CN104754340A - Reconnaissance image compression method for unmanned aerial vehicle - Google Patents

Abstract

The invention discloses a reconnaissance image compression method for an unmanned aerial vehicle. On the premise that most of the target objects of the unmanned aerial vehicle are objects with appearance in regular shapes, for example, military vehicles and pillboxes, before compression is conducted, an interested reconnaissance target area and a background area of a reconnaissance image of the unmanned aerial vehicle have been completely separated without influencing each other, and an appropriate compression method can be freely selected for the two parts. According to the reconnaissance image compression method, different bit rates are separately selected for the interested area and the background area, and it is ensured that the interested reconnaissance target area is compressed according to the small compression ratio and the background area is compressed according to the large compression ratio. Meanwhile, the compression ratios can be freely selected according to the segmentation compression method, and the image quality of reconnaissance targets and backgrounds is freely regulated so that the compression requirements of the reconnaissance images of the unmanned aerial vehicle can be met. Meanwhile, the better the area recovery quality is, the larger the peak signal to noise ratio of the corresponding area is.

Description

A kind of unmanned plane reconnaissance image compression method

Technical field:

The present invention relates to a kind of unmanned plane reconnaissance image compression method, it belongs to unmanned plane and scouts field.

Background technology:

Why image can be because in the data message of image except effective information, also include a large amount of redundant informations by compression.In actual applications, people often only compare care to certain part wherein and note, and are not very interested to the content of remainder.That is, when processing image, because the object of user's subjectivity application is different, then the interest level of corresponding whole image zones of different is also different.But, still need the bulk information of regions of non-interest when obtaining useful information from image.Then when encoding, identical coded system equivalent processes can not be adopted to entire image, only can not transmit content of interest.So, different coded treatment can only be carried out to the area-of-interest of image and regions of non-interest, just may obtain best encoding efficiency.During actual treatment, people always wish that the content oneself be concerned about can obtain better compression effectiveness relative to remainder.

Image compression encoding based on area-of-interest (ROI) is study hotspot and the difficult point in current compressed encoding field.It can realize the compression quality of compression quality higher than image background regions of interesting image regions, can meet the requirement of people to image.Research based on interesting sets has achieved many valuable achievements in research, area-of-interest (ROI) coding standard particularly proposed in JPEG2000 standard, i.e. maximum shift method (Max-shift) and generic scaling based method (Generic-scaling-based), this has established solid foundation to the development based on interesting sets.

Research abroad about area-of-interest compression is more, mainly concentrates on the improvement of wavelet coefficient translation thought and kind of the method for two in standard is combined with different coding algorithm.The domestic research about encoding region of interest is also growing, but the overwhelming majority concentrates in the research of JPEG2000 standard, is scarcely out of swaddling-clothes.At present, mainly contain for the compression algorithm in interesting target region both at home and abroad:

1. two kinds of area-of-interest (ROI) encryption algorithm: Max-shift and Generic-scaling-based proposed in JPEG2000 standard are adopted.The former needs to carry out precision encoding to the shape of area-of-interest, adds the complexity of coding/decoding; The latter lacks the flexibility of definition area-of-interest any side-play amount, can not the contrast of flexible area-of-interest and background area reconstructed image quality.

2. in order to overcome the shortcoming of first two algorithm, the people such as Zhou Wang propose improved algorithm---the Bitplane-by-Bitplane Shift according to these two kinds of algorithm ideas, improved algorithm can realize the encoding region of interest to arbitrary shape, and the contrast of area-of-interest and background area reconstructed image quality can be regulated, but the popularization in essence still to Max-shift algorithm.

3. the people such as Lijie Liu in 2003 also been proposed a kind of new modified model translation algorithm (Significant BitplaneShift), it is except possessing the advantage of Bitplane-by-Bitplane Shift algorithm, can also according to the coding of different priority realization to multiple semi-cylindrical hills.

4. Claudio.M proposes the six kinds of image processing algorithms of extraction local maximum taked for obtaining area-of-interest according to the visual characteristic of human eye.Afterwards he supplement again the algorithm based on wavelet transformation, the algorithm based on discrete dct transform, based on the first Gaussian filtering algorithm of Laplace conversion and the algorithm of result of statistics 7 × 7 region templates coupling again.

