CN104408482A - Detecting method for high-resolution SAR (Synthetic Aperture Radar) image object - Google Patents

Abstract

The embodiment of the invention discloses a detection method for a high-resolution SAR (Synthetic Aperture Radar) image object and is applied to the technical field of image processing. According to the detection method, the defects that an existing detection method for the SAR image object is higher in detection false alarm and is inaccurate in positioning are mainly solved. The detection method comprises the following steps: performing overlapped blocking on an original SAR image to obtain a plurality of subimages; extracting the characteristics of a training sample for training a classifier, and performing catalog classification on the subimages by using the trained classifier; obtaining weighted scarcity characteristics of pixels of the subimages containing the objects in the catalog class, performing threshold segmentation on the subimages containing the objects, performing morphological processing on the segmented binary images to obtain a region of interest (ROI) of the subimages containing the objects in the object class, and correcting the ROI to obtain an object detection result. The detection method disclosed by the invention is applied to the detection of the high-resolution SAR image object with multiple object types.

Description

A kind of High Resolution SAR Images object detection method

Technical field

The invention belongs to technical field of image processing, relate to a kind of High Resolution SAR Images object detection method.Can be used for the object detection and recognition in High Resolution SAR image, computer vision and field of intelligent control.

Background technology

It is utilize the information such as the gray scale of target, texture, shape, edge and direction in SAR image, determine its position that synthetic-aperture radar (SAR) image object detects, by the image processing techniques that object and background is separated.SAR image target detection is the prerequisite of further target identification, is the hot issue of current Based Intelligent Control and computer vision field research.Due to the singularity of SAR imaging, SAR image and optical imagery are very different, and such as SAR image comprises more redundant information, there is speckle noise, and SAR target azimuthal is very responsive.So, from SAR image, how accurately to detect that target is the difficult point of SAR image decipher always.

Along with SAR image resolution improves constantly, target information presents explosive increase, and the difficult point of target detection mainly contains: (1) image scene becomes increasingly complex, and the data volume that need process is increasing; (2) target is by the point target on low-resolution image in original single channel single polarization, becomes the Area Objects having and enrich minutia and scattering signatures, and the kind of target signature and instability increase; (3) redundant information that image comprises increases, if processed all pixels of image, arithmetic speed can be affected.

CFAR method based on CFAR is classical SAR image object detection method.The method is that the clutter distribution parameter of estimated background, chooses the threshold value of detection adaptively, then utilizes the moving window traversing graph picture of fixed size, judges whether pixel belongs to target under the known condition of image overall or local background.When image detail information is not obvious, when background is more even, the method can realize the detection to target when not having the prior imformations such as target type, target property and background characteristics, and can obtain good testing result.

Although CFAR method can detect target more accurately when simply utilizing picture amplitude information, for High Resolution SAR image, detailed information is very abundant, and CFAR mainly exists following defect in the case:

(1) target detection false-alarm is higher, is unsuitable for the High Resolution SAR image that detailed information is abundant.CFAR method simply utilizes the amplitude information of image to carry out target detection, when the gray-scale value of background pixel and the gray-scale value of target close to time can be misjudged as target, cause testing result to occur higher false-alarm.Meanwhile, testing result is easily affected by noise, is unfavorable for the subsequent treatment of image, as target identification.

(2) what detect is suspected target regions all in image, inaccurate for real target localization.

Summary of the invention

The embodiment of the present invention provides a kind of High Resolution SAR Images object detection method, higher and locate inaccurate defect with the detection false-alarm solving existing SAR image object detection method.

A first aspect of the present invention provides a kind of High Resolution SAR image object detection method, comprising:

Overlap partition is carried out to the original SAR image of input, obtains multiple subimage;

The feature extracting training sample is trained sorter, carries out catalog classification with the sorter after training to multiple described subimage;

Obtain the rarity feature of weighting that catalogue classification is the pixel of the subimage comprising target, Threshold segmentation is carried out to the described subimage comprising target, Morphological scale-space is carried out to the bianry image after segmentation, obtains the region of interest ROI that catalogue classification is the subimage comprising target;

Described ROI is revised, obtains object detection results.

According to first aspect, in the implementation that the first is possible, the described original SAR image to input is carried out overlap partition and is comprised:

According to formula 1, overlap partition is carried out to original SAR image;

Wherein, m, n are respectively the length of original SAR image and wide, and piecemeal is of a size of x × x, and p% is Duplication, and the subimage obtained is designated as S _i, i=1,2 ..., N, N are the number obtaining multiple subimage.

