CN103456011A - Improved hyperspectral RX abnormal detection method by utilization of complementary information - Google Patents

Abstract

The invention relates to an improved hyperspectral RX abnormal detection method by the utilization of complementary information. A classic RX operator is respectively used for performing abnormal detection on visible near-infrared data and short wave infrared data in the same scene, and target judgment of the initial abnormal detection is obtained. On the basis, sensors are used for obtaining redundancy and complementarity of the information. Combined with a decision fusion method based on a rule, a final judged result of the RX abnormal detection is obtained. Experimental simulation is performed on actually-measured hyperspectral data, and the effectiveness of the algorithm is verified. The algorithm can effectively utilize redundancy and the complementary information of the two sets of data, and therefore the target detection rate is finally improved.

Description

A kind of high spectrum RX of modified method for detecting abnormality that utilizes complementary information

Technical field

The invention belongs to a kind of high spectrum RX of modified method for detecting abnormality that utilizes complementary information, belong to hyperspectral data processing method and applied technical field, be applicable to the high-spectral data object detection field.

Background technology

Target in hyperspectral remotely sensed image organically combines traditional spatial image dimension information and the spectrum dimension information of reflection atural object radiation characteristic together, has very high spectral resolution, for the meticulous target detection of atural object provides good condition.The Hyperspectral imaging target detection technique is accompanied by the development of imaging spectral technology and rises, and in the remote sensing image application, has very important significance.Usually algorithm of target detection need to be used scene priori to a certain degree, comprises the priori image informations such as expertise, the curve of spectrum.Yet, in practice, these prioris are difficult to obtain.And still lack at present effective spectra inversion algorithm, curve of spectrum storehouse also is left to be desired, therefore, without the Outlier Detection Algorithm of priori, in the remote sensing image application, have very important significance.

Outlier Detection Algorithm is considered as the singular point under certain distribution occasion by the little target in high spectrum image, and the impact point that makes to convert in rear data by particular procedure is outstanding, and then automatically detects abnormal object.Wherein the most representative method is exactly the RX method that Reed and Yu propose.Its hypothesis local background obedience multivariate normal distribution, describe background model by the covariance matrix of estimating local background.Similar algorithm also has the UTD algorithm, and this algorithm replaces with even vector to the matched signal in the RX algorithm by measurement signal.The people such as Chang have proposed a lot of improvement algorithms on the basis of classical RX algorithm, such as methods such as adopting correlation matrix, describe background model, and then the detection operator be improved, and they mainly include CRXD, NRXD etc.In practical application, directly use RX to detect operator, while carrying out the high resolution image Anomaly target detection, often detect poor effect, can produce a lot of false-alarms zone.The non-singularity of not enough for the number of samples of estimate covariance matrix in local background, sample and the factors such as " pollutions " that is vulnerable to abnormal data all can have a strong impact on the estimation of background covariance matrix, make the covariance matrix of estimation can not the accurate description background model, thereby cause the detection hydraulic performance decline of RX algorithm.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, propose a kind of RX Hyperspectral imaging method for detecting abnormality based on decision level fusion.

The inventive method comprises the following steps:

1) pre-service of high-spectral data.

Each wave band of spectrum picture is realized showing with gray-scale map, spectrum picture to each wave band carries out visual examination, determine that the effect due to Atmospheric Absorption, refraction and Scattering Factors produces the abnormal wave band of considerable influence and due to other factors, spectroscopic data produced the abnormal wave band of larger noise spectroscopic data, described abnormal wave band refers to imaging to show the atural object characteristic wave bands, abnormal wave band is done to direct rejecting and process.The pretreated method of data is:

data ₁＝pre_processing(data ₁)，data ₂＝pre_processing(data ₂)

2) registration of image.

At first visible near-infrared and short-wave infrared image are carried out to image registration and image filtering.Then the short-wave infrared image after registration and visible near-infrared image are carried out respectively to the processing of RX abnormality detection, obtain a preliminary decision-making, obtain the bianry image that two width detect.Select the fusion standard of identical false alarm rate as preliminary decision-making, use the RX method for detecting abnormality:

out ₁＝RX(data ₁，XJ)，out ₂＝RX(data ₂,XJ)

The preliminary target decision result obtained.

Be below the RX method for detecting abnormality:

Be the high resolution image data that N, image dimension are B for the pixel number, suppose that the observation pixel means the spectrum vector x that a dimension is B, wherein x=[x ₁, x ₂..., x _b], corresponding binary object detection hypothesized model is as follows:

H ₀:x(n)＝x ⁰(n) （1）

H ₁:x(n)＝x ⁰(n)+t

Wherein, x ⁰(n) represent non-target background pixel, t represents goal pels, H ₀be expressed as the hypothesis of background clutter, H ₁mean the hypothesis that contains target.