5. Jerome M.Shapiro proposes Embedded Wavelet position encryption algorithm, mainly contain embedded Zerotree wavelet transform algorithm (EZW) and multistage manifold conjunction splitting algorithm (SPIHT), encoding stream after interesting target wavelet coefficient compression process is placed on the front end of whole compressed file, prior to the encoding stream after background area wavelet coefficient compression process.

Summary of the invention:

The invention provides a kind of unmanned plane reconnaissance image compression method, it can solve the contradiction between the Quality of recovery of target area and image compression rate, can meet the demand that unmanned plane is scouted to a great extent.

The present invention adopts following technical scheme: a kind of unmanned plane reconnaissance image compression method, it comprises the steps

Step 1, initialization

Calculate the value of n, initialization significant coefficient list LSP is empty list, belongs to set R by all _rOOT∪ R _lLnode (i, j) deposit in inessential coefficient list LIP, then by all belong to set R _rOOTnode (i, j) deposit in inessential aggregate list LIS, and regard category-A element as;

Step 2, sequence

2.1, the important function S of all nodes (i, j) in calculations list LIP _n(i, j)

If S _n(i, j)=1, then add in list LSP by node (i, j), exports the sign bit of its corresponding wavelet coefficient C (i, j) simultaneously;

2.2, all nodes (i, j) in list LIS are analyzed

If A. node (i, j) is a category-A element, calculate its important function value S _n(D (i, j)); If S _n(D (i, j))=1, then:

A. the important function value S of all nodes (k, l) in set of computations O (i, j) _n(k, l)

If S _n(k, l)=1, then deposit node (k, l) in list LSP, exports the sign bit of its corresponding wavelet coefficient C (k, l) simultaneously,

If S _n(k, l)=0, then deposit the afterbody to list LIP by node (k, l);

If b. gathering L (i, j) is an empty set, i.e. L (i, j) ≠ φ, just node (i, j) is stored in the afterbody of list LIS, and regards category-B element as, then proceed step B process; Otherwise just node (i, j) is deleted from list LIS;

If B. node (i, j) is a category-B element, calculate its important function value S _n(L (i, j)), if S _n(L (i, j))=1, just deposits afterbody to LIS by all nodes (k, l) belonging to set O (i, j) with category-A element, node (i, j) is shifted out from list LIS simultaneously;

Step 3, progressive

The absolute value of the wavelet coefficient C (i, j) in calculations list LSP corresponding to all nodes (i, j) | C (i, j) |, and its bit value of n-th is exported;

Step 4, threshold value upgrade

If n=n-1, going back to step 2 again carries out next stage coded scanning simultaneously.

Further, after completing compressed encoding, the co-ordinate position information of two-part compressed data information and area-of-interest is carried out volume frame according to certain data frame structure order, then these image information datas is sent to grounded receiving station by wireless channel.

The present invention has following beneficial effect: unmanned plane reconnaissance image compression method selecting of the present invention be a kind of segmentation compression method, before compression process, the spot region interested of unmanned plane reconnaissance image has fully achieved with background area and has been separated, be independent of each other between the two, therefore can the suitable compression method of unrestricted choice to these two parts.Unmanned plane reconnaissance image compression method of the present invention separately selects different bit rates to area-of-interest and background area, ensures spot region interested small reduction ratio compression, and background area is large compression ratio compression then.Meanwhile, this segmentation compression method also can the size of unrestricted choice compression ratio, achieves the picture quality of free adjustment aim and background.Also according to the requirement of device transmission bandwidth, can freely adjust the compression bit rate of spot and background area, meet the compression requirement of unmanned plane reconnaissance image.Better with time domain Quality of recovery, the Y-PSNR of its corresponding region is also larger.

Accompanying drawing illustrates:

Fig. 1 is for adding noisy image.

Fig. 2 is muting image.

Fig. 3 is closed curve.

Fig. 4 is for extracting area-of-interest.

Fig. 5 is the theory diagram based on region of interest regional partition compression method.

Fig. 6 is unmanned plane reconnaissance image compression method simulation result of the present invention.

Fig. 7 is the contrast with generic scaling based method and SPIHT method simulation result.

Fig. 8 is the process chart of unmanned plane reconnaissance image compression method.

