate's meteorite crater, and β is core weight vector, and c is the category attribute of candidate's meteorite crater, c ∈ { c ₀, c ₁, c ₀represent non-meteorite crater classification, c ₁represent meteorite crater classification; Then use Bayes rule (2) to calculate posterior probability, final test sample is assigned to the category attribute with maximum a posteriori probability;

$\hat{p}\left(c\right|x;\mathrm{\&beta;})=\frac{\hat{p}(x;\mathrm{\&beta;}|c)p\left(c\right)}{{\mathrm{\&Sigma;}}_{i=0}^1\hat{p}(x;\mathrm{\&beta;}|c_i)p\left(c_i\right)}---\left(2\right)$

For classification problem, first estimate two conditional probability densities under given classification with these two density are obtained by follow-up sparse Density Estimator RSDE-WL1; Then according to formula (2), corresponding posterior probability is calculated respectively with according to the sample size of each category attribute, calculate the prior probability p (c of two kinds ₀) and p (c ₁): p (c ₀)=n ₀/ n, p (c ₁)=n ₁/ n, p (c ₀)+p (c ₁)=1; The sample of Bayes decision rule (3) to input is finally utilized to classify:

$\left.\begin{array}{cc}\mathrm{if}\hat{p}(x;{\mathrm{\&beta;}}_{n_1},h_1|c_1)p\left(c_1\right)\&GreaterEqual;\hat{p}(x;{\mathrm{\&beta;}}_{n_0},h_0|c_0)p\left(c_0\right),& x\&Element;c_1\\ \mathrm{else}& x\&Element;c_0\end{array}\right\}---\left(3\right)$

Therefore, Weak Classifier h _tx the expression formula of () is:

$h_t\left(x\right)=\left\{\begin{array}{cc}1& \mathrm{if}\hat{p}(x;{\mathrm{\&beta;}}_{n_1},h_1|c_1)p\left(c_1\right)\&GreaterEqual;\hat{p}(x;{\mathrm{\&beta;}}_{n_0},h_0|c_0)p\left(c_0\right)\\ 0& \mathrm{otherwise}\end{array}\right.---\left(4\right)$

Wherein two kind attribute lower probability density with sparse estimation expression be respectively:

$\hat{p}(x;{\mathrm{\&beta;}}_{n_0},h_0|c_0)={\mathrm{\&Sigma;}}_{k=1}^{m_0}{\mathrm{\&beta;}}_k{K}_{h_0}(x,x_k)---\left(5\right)$

$\hat{p}(x;{\mathrm{\&beta;}}_{n_1},h_1|c_1)={\mathrm{\&Sigma;}}_{k=1}^{m_1}{\mathrm{\&beta;}}_k{K}_{h_1}(x,x_k)---\left(6\right)$

M ₀and m ₁the number of non-zero core weights in sparse Density Estimator RSDE-WL1 expression formula under correspond to two kind attributes respectively, n ₀and n ₁correspond to the number of non-meteorite crater and meteorite crater example respectively, m ₀< n ₀and m ₁< n ₁; with for core weight vector, β _kfor core weights coefficient, 0≤β _k≤ 1; h ₀and h ₁for nucleus band is wide, h ₀> 0, h ₁> 0; with for kernel function;

Wherein, the implementation procedure of the sparse Density Estimator RSDE-WL1 of improvement is as follows:

First compression collection density Estimation algorithm RSDE is introduced; RSDE based on experience points square error criterion ISE, with total regression matrix based on, make core weights be tending towards 0, thus obtain the sparse expression formula of density p (x), wherein K _i,k=K _h(x _i, x _k) be Φ _nthe i-th row k column element; RSDE with gaussian kernel estimates, its core weight vector β can obtain by minimizing integrated square error, as follows: wherein, represent N × N and tie up space of matrices;

In formula (7), parameter beta is consistent with the meaning of parameters in formula (2), and dx represents differential term, E _{p (x)}represent expectation value;

Wherein, item has nothing to do due to itself and parameter beta, need not consider, E _{p (x)}{ } represents the expectation value about density p (x); By Density Estimator expression formula substitution formula (7), through series of transformations, obtains the non-negative double optimization problem of belt restraining

$\mathrm{\&beta;}=\underset{\mathrm{\&beta;}}{\arg \min }\{\frac{1}{2}{\mathrm{\&beta;}}_N^T{C}_N{\mathrm{\&beta;}}_N-{\hat{P}}_N^T{\mathrm{\&beta;}}_N\}---\left(8\right)$

Constraint condition β _k>=0,1≤k≤N and wherein, matrix element definition is g _h() is gaussian kernel function, and h is that nucleus band is wide, ${\hat{p}}_N={[p_1{p}_2\&CenterDot;\&CenterDot;\&CenterDot;p_N]}^T={[\hat{p}(x_1;h)\hat{p}(x_2;h)\&CenterDot;\&CenterDot;\&CenterDot;\hat{p}(x_N;h)]}^T$ The Parzen window estimated value vector of each sample point, β _n=[β ₁, β ₂..., β _n] ^t;

In order to reduce the degree of rarefication of the aggregation extent of weights coefficient in some region and raising density Estimation, introduce the weighting l of weights coefficient ₁norm as penalty term, the sparse Density Estimator algorithm RSDE-WL1 be improved; also regularization term is called, wherein for diagonal matrix; Definition w=[w ₁, w ₂..., w _n] ^t, β _n=[β ₁, β ₂..., β _n] ^t, adding the new double optimization problem after penalty term is

$\begin{array}{c}\mathrm{\&beta;}=\underset{\mathrm{\&beta;}}{\arg \min }\{\frac{1}{2}{\mathrm{\&beta;}}_N^T{C}_N{\mathrm{\&beta;}}_N-{\hat{P}}_N^T{\mathrm{\&beta;}}_N+{\mathrm{\&lambda;w}}^T{\mathrm{\&beta;}}_N\}\\ =\underset{\mathrm{\&beta;}}{\arg \min }\{\frac{1}{2}{\mathrm{\&beta;}}_N^T{C}_N{\mathrm{\&beta;}}_N-{({\hat{P}}_N-\mathrm{\&lambda;w})}^T{\mathrm{\&beta;}}_N\}\end{array}---\left(9\right)$ Notice that problem (9) is non-convex, utilize corresponding iterative algorithm to solve sparse solution that the problems referred to above obtain weights coefficient;

(2) optimal feature selection

Calculate Weak Classifier h _tx the weighted error summation of (), selects the single optimal characteristics f meeting least error ^tfor building the optimum Weak Classifier of current iteration

$f^t=\underset{f}{\arg \min }{\mathrm{\&Sigma;}}_{i=1}^n{w}_i^t|h(x_i,f)-y_i|---\left(10\right)$

h _t(x)＝h(x,f ^t) (11)

(3) next iteration process sample weights upgrades

SparseBoost algorithm combines classification results this information relevant of current sample weights in AdaBoost algorithm and selected feature in the past, and this information contributes to selecting current optimal characteristics; In implementation procedure, strengthen by the sample weights of mis-classification, and reduce by the sample weights of correctly classifying; When calculating weighted error summation, can be selected in next iteration process by the sample of mis-classification; It is as follows that weight upgrades expression formula

$w_i^{t+1}=w_i^t{\mathrm{\&gamma;}}_t^{1-|h_t\left(x_i\right)-y_i|}.---\left(12\right)$