According to first aspect, in the implementation that the second is possible, the feature of described extraction training sample is trained sorter, carries out catalog classification, comprising with the sorter after training to multiple described subimage:

Determine catalogue classification number, the target type number that wherein can be obtained comprising in SAR image by prior imformation is M, the corresponding two kinds of catalogue classifications of each target type, namely comprise this target and do not comprise this target, so total catalogue classification has 2M, and each catalogue classification is designated as C _q, q ∈ 2M, wherein C ₁, C ₂..., C _mfor comprising the catalogue category label of target;

Extract proper vector for described 2M catalogue classification random selecting training sample respectively;

Extract the proper vector of described subimage, catalog classification is carried out to described subimage.

The implementation possible according to the second of first aspect, in the implementation that the third is possible, describedly extract proper vector for described 2M catalogue classification random selecting training sample respectively, comprising:

Weber partial descriptions feature WLD is extracted as proper vector to each training sample, is designated as and sorter is trained;

Accordingly, the proper vector of the described subimage of described extraction, catalog classification is carried out to described subimage, comprising:

Extract the WLD feature of described subimage, utilize the sorter trained to carry out catalog classification to described subimage.

According to first aspect, in the 4th kind of possible implementation, described Threshold segmentation is carried out to the described subimage comprising target, comprising:

Carry out Threshold segmentation to the described subimage comprising target, setting threshold value is Th _u, by following formula to Image Segmentation Using, the image after segmentation is designated as BW, as formula 2:

$\mathrm{BW}=\left\{\begin{array}{c}1,U_k\≥{\mathrm{Th}}_u\\ 0,U_k\≥{\mathrm{Th}}_u\end{array}\right.---\left(2\right)$

In formula 2, U _kfor the rare value of weighting of a kth pixel.

According to first aspect, in the 5th kind of possible implementation, the rarity feature of described weighting, comprising:

Calculate the space distribution weights of a kth pixel, be designated as ω _k, k=m _sub* n _sub, ω _kcalculate as formula 3:

${\mathrm{\ω}}_k=\sqrt{{(I_k-I_0)}^2+{(I_k-I_0)}^2}+\frac{\underset{j=1}{\overset{\mathrm{num}}{\mathrm{\Σ}}}\sqrt{{(I_k-I_j)}^2+{(I_k-I_j)}^2}}{\mathrm{num}}---\left(3\right)$

Wherein, in formula 3, m _sub, n _subbe respectively subimage long and wide, I _kfor the gray-scale value of a subimage kth pixel, I ₀for the pixel value of subimage center pixel, num be in subimage with I _kthe number of equal pixel;

Calculate the rare value of a kth pixel, be designated as u _k, k=m _sub* n _sub, u _kcalculate as formula 4:

$u_k=\left\{\begin{array}{cc}255,& \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k}\&GreaterEqual;255\\ \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k},& \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k}<255\end{array}\right.---\left(4\right)$

Wherein, in formula 4, m _sub, n _subbe respectively subimage long and wide, un _kfor the number of times that the pixel of kth in subimage occurs;

Calculate the rare value of weighting of a kth pixel, be designated as U _k, k ∈ m*n, U _kcalculate as formula 5;

U _k＝ω _k*u _k(5)

According to first aspect, in the 6th kind of possible implementation, describedly correction carried out to described ROI comprise:

The ROI number obtained is designated as L, and t ROI is ROI _t;

Judge ROI _s, s ∈ L and ROI _t, whether t ∈ L is connected region, wherein, and s ≠ t.If ROI _swith ROI _tfor connected region, then they are merged, obtain a new ROI; Repeat this step, until traveled through all ROI.

The High Resolution SAR Images object detection method that the embodiment of the present invention provides, by carrying out overlap partition to image, obtain the subimage that size is unified, then carry out cataloguing to all subimages, the type namely whether comprising target according to subimage and comprise target carries out catalog classification to it, obtains the catalogue that comprises all subimages after image directory.The last subimage comprising target of selecting from catalogue carries out target detection, obtain interested region (ROI), ROI is revised and obtains final testing result, remove redundant information, decrease the impact of complex background on target detection, the specific aim of target detection is strengthened, and the result obtained is more accurate.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below.