On two-value statistical model basis, adopt the generalized likelihood-ratio test method to deduce and show that RX detects operator, its expression formula is:

$\mathrm{\δ}\left(x\right)={(x-\mathrm{\μ})}^T{(\frac{N}{N+1}{C}^{-1}+\frac{1}{N+1}(x-\mathrm{\μ}){(x-\mathrm{\μ})}^T)}^{-1}(x-\mathrm{\μ})---\left(2\right)$

Adopt the reduced form of RX operator:

δ _RXD(x)＝(x-μ) ^TC ^-1(x-μ)

$C=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}(x-u){(x-u)}^T---\left(4\right)$

$u=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}x---\left(5\right)$

In formula: the decision threshold that η is operator; The estimated value that μ is the background mean value vector; The estimated value of the covariance matrix that C is the background sample.

3) the decision level rule merges.

According to the Expert opinion zone, divide target area and background area, carry out rule and merge preparation.Target area is designated as to area ₁, background area is designated as area ₂; Then carry out the "or" logic in target area, in background area, carry out " with " logic." from " and "or" concrete operations and different remote sensing instrument between complementation and redundancy closely related, be that the arithmetical logic of reflection multi-sensor information fusion is described.Total target is when removing partial redundance information, give prominence to and embody the message complementary sense that the two sensors foreign matter detects.Carrying out rule fusion concrete grammar at fusion center is:

Obtain fusion results.

Above-mentioned various in, data ₁, data ₂be respectively the visible near-infrared data of 75 wave band and 80 wave band short-wave infrared data; Pre_processing is the pre-processing image data function; XJ means false alarm rate, as the bridge that connects two groups of preliminary rulings; out ₁, out ₂be respectively the preliminary target decision result after the RX abnormality detection of visible near-infrared, short-wave infrared data;

with

mean respectively the bianry image after the visible near-infrared and short-wave infrared RX in target area detects,

with

mean respectively the bianry image after the visible near-infrared and short-wave infrared RX in background area detects, fusion (x, y) is the fusion center court verdict.

The present invention is directed to the data characteristics of high-spectrum remote-sensing, at first experimental data is carried out to pre-service, the image after registration is carried out respectively to the RX abnormality detection, obtain preliminary decision-making, finally at fusion center, carry out decision level fusion, obtain final testing result.Parameter in the present invention arranges simply, has obtained and has detected preferably effect, has made up the limitation of local background's covariance matrix in traditional RX algorithm.

The accompanying drawing explanation

Fig. 1 is decision-making and fusion figure;

Fig. 2 is the inventive method process flow diagram;

The visible near-infrared gray-scale map that Fig. 3 (a) is the 30th wave band;

The short-wave infrared gray-scale map that Fig. 3 (b) is the 30th wave band;

Fig. 4 (a) is visible near-infrared Data Detection figure as a result;

Fig. 4 (b) is short-wave infrared Data Detection figure as a result;

Fig. 4 (c) is Pixel-level fusion detection figure as a result;

Fig. 4 (d) is decision level rule fusion detection figure as a result.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

What the present embodiment adopted is one group of actual measurement high-spectral data, after contextual data is carried out to image registration, Band fusion, these group data are 226 Χ 502 pixels, there are two kinds of describing modes, respectively the visible and near infrared spectrum data of 80 wave bands, the short-wave infrared data of 75 wave bands, Fig. 3 (a) is the visible near-infrared gray-scale map of the 30th wave band, Fig. 3 (b) is the short-wave infrared gray-scale map of the 30th wave band.

Experiment scene has comprised take 1 van, 2 semitrailers of a large amount of vegetation (willow woods and harvested milpa) under overall background.Include 10 two class targets (being applied as respectively green material A, material B) that scribble the camouflage paint vehicle in scene, they are placed respectively up and down.

By the inventive method process flow diagram, as Fig. 2, can find out clearly design procedure of the present invention.

At first, the good visible near-infrared and short-wave infrared data to registration, utilize the airspace filter method to due to the effect of Atmospheric Absorption, refraction and Scattering Factors, spectroscopic data being produced to the abnormal wave band of considerable influence in data and abnormal wave band that other factors produce larger noise to spectroscopic data carries out filtering, obtain the data data after airspace filter ₁and data ₂.Again, the image after airspace filter is carried out to the RX abnormality detection, using identical false alarm rate as the judgment condition that concerns of decision level fusion, the result data after the RX abnormality detection is out ₁and out ₂.For ease of the superiority of comparison algorithm, experiment is all to be controlled under an identical false alarm rate (0.47%), and the foreign matter target of then selecting thresholding to complete Various types of data detects automatically.Testing result is as Fig. 4 (a) and Fig. 4 (b).For outstanding decision level fusion effect is very good, added one group of testing result that Pixel-level merges in experiment, as Fig. 4 (c), finally RX testing result figure is carried out to the decision level rule and merge, the decision level fusion process flow diagram, as Fig. 1.Obtain fusion results, as Fig. 4 (d).Table 1 is the experimental data statistics table of comparisons.