Embodiment:

Below in conjunction with drawings and Examples, the present invention is described in further detail.

Unmanned plane reconnaissance image is that drone is aloft taken earthward, and the picture got, be also just similar to people on the aircraft of high-altitude flying, take pictures earthward.Concerning have man-machine, people directly can obtain relevant information from clapping the picture got; And concerning unmanned plane, to need the picture that photographs, by after air transmission to ground station, artificially therefrom to get relevant information.Therefore, for the picture that unmanned plane photographs, we are resent to ground, obtain information wherein by surface personnel after being processed again after needing to be compressed process.

Because unmanned plane during flying height is higher, speed is fast, and the airborne sensor angle of visual field is comparatively large, so object pixel is little and quantity large in unmanned plane reconnaissance image, comparatively other normal image are poor for in-frame correlation and spatial coherence, and frame-to-frame correlation and temporal correlation are not strong yet.Unmanned plane is in some application of military aspect, and be want to obtain some Military Informations from the picture photographed, simultaneously concerning entire image, people compare care a certain subregion wherein.In the image that a width is photographed by unmanned plane, can be generally that some military equipment objects contain Military Information, such as military vehicle, tank, pillbox etc.Therefore, for unmanned plane reconnaissance image, people can compare care military target part wherein.But due to this unmanned plane reconnaissance image be that unmanned plane is at high aerial photographing, it is very little that these military targets such as military vehicle, tank, pillbox can show in picture, that is, in unmanned plane reconnaissance image, the military surveillance target that people are concerned about only occupies the fraction region of entire image.

Concerning the unmanned plane reconnaissance image containing military information, military surveillance target is wherein smaller, only account for fraction region.Remaining extensive area is then the natural environment landscape facilities such as some mountains and rivers, road, trees, and people are not be concerned about very much to these contents, and these contents are general as reference information, the geographical position residing for follow-up judgement military surveillance target.During this kind of unmanned plane reconnaissance image of actual acquisition, after people wish decompression, spot part interested wherein can know identification, and remaining background parts then requires it is not very high, can have individual reference value.

Concerning this unmanned plane reconnaissance image, we wish to adopt a kind of compression processing method to it, and its spot part high-quality interested is reappeared, and the data volume of entire image also obtains certain minimizing simultaneously.Generally common compression method is compression entire image being carried out to equal extent substantially, due in unmanned plane reconnaissance image, spot region is wherein a region among a small circle, to ensure that this spot high-quality reappears, the compression ratio then selected is just very limited, is difficult to meet a large amount of data volume reducing image; To make the data volume of image be reduced in a large number, then the compression ratio selected will be comparatively large, is difficult to ensure that spot wherein can be reappeared by high-quality.Therefore, Ordinary Compression method is not suitable for unmanned plane reconnaissance image, can not meet the requirement of unmanned plane reconnaissance image.For this requirement, wish there is a kind of compression method for area-of-interest, when processing, small reduction ratio compression or even Lossless Compression are adopted to the spot region interested of unmanned plane reconnaissance image, ensure that the high-quality of spot interested reappears, then adopt large compression ratio to compress to its background area, the data volume of image is greatly reduced.

Generally speaking, General image is that extensive area wherein carries main information, and unmanned plane reconnaissance image is region among a small circle wherein carries main information.According to the compression requirement of unmanned plane reconnaissance image, adopt a kind of applicable compression method.Concerning unmanned plane reconnaissance image, require that, while the clear reconstruct of the spot of unmanned plane reconnaissance image, the data volume of view picture unmanned plane reconnaissance image have also been obtained and reduces greatly.

The principle that the present invention is based on region of interest regional partition compression method mainly comprises the following aspects:

One. the extraction of area-of-interest.The present invention is directed unmanned plane reconnaissance image, and for the spot of these unmanned plane reconnaissance images be all the military equipments such as military vehicle, pillbox, tank, these targets are all regular pattern composite figures, and the large regions natural background in image does not then possess systematicness.In the region of interesting extraction stage, need to utilize the interesting target object finding certain features in view picture unmanned plane reconnaissance image someway, after finding this object, determine that a pocket at object place is the area-of-interest of unmanned plane reconnaissance image, thus extract the spot region interested of unmanned plane reconnaissance image.

The extraction of area-of-interest mainly contains Roberts rim detection, Sobel rim detection, Prewitt rim detection, Canny rim detection etc.By comparative analysis, Selection utilization Canny edge detection operator of the present invention finds the edge contour curve of image object, then the systematicness feature of foundation spot figure, processes these edge contour curves, determines the spot of unmanned plane reconnaissance image.Classical Canny edge detection algorithm is with the smoothing filtering of Gaussian function.The object of image smoothing filtering technique improves signal to noise ratio, stress release treatment.But when making Gaussian smoothing to unmanned plane reconnaissance image, edge can be smoothed out as radio-frequency component, some low intensive edges also can be filtered simultaneously.Meanwhile, in Canny testing process, need high-low threshold value process, connect edge, therefore high-low threshold value choose also very important.Select the Canny edge detection algorithm slightly made improvements to the process of unmanned plane reconnaissance image, find the pattern edge profile information of unmanned plane reconnaissance image.

When 1. first making the disposal of gentle filter to original unmanned plane reconnaissance image, two-dimensional Gaussian function G (x, y) and first derivative G ' (x, y) thereof is selected to make convolution algorithm with original unmanned plane reconnaissance image f (x, y).In Practical Calculation process, to G (x, y) discretization, G (i, j) to be denoted as.

P(i,j)＝G(i,j)*f(i,j) (1)

Q(i,j)＝G′(i,j)*f(i,j) (2)

2. amplitude and the direction of gradient is asked.Gradient is asked to above-mentioned convolution results, then extracts x respectively, the one dimension gradient vector on y direction.

$\left[\begin{array}{c}{P}_x(x,y)\\ P_y(x,y)\end{array}\right]=\▿P(i,j)=\left[\begin{array}{c}\∂P/\∂x\\ \∂P/\∂y\end{array}\right]---\left(3\right)$

$\left[\begin{array}{c}{Q}_x(x,y)\\ Q_y(x,y)\end{array}\right]=\▿Q(i,j)=\left[\begin{array}{c}\∂Q/\∂x\\ \∂Q/\∂y\end{array}\right]---\left(4\right)$

The gradient direction θ (i, j) at point (i, j) place can be obtained by formula (3).

$\mathrm{\θ}(i,j)=\arctan \left(\frac{P_x(i,j)}{P_y(i,j)}\right)---\left(5\right)$

The gradient magnitude M (i, j) at point (i, j) place is obtained by formula (4) and formula (5).

M(i,j)＝[Q _y·sin(θ(i,j))+Q _x·cos(θ(i,j))] (6)

3. non-maxima suppression is carried out to gradient magnitude.

4. dual threshold process and edge conjunction.Choosing of high-low threshold value is very important, directly determines the number of the marginal information detected and the continuity at edge.Local variance reflection be the localized variation situation of image, for obtaining the overall variation situation of image, utilize the average variance of image as one of high threshold parameter.If only consider the grey scale change of image, and do not consider the gray value of image self pixel, rim detection effect is also bad.So the average gray value of image also can be used as one of high threshold parameter.

T _h＝k×(E _ave/F _ave) (7)

$E_{\mathrm{ave}}=\frac{1}{L_w\×L_h}\underset{m=1}{\overset{L_w\×L_h}{\mathrm{\Σ}}}{(f_m(i,j)-F_{\mathrm{ave}})}^2---\left(8\right)$

$F_{\mathrm{ave}}=\frac{1}{L_w\×L_h}\underset{m=1}{\overset{L_w\×L_h}{\mathrm{\Σ}}}{f}_m(i,j)---\left(9\right)$

Wherein, T _hfor required high threshold, f _m(i, j) is the gray value of pixel (i, j), E _ave, F _ave, L _wand L _hbe respectively the average variance of image, average gray value, image width and height.Parameter k is positioned at (0.2,0.5), Low threshold choose T _l=0.4 × T _h.

After determining high-low threshold value, according to these two threshold value connection layouts as edge contour curve.

With the Canny algorithm after improvement to military vehicle image, in original image, also add the salt-pepper noise that noise density is 0.02 simultaneously.Shown in following Fig. 1 and Fig. 2 of testing result.By Fig. 1 and 2 result, the Canny algorithm of improvement can be found out, while maintenance traditional C anny algorithm advantage, avoid the loss at some weak edges, and it be in restraint speckle, also have some improvement.When selecting high-low threshold value, according to average and the variance determination high-low threshold value of image, this avoid the limitation of artificial setting, automatically can calculate high-low threshold value.

With improving the Canny algorithm of front and back to unmanned plane reconnaissance image, Lena image and pillbox image procossing, in original image, also add the salt-pepper noise that noise density is 0.02 simultaneously.First the edge contour curve of unmanned plane reconnaissance image is found, after finding the edge contour curve of figure by the above-mentioned Canny algorithm slightly made improvements, eight neighborhood search method is utilized to judge the closure of these edge contour curves found, finally judge whether the Confined outline curve found possesses systematicness feature again, finally determine that the region at this regular closed curve place is the area-of-interest of view picture unmanned plane reconnaissance image, then determine closed curve.Close, as the term suggests, namely close, there is no breach.In the present invention, be the region wanting first to find regular military target place, the then target and background of separate picture, then implement follow-up compression process.After the process of the above-mentioned Canny algorithm slightly made improvements, get the edge contour curve of unmanned plane reconnaissance image.In order to find regular figure region wherein, need first to find certain contour curve, more whether go to judge that this contour curve surrounds be a regular figure.Therefore need from all edge contour curves found, only retain the contour curve that some are closed.

During experiment process, subsequent treatment is done to the pattern edge contour curve image of unmanned plane reconnaissance image, find closed curve wherein according to eight realm informations determined a little.Its specific algorithm is according to being: the p (x, y) that 1. sets up an office is the starting point A of image border ₀; 2. by 3 × 3 neighborhoods of counterclockwise scanning current pixel point, until run into the pixel consistent with fixed point gray value, Here it is new image border point A _n; 3. as the subsequent border point A scanned _nbe initial fixed point (closed outline), or subsequent border point A _nwithout the subsequent border point (non-close profile) be connected, terminate scanning; 4. record the closed outline obtained, give up non-close profile, then search for other edge contours until image terminates.

But when actual treatment, because unmanned plane reconnaissance image is the picture of unmanned plane at high aerial photographing, the region shared by military surveillance target is wherein also smaller, and like this when doing contour detecting process, its profile possible is not give prominence to very much.Likely can cause the actual contour curve found and not exclusively close, but military surveillance target wherein also possesses certain size dimension, therefore, when searching for closed curve, to above-mentioned searching algorithm according to having done some flexible process.When the terminal of certain curve and starting point not overlapping time, when meeting following formula, closed when tentatively can determine this curve.

$\left\{\begin{array}{c}d\≤10\\ \frac{\mathrm{sum}}{d}\≥7\end{array}\right.---\left(10\right)$

In formula, d is the distance between terminal and starting point, and distance here calculates according to pixel coordinate, and sum is all pixel sums from this curve of origin-to-destination.The scope restriction 10 of its middle distance d is a numerical value of preliminary election in actual process.Although the regional compare shared by military surveillance target of unmanned plane reconnaissance image is little, the edge contour of this military target also possesses certain size.In actual process, for it sets a size scope, namely the size range of its length and width or radius is [20,80].According to size range, the minimum perimeter polygon of known military target contour of object is 80, and remove maximum vacancy distance 10, then the sum in above formula is at least 70, and therefore the ratio of sum and d is at least 7.

Result is illustrated in fig. 3 shown below.

Estimation of structural regularity again.After obtaining closed curve, just by judging that the systematicness of these closed curves determines regular figure wherein, thus area-of-interest can be found.According to the geometric invariant moment character of figure, have certain proportionate relationship between the girth of known regular figure and area, therefore we can judge to detect whether the closed outline obtained is regular figure according to this characteristic.

The moment characteristics main geometric properties of image-region.If piece image is f (x, y), then (p+q) rank geometric moment of f (x, y) is:

M _pq＝∫∫x ^py ^qf(x,y)dxdy,p,q＝0,1,2,…(11)

The then centre of form of this object:

$\stackrel{\‾}{x}=\frac{M_{10}}{M_{00}},\stackrel{\‾}{y}=\frac{M_{01}}{M_{00}}---\left(12\right)$

Center square:

${\mathrm{\μ}}_{\mathrm{pq}}={\∫\∫(x-\stackrel{\‾}{x})}^p{(y-\stackrel{\‾}{y})}^{q}f(x,y)\mathrm{dxdy},p,q=\mathrm{0,1,2},\·\·\·---\left(13\right)$

If process image is discrete picture, then its geometric moment:

$M_{\mathrm{pq}}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{i}^p{j}^{q}f(i,j)---\left(14\right)$

${\mathrm{\μ}}_{\mathrm{pq}}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{(i-\stackrel{\‾}{x})}^p{(j-\stackrel{\‾}{y})}^{q}f(i,j)---\left(15\right)$

For bianry image, when (i, j) is for interior of articles point, f (i, j)=1, then put as f (i, j)=0 other.Wherein M ₀₀represent the area of object in bianry image.

$M_{00}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}f(i,j)---\left(16\right)$

Namely all pixel sums that closed outline graphics package contains are the area s of figure.

Because the geometric invariant moment character of figure, exists certain proportion relation between the girth of regular figure and area, the ratio formula that this patent adopts is l ²/ s, wherein l is girth, namely closed outline figure neighboring pixel and, s is graphics area, i.e. all pixel sums of containing of closed outline graphics package.From top to bottom, from left to right, point by point scanning is carried out to the enclosed curve regions of reading in, the girth of inner all pixel sum m, the l presentation graphics of record edge contour pixel point sum l and closed area, the area of s=(l+m) presentation graphic.Then ratio computing formula is:

$r=\frac{l^2}{l+m}---\left(17\right)$

During actual treatment, the closed curve that may read not is close completely, may there is breach, before its girth of calculating and area, needs first to be filled by this vacancy mouth, makes it to close completely.Although military surveillance target wherein only occupies pocket in view picture unmanned plane reconnaissance image, the edge contour of this military target also possesses certain size.This patent for be the unmanned plane reconnaissance image comprising military surveillance target (tank, military vehicle, pillbox), during this kind of unmanned plane reconnaissance image of actual treatment, for military target wherein sets a size scope, namely the size range of its length and width or radius is [20,80].Under this size range condition, concerning rectangle, its peripheral extent [80,320], areal extent [400,6400]; Concerning circle, its peripheral extent [40 π, 160 π], areal extent [400 π, 6400 π].Both comprehensive, the scope [80,160 π] of girth, the scope [400,6400 π] of area.Therefore, before the proportionate relationship judging girth and area, formula (18) need first be met

$\left\{\begin{array}{c}160\mathrm{\π}\≥l\≥80\\ 6400\mathrm{\π}\≥l+m\≥400\end{array}\right.---\left(18\right)$

In experimentation, can judge whether the closed area detected is regular figure according to the size of r value.In this patent, user is interested in the military surveillance target in unmanned plane reconnaissance image, wishes the region finding military surveillance target place, and this region is exactly the area-of-interest of unmanned plane reconnaissance image.When 16≤r≤25, infer that processed closed outline is a rectangular shape, i.e. corresponding military vehicle or tank.As r=4 π, infer processed closed outline be one round-shaped, i.e. corresponding pillbox.

The closed outline figure of Fig. 3 is judged, judges whether these figures are regular figure, and retention discipline closed outline, abandons irregular closed outline.After obtaining the regular domain in image, according to the coordinate position of all pixels of selected regular figure, obtain the pixel (minx formed with minimum abscissa and minimum ordinate, miny) pixel (maxx, maxy) and with maximum abscissa and maximum ordinate formed.Determine that the rectangular area that formed by upper summit (minx, miny) and lower summit (maxx, maxy) is as the area-of-interest needing to extract, as the military target region interested of the unmanned plane reconnaissance image required for us.

Result as shown in Figure 4.As shown above, Fig. 4 (a) represents in military vehicle image, finally determines military vehicle target area interested wherein, and Fig. 4 (b) represents in pillbox image, finally determine pillbox target area interested wherein, namely determine area-of-interest.

Then interested region is separated the military surveillance target area interested that have found in unmanned plane reconnaissance image with background area, determine the area-of-interest of a rectangular area as image at this military target place, namely specify that area-of-interest and the background area of image, just need the area-of-interest of original image (ROI) and background area (BG) to be separated below.

Two, the separation of image.After finding spot, determine the area-of-interest R of zonule rectangular extent as unmanned plane reconnaissance image at spot place, write the co-ordinate position information of area-of-interest simultaneously.Judge whether each pixel of unmanned plane reconnaissance image belongs to region R, original unmanned plane reconnaissance image is separated into area-of-interest (ROI) and background area (BG) two parts.

The separation method that the present invention selects is a kind of separation method based on regional concept.

Carry out separated region based on regional concept namely some pixels possessing similar quality to be communicated with, realize the separation of each zones of different of image.In experiment processing procedure, the specific descriptions of this separation are as follows:

(1) area-of-interest is isolated.To the sequential scanning that original unmanned plane reconnaissance image carries out from left to right, from top to bottom, judge each pixel (x of image, y) (wherein R is the rectangle area-of-interest obtained according to the positional information of the spot detected whether to belong to region R, the upper summit of this rectangle is (minx, miny), lower summit is (maxx, maxy).

To scanning each pixel (x, y) got, whether discriminatory analysis belongs to R.If (x, y) ∈ is R, then keep the gray value f (x, y) of its script; Otherwise its gray value f (x, y) is then made to be 0.Be scanned up to picture to terminate always, finally show processed image.Namely the subgraph of interesting image regions is only comprised.

$\left\{\begin{array}{c}f(x,y)=f(x,y),(x,y)\∈R\\ f(x,y)=0,(x,y)\∉R\end{array}\right.---\left(19\right)$

(2) background area is isolated.In like manner according to step process above, just can keep the script gray value of background area pixels point, and the gray value of area-of-interest pixel set to 0.But because area-of-interest scope is less, namely need the pixel limited amount set to 0, said method slightly can be made improvements, reduce some amounts of calculation.

In this process, to original unmanned plane reconnaissance image, only can scan area-of-interest, namely start from left to right from point (minx, miny), scan according to this from top to bottom.When scanning point (maxx, miny), just consider line feed, namely start from point (minx, miny+1), the rest may be inferred again, until till scanning point (maxx, maxy).Due in scanning process, the pixel got is all in interesting image regions, then the gray value f (x, y) of the pixel only these need got sets to 0.Image after last Graphics Processing, namely only comprises the subgraph of background area.

$\left\{\begin{array}{c}f(x,y)=f(x,y),(x,y)\∉R\\ f(x,y)=0,(x,y)\∈R\end{array}\right.---\left(20\right)$

Through above separating step, the area-of-interest (ROI) and background area (BG) that achieve unmanned plane reconnaissance image are separated.In the subgraph of background area, give up interested area information, remain background area information.In area-of-interest subgraph, give up background area information, remained interested area information.Image, after N level wavelet transform process, has 1 low-frequency band, 3N high frequency band.The wavelet coefficient of its higher frequency band is smaller, which represent the information (detail coefficients) of image detail part; The wavelet coefficient of low-frequency band is larger, which represent the information (silhouette coefficient) of image outline part.Although wavelet transformation does not reach itself do compression process to image, the wavelet coefficient after conversion also exists Some features, and these features make image be beneficial to compression.

In this patent, EZW algorithm and spiht algorithm are two kinds of conventional at present embedded Coding Compression Algorithm based on wavelet transformation, and EZW and SPIHT all employ successive approximation to quantification method, namely the threshold value of each scanning gets the half of last threshold value, and their initialization procedure and thinning process are also similar.

But the structure of SPIHT to zero tree is more efficient compared with EZW.Illustrate: suppose that c (i, j) is important, its descendants (comprising child) is unessential, in EZW, first time scanning gives symbol P c (i, j), four class trees are formed to descendants thereafter, represents with PZZZZ five symbols; In SPIHT, c (i, j) to move on to LSP from LIP and exports one 1 by first time scanning, to (i, the j) in LIS relatively after export one 0, namely illustrate with two symbol tables.Therefore, spiht algorithm can utilize the feature of wavelet coefficient better, has better coding efficiency.

Three, the compression of image.Unmanned plane reconnaissance image is after dividing processing, and unmanned plane reconnaissance image has been divided into two subgraphs, and we can carry out the compression of different compression ratio respectively to these two parts.Adopt low compression ratio compression or even Lossless Compression to area-of-interest, then adopt high compression ratio to compress to background area, we also can select different compression methods to compress background area according to the requirement of transmission bandwidth simultaneously.So both ensure that the high efficiency of compression of image, also ensure that the high-quality of area-of-interest reappears.

By analysis above, this patent selects spiht algorithm to do compression process to the spot region of unmanned plane reconnaissance image and background area.Again according to this two-part difference compression requirement, select these two subgraphs of SPIHT process of different bit rates.The compression ratio that different bit rates is corresponding different, bit rate is larger, and compression ratio is less.

After operational processes above, unmanned plane reconnaissance image is originally separated into the area-of-interest (ROI) only comprising spot information and the background area (BG) only comprising background information, traditional spiht algorithm scanning encoding all from the beginning to the end, can cause the waste of code stream like this.Comprise the feature of information according to this two parts subgraph, suitable improvement has been done to spiht algorithm.Basic thought: to the subgraph only comprising ROI coefficient, encoding those is not positioned at point and the set of BG.To the subgraph only comprising BG coefficient, encoding those is not positioned at point and the set of ROI.

Only to comprise the area-of-interest subgraph of military surveillance target information, describe compression handling process, the background area subgraph only comprising background information can be analogized.

1, initialization

Calculate the value of n, initialization significant coefficient list LSP is empty list, belongs to set R by all _rOOT∪ R _lLnode (i, j) deposit in inessential coefficient list LIP, then by all belong to set R _rOOTnode (i, j) deposit in inessential aggregate list LIS, and regard category-A element as.

2, sort

(1) the important function S of all nodes (i, j) in calculations list LIP _n(i, j).

If S _n(i, j)=1, then add in list LSP by node (i, j), exports the sign bit of its corresponding wavelet coefficient C (i, j) simultaneously.

(2) all nodes (i, j) in list LIS are analyzed.

If A. node (i, j) is a category-A element, calculate its important function value S _n(D (i, j)).If S _n(D (i, j))=1, then:

A. the important function value S of all nodes (k, l) in set of computations O (i, j) _n(k, l).

If S _n(k, l)=1, then deposit node (k, l) in list LSP, exports the sign bit of its corresponding wavelet coefficient C (k, l) simultaneously.

If S _n(k, l)=0, then deposit the afterbody to list LIP by node (k, l).

If b. gathering L (i, j) is an empty set, i.e. L (i, j) ≠ φ, just node (i, j) is stored in the afterbody of list LIS, and regards category-B element as, then proceed step B process; Otherwise just node (i, j) is deleted from list LIS.

If B. node (i, j) is a category-B element, calculate its important function value S _n(L (i, j)).If S _n(L (i, j))=1, just deposits afterbody to LIS by all nodes (k, l) belonging to set O (i, j) with category-A element, node (i, j) is shifted out from list LIS simultaneously.

3, progressive

The absolute value of the wavelet coefficient C (i, j) in calculations list LSP corresponding to all nodes (i, j) | C (i, j) |, and its bit value of n-th is exported.

4, threshold value upgrades

If n=n-1, going back to step 2 again carries out next stage coded scanning simultaneously.In experiment processing procedure, in order to make area-of-interest quality be better than background area quality, the bit rate (R of area-of-interest _roi) be higher than the bit rate (R of background area _bg).After completing compressed encoding, the co-ordinate position information of two-part compressed data information and area-of-interest is carried out volume frame according to certain data frame structure order, then these image information datas is sent to grounded receiving station by wireless channel.After receiving image data information, first the information received is carried out to SPIHT decoding and the wavelet inverse transformation of corresponding coding bit rate, decode procedure is the inverse process of coding, again according to the co-ordinate position information of the area-of-interest received, area-of-interest subgraph correspondence after decompression is inlayed the relative position of the background area subgraph to decompression, thus reconstruct original image.

Four, transmission or storage.After compression, the co-ordinate position information combination of two-part compressed data information and area-of-interest is sent to grounded receiving station by wireless channel together.

Five, the reconstruct of image.After receiving image data information, first carry out decompression operation, then according to the coordinate position of area-of-interest, synthesis area-of-interest subgraph and background area subgraph, reconstruct original unmanned plane reconnaissance image.

Under normal circumstances, we understand the treatment effect of the first l-G simulation tests of some emulation tools of