Fig. 1 is the process flow diagram of the High Resolution SAR image object detection that the embodiment of the present invention provides;

Fig. 2 is the original SAR image that emulation experiment of the present invention uses;

Fig. 3 is the result using CFAR algorithm image to be carried out to target detection;

Fig. 4 is the present invention's image block schematic diagram used;

Fig. 5 is the μ that the present invention uses _l, μ _r, μ _d, μ _ucomprise the particular location schematic diagram of pixel;

Fig. 6 is antithetical phrase image directory of the present invention sorted result figure, C ₁class is that meadow comprises target, C ₂class is that meadow is not containing target, C ₄class is that field is not containing target, C ₆class is that forest is not containing target, C ₈class is that lake is not containing target;

Fig. 7 is that the present invention is to class C ₁the subimage comprising target carries out the result of target detection.

Embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, be clearly and completely described the technical scheme in the embodiment of the present invention, obviously, described embodiment is only the present invention's part embodiment, instead of whole embodiments.

The object of the invention is to overcome the detection false-alarm of existing SAR image object detection method higher and locate inaccurate deficiency, propose a kind of object detection method based on the classification of High Resolution SAR image directory.Image directory classification is on the basis of original image piecemeal, according in the subimage that piecemeal obtains comprise the difference of information, under sub-image assigns to different catalogue classifications.The Advantages found of catalog classification exists: (1) removes redundant information.After image carries out catalog classification, only need carry out target detection to the subimage comprising target, decrease and need data volume to be processed, ensure that the real-time of detection algorithm; (2) decrease the impact of complex background on target detection, the specific aim of target detection is strengthened, and the result obtained is more accurate.

Technical scheme of the present invention first carries out overlap partition to image, obtain the subimage that size is unified, then carry out cataloguing to all subimages, the type namely whether comprising target according to subimage and comprise target carries out catalog classification to it, obtains the catalogue that comprises all subimages after image directory.The last subimage comprising target of selecting from catalogue carries out target detection, obtains interested region (ROI), revises obtain final testing result to ROI.Below the present embodiment is described in detail.Fig. 1 is the process flow diagram of the High Resolution SAR image object detection method that the embodiment of the present invention provides.With reference to figure 1, the method mainly comprises the following steps:

Step 10, carries out overlap partition to the original SAR image of input, obtains multiple subimage.

Reference formula 1,

In formula 1, m, n are respectively the length of original SAR image and wide, and piecemeal is of a size of x × x, and p% is Duplication, and the subimage obtained is designated as S _i, i=1,2 ..., N, N are the number obtaining multiple subimage.

Step 20, the feature extracting training sample is trained sorter, and carry out catalog classification with the sorter after training to multiple subimage, the classification of catalog classification comprises and comprises target and do not comprise target.

First, determine catalogue classification number: the target type number that can be obtained comprising in SAR image by prior imformation is M, the corresponding two kinds of catalogue classifications of each target type, namely comprise this target and do not comprise this target, so total catalogue classification has 2M, and each catalogue classification is designated as C _q, q ∈ 2M, wherein C ₁, C ₂..., C _mfor comprising the catalogue category label of target.

Secondly, select training sample and extract proper vector: respectively for 2M catalogue classification random selecting training sample.Weber partial descriptions feature (WLD) is extracted as proper vector to each training sample, is designated as and sorter is trained.For the pixel in W × W window, WLD feature calculation is as formula 2:

$\mathrm{WLD}\left(w_c\right)=\arctan \left[\underset{j=0}{\overset{W^2-1}{\mathrm{\&Sigma;}}}\frac{w_j-w_c}{w_c}\right]+\arctan 2\left(\frac{{\mathrm{\&mu;}}_l-{\mathrm{\&mu;}}_r}{{\mathrm{\&mu;}}_d-{\mathrm{\&mu;}}_u}\right)---\left(2\right)$

In formula 2, w _cfor the center pixel in window, w _jfor the pixel of the jth in window, μ _l, μ _r, μ _d, μ _ube respectively the average of the average of all pixels in center pixel left, the average of all pixels in right, all pixels of average and top of all pixels in below.By using WLD feature, not only containing azimuth information, and embodying the otherness of pixel and its neighborhood territory pixel, target can be characterized more accurately.

Finally, extract the WLD feature of subimage, utilize the sorter trained to carry out catalog classification to it.After catalog classification terminates, the subimage not belonging to arbitrary classification is given up.

Step 30, acquisition catalogue classification are the rarity feature of weighting of the pixel of the subimage comprising target, Threshold segmentation is carried out to the subimage comprising target, Morphological scale-space is carried out to the bianry image after segmentation, obtains the area-of-interest (Region OfInterest) that catalogue classification is the subimage comprising target.

In catalogue classification, the subimage comprising target is label is C ₁, C ₂..., C _msubimage.

For piece image, the number of object pixel is less relative to the number of background pixel, and object pixel normally flocks together instead of the distribution that disperses in the picture, the gray-scale value of target is greater than the gray-scale value of background simultaneously.According to above-mentioned analysis, calculate the rarity feature of weighting of pixel, be designated as U.After obtaining the rare value of weighting of each pixel, carry out Threshold segmentation to image, setting threshold value is Th _u, by following formula to Image Segmentation Using, the image after segmentation is designated as BW, as formula 3:

$\mathrm{BW}=\left\{\begin{array}{c}1,U_k\&GreaterEqual;{\mathrm{Th}}_u\\ 0,U_k\&GreaterEqual;{\mathrm{Th}}_u\end{array}\right.---\left(3\right)$

In formula 3, U _kfor the rare value of weighting of a kth pixel.Carrying out the process such as morphology opening and closing operation, corrosion expansion to splitting the bianry image BW obtained, obtaining the two-value template of accurate characterization interesting image regions.In this template, pixel value is the region of interest ROI in the corresponding original image in region of 1.

Step 40, revises ROI, obtains object detection results.

The ROI number obtained is designated as L, and t ROI is ROI _t.

Judge ROI _s, s ∈ L and ROI _t, whether t ∈ L is connected region, wherein, and s ≠ t.If ROI _swith ROI _tfor connected region, then they are merged, obtain a new ROI.Repeat this step, until traveled through all ROI.

On such scheme basis, in step 30, the calculating of the rarity feature of weighting is as formula 4:

Calculate the space distribution weights of a kth pixel, be designated as ω _k, k=m _sub* n _sub, ω _kcalculate as formula 4:

${\mathrm{\&omega;}}_k=\sqrt{{(I_k-I_0)}^2+{(I_k-I_0)}^2}+\frac{\underset{j=1}{\overset{\mathrm{num}}{\mathrm{\&Sigma;}}}\sqrt{{(I_k-I_j)}^2+{(I_k-I_j)}^2}}{\mathrm{num}}---\left(4\right)$

In formula 4, m _sub, n _subbe respectively subimage long and wide, I _kfor the gray-scale value of a subimage kth pixel, I ₀for the pixel value of subimage center pixel, num be in subimage with I _kthe number of equal pixel.

Calculate the rare value of a kth pixel, be designated as u _k, k=m _sub* n _sub, u _kcalculate as formula 5:

$u_k=\left\{\begin{array}{cc}255,& \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k}\&GreaterEqual;255\\ \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k},& \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k}<255\end{array}\right.---\left(5\right)$

In formula 5, m _sub, n _subbe respectively subimage long and wide, un _kfor the number of times that the pixel of kth in subimage occurs.

Calculate the rare value of weighting of a kth pixel, be designated as U _k, k ∈ m*n, U _kcalculate as formula 6:

U _k＝ω _k*u _k(6)

The present invention compared with prior art tool has the following advantages:

1, detection method reduces the redundant information of image, shortens detection time, improves the execution efficiency of algorithm.

The present invention processes from forwarding to the process of global image subimage, introducing image directory is classified, under subimage unified for size is categorized into different catalogues, priori is utilized to carry out filtering to the subimage not comprising target, greatly reduce the data volume that need process, ensure that the efficiency that algorithm performs.

2, object detection results is accurate, and false alarm rate reduces.

The present invention carries out overlap partition to image, High Resolution SAR image is divided into the unified subimage of size, and relatively uniform concerning background each subimage, carry out target detection to image on this basis, the testing result obtained is more accurate.

3, object detection method highly versatile.

The image directory sorting technique that the present invention introduces can be divided into corresponding catalogue classification by sub-image, artificially can specify and need target classification to be processed, and the subimage taken out wherein carries out target detection.Be applicable to the High Resolution SAR image of various complex scene.

Lift specific embodiment below and illustrate effect of the present invention.

Fig. 1 is the process flow diagram of the High Resolution SAR image object detection that the embodiment of the present invention provides;

Fig. 2 is the original SAR image that emulation experiment of the present invention uses;

Fig. 3 is the result using CFAR algorithm image to be carried out to target detection;

Fig. 4 is the present invention's image block schematic diagram used;

Fig. 5 is the μ that the present invention uses _l, μ _r, μ _d, μ _ucomprise the particular location schematic diagram of pixel;

Fig. 6 is antithetical phrase image directory of the present invention sorted result figure, C ₁class is that meadow comprises target, C ₂class is that meadow is not containing target, C ₄class is that field is not containing target, C ₆class is that forest is not containing target, C ₈class is that lake is not containing target;

Fig. 7 is that the present invention is to class C ₁the subimage comprising target carries out the result of target detection.

With reference to Fig. 1, the specific embodiment of the present invention is as follows:

Step 1, carry out overlap partition to the SAR image of input, block is of a size of x × x, and Duplication people is for being set as p%, and the subimage obtained is S _i, i=1,2 ..., N, N are subimage sum.In the present embodiment, x=200, p=25.

Step 2, catalog classification is carried out to subimage:

(2a) the weber partial descriptions feature (WLD) of calculation training sample is trained SVM classifier as the proper vector of token image.For the pixel in W × W window, the computing formula of WLD is as follows:

$\mathrm{WLD}\left(w_c\right)=\arctan \left[\underset{j=0}{\overset{W^2-1}{\mathrm{\&Sigma;}}}\frac{w_j-w_c}{w_c}\right]+\arctan 2\left(\frac{{\mathrm{\&mu;}}_l-{\mathrm{\&mu;}}_r}{{\mathrm{\&mu;}}_d-{\mathrm{\&mu;}}_u}\right)$

In the present embodiment, High Resolution SAR image scene classification M=4, is respectively meadow, field, forest and lake.Catalogue classification is 2*M=8, is respectively meadow and comprises target (C ₁), meadow is containing target (C ₂), field comprises target (C ₃), field is containing target (C ₄), forest comprises target (C ₅), forest is containing target (C ₆), lake comprises target (C ₇) and lake not containing target (C ₈).

(2b) subimage S is extracted _iwLD feature input SVM, obtain S _igeneric label.After classification terminates, the subimage not belonging to arbitrary classification is given up.After classification terminates, C ₁class 1 width subimage, C ₂class comprises 12 width subimages, C ₃, C ₅, C ₇class comprises 0 width subimage, C ₄class comprises 6 width subimages, C ₆class comprises 9 width subimages, C ₈class comprises 5 width subimages.

Step 3, selects class C according to priori ₁subimage under catalogue.Calculate the rare value of weighting of sub-image pixels, be designated as:

$u_k=\left\{\begin{array}{cc}255,& \frac{{200}^2}{{\mathrm{un}}_k}\&GreaterEqual;255\\ \frac{{200}^2}{{\mathrm{un}}_k},& \frac{{200}^2}{{\mathrm{un}}_k}<255\end{array}\right.$

${\mathrm{\&omega;}}_k=\sqrt{{(I_k-I_0)}^2+{(I_k-I_0)}^2}+\frac{\underset{j=1}{\overset{\mathrm{num}}{\mathrm{\&Sigma;}}}\sqrt{{(I_k-I_j)}^2+{(I_k-I_j)}^2}}{\mathrm{num}}$

U _k＝ω _k*u _k

Wherein, un _kfor the number of times that the pixel of kth in subimage occurs, I _kfor the gray-scale value of a subimage kth pixel, I ₀for the pixel value of subimage center pixel, num be in subimage with I _kthe number of equal pixel, k ∈ 200 ².Setting threshold value is Th _u=180, by following formula to Image Segmentation Using, the image after segmentation is designated as BW:

$\mathrm{BW}=\left\{\begin{array}{c}1,U_k\&GreaterEqual;{\mathrm{Th}}_u\\ 0,U_k\&GreaterEqual;{\mathrm{Th}}_u\end{array}\right.$

The process such as morphology opening and closing operation, corrosion expansion is carried out to BW, obtains the ROI template of accurate characterization image.

Step 4, revises ROI, obtains final object detection results:

(4a) hypothesis is after step 3, and the ROI number obtained is L, remembers that t ROI is ROI _t;

(4b) ROI is judged _s, s ∈ L and ROI _t, whether t ∈ L is connected region, wherein, and s ≠ t.If ROI _swith ROI _tfor connected region, then they are merged into a new connected region.Repeat this step, until traveled through all ROI.ROI blue rectangle frame table final in this experiment shows.

The above; be only the specific embodiment of the present invention, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; change can be expected easily or replace, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should described be as the criterion with the protection domain of claim.

Claims (7)

1. a High Resolution SAR Images object detection method, is characterized in that, comprising:

Overlap partition is carried out to the original SAR image of input, obtains multiple subimage;

The feature extracting training sample is trained sorter, carries out catalog classification with the sorter after training to multiple described subimage;

Obtain the rarity feature of weighting that catalogue classification is the pixel of the subimage comprising target, Threshold segmentation is carried out to the described subimage comprising target, Morphological scale-space is carried out to the bianry image after segmentation, obtains the region of interest ROI that catalogue classification is the subimage comprising target;

Described ROI is revised, obtains object detection results.

2. method according to claim 1, is characterized in that, the described original SAR image to input is carried out overlap partition and comprised:

According to formula 1, overlap partition is carried out to original SAR image;

Wherein, m, n are respectively the length of original SAR image and wide, and piecemeal is of a size of x × x, and p% is Duplication, and the subimage obtained is designated as S _i, i=1,2 ..., N, N are the number obtaining multiple subimage.

3. method according to claim 1, is characterized in that, the feature of described extraction training sample is trained sorter, carries out catalog classification, comprising with the sorter after training to multiple described subimage:

Determine catalogue classification number, the target type number wherein being obtained comprising in SAR image by prior imformation is M, the corresponding two kinds of catalogue classifications of each target type, namely comprise this target and do not comprise this target, so total catalogue classification has 2M, and each catalogue classification is designated as C _q, q ∈ 2M, wherein C ₁, C ₂..., C _mfor comprising the catalogue category label of target;

Extract proper vector for described 2M catalogue classification random selecting training sample respectively;

Extract the proper vector of described subimage, catalog classification is carried out to described subimage.

4. method according to claim 3, is characterized in that, describedly extracts proper vector for described 2M catalogue classification random selecting training sample respectively, comprising:

Weber partial descriptions feature WLD is extracted as proper vector to each training sample, is designated as and sorter is trained;

Accordingly, the proper vector of the described subimage of described extraction, catalog classification is carried out to described subimage, comprising:

Extract the WLD feature of described subimage, utilize the sorter trained to carry out catalog classification to described subimage.

5. method according to claim 1, is characterized in that, describedly carries out Threshold segmentation to the described subimage comprising target, comprising:

Carry out Threshold segmentation to the described subimage comprising target, setting threshold value is Th _u, by following formula to Image Segmentation Using, the image after segmentation is designated as BW, as formula 2:

$\mathrm{BW}=\left\{\begin{array}{cc}1,& U_k\&GreaterEqual;{\mathrm{Th}}_u\\ 0,& U_k\&GreaterEqual;{\mathrm{Th}}_u\end{array}\right.---\left(2\right)$

In formula 2, U _kfor the rare value of weighting of a kth pixel.

6. method according to claim 1, is characterized in that, the rarity feature of described weighting, comprising:

Calculate the space distribution weights of a kth pixel, be designated as ω ^k, k=m _sub* n _sub, ω ^kcalculate as formula 3:

${\mathrm{\&omega;}}_k=\sqrt{{(I_k-I_0)}^2+{(I_k-I_0)}^2}+\frac{\underset{j=1}{\overset{\mathrm{mun}}{\mathrm{\&Sigma;}}}\sqrt{{(I_k-I_j)}^2+{(I_k-I_j)}^2}}{\mathrm{num}}---\left(3\right)$

Wherein, in formula 3, m _sub, n _subbe respectively subimage long and wide, I _kfor the gray-scale value of a subimage kth pixel, I ₀for the pixel value of subimage center pixel, num be in subimage with I _kthe number of equal pixel;

Calculate the rare value of a kth pixel, be designated as u ^k, k=m _sub* n _sub, u ^kcalculate as formula 4:

$u_k=\left\{\begin{array}{cc}255,& \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k}\&GreaterEqual;255\\ \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k},& \frac{m_{\mathrm{sub}}*n_{\mathrm{sub}}}{{\mathrm{un}}_k}<255\end{array}\right.---\left(4\right)$

Wherein, in formula 4, m _sub, n _subbe respectively subimage long and wide, un ^kfor the number of times that the pixel of kth in subimage occurs;

Calculate the rare value of weighting of a kth pixel, be designated as U _k, k ∈ m*n, U _kcalculate as formula 5;

U _k＝ω _k*u _k(5)

7. method according to claim 1, is characterized in that, describedly carries out correction to described ROI and comprises:

The ROI number obtained is designated as L, and t ROI is ROI _t;

Judge ROI _s, s ∈ L and ROI _t, whether t ∈ L is connected region, wherein, and s ≠ t.If ROI _swith ROI _tfor connected region, then they are merged, obtain a new ROI; Repeat this step, until traveled through all ROI.