The table 1 experimental result statistics table of comparisons

From the statistics table of comparisons 1, can find out through rule-based decision level fusion, under identical false alarm rate (0.47%), 53.59% of the 15.93% and 75 wave band short-wave infrareds that the target detection rate is visible near-infrared from 80 wave bands are promoted to and merge 86.91%, and simple Pixel-level merges the complementary information that can not effectively utilize two groups of data (visible near-infrared with short-wave infrared data) under certain condition.In experiment, the verification and measurement ratio that Pixel-level merges is than the verification and measurement ratio of 75 wave band short-wave infrareds also low (53.06% pair 53.59%).This is mainly that this can affect the estimation of background covariance matrix, can not bring into play the usefulness that Pixel-level merges because there are some abnormal datas (because the band imaging limitation of infrared sensor causes) in the short-wave infrared data.Compare that Pixel-level merges and 80 wave bands are visible near-infrared, the testing result of 75 wave band short-wave infrareds, the algorithm that the present invention proposes can effectively utilize redundancy and the complementary information of two groups of data.By mutually revising by short-wave infrared and visible near-infrared data, utilized message complementary sense, finally improved the target detection rate.In addition, also illustrated that the decision level visual fusion has higher serious forgiveness, and there is stronger antijamming capability.

Claims (1)

1. the high spectrum RX of a modified method for detecting abnormality that utilizes complementary information is characterized in that the method comprises the following steps:

1) pre-service of high-spectral data;

Each wave band of spectrum picture is realized showing with gray-scale map, spectrum picture to each wave band carries out visual examination, determine that the effect due to Atmospheric Absorption, refraction and Scattering Factors produces the abnormal wave band of considerable influence and due to other factors, spectroscopic data produced the abnormal wave band of larger noise spectroscopic data, described abnormal wave band refers to imaging to show the atural object characteristic wave bands, abnormal wave band is done to direct rejecting and process; The pretreated method of data is:

，

2) registration of image;

At first visible near-infrared and short-wave infrared image are carried out to image registration and image filtering; Then the short-wave infrared image after registration and visible near-infrared image are carried out respectively to the processing of RX abnormality detection, obtain a preliminary decision-making, obtain the bianry image that two width detect; Select the fusion standard of identical false alarm rate as preliminary decision-making, use the RX method for detecting abnormality:

，

The preliminary target decision result obtained;

Be below the RX method for detecting abnormality:

For the pixel number, be , the image dimension is

high resolution image data, suppose observation pixel mean that a dimension is

the spectrum vector

, wherein

, corresponding binary object detection hypothesized model is as follows:

（1）

Wherein,

represent non-target background pixel,

represent goal pels,

be expressed as the hypothesis of background clutter,

mean the hypothesis that contains target;

On two-value statistical model basis, adopt the generalized likelihood-ratio test method to deduce and show that RX detects operator, its expression formula is:

（2）

Adopt the reduced form of RX operator:

（3）

（4）

（5）

In formula: decision threshold for operator; estimated value for the background mean value vector;

estimated value for the covariance matrix of background sample;

3) the decision level rule merges;

According to the Expert opinion zone, divide target area and background area, carry out rule and merge preparation; Target area is designated as , background area is designated as ; Then carry out the "or" logic in target area, in background area, carry out " with " logic; " from " and "or" concrete operations and different remote sensing instrument between complementation and redundancy closely related, be that the arithmetical logic of reflection multi-sensor information fusion is described; Total target is when removing partial redundance information, give prominence to and embody the message complementary sense that the two sensors foreign matter detects; Carrying out rule fusion concrete grammar at fusion center is:

Obtain fusion results;

Above-mentioned various in,

,

be respectively the visible near-infrared data of 75 wave band and 80 wave band short-wave infrared data;

for the pre-processing image data function;

mean false alarm rate, as the bridge that connects two groups of preliminary rulings;

,

be respectively the preliminary target decision result after the RX abnormality detection of visible near-infrared, short-wave infrared data;

with

mean respectively the bianry image after the visible near-infrared and short-wave infrared RX in target area detects,

with

mean respectively the bianry image after the visible near-infrared and short-wave infrared RX in background area detects,

for the fusion center court verdict.

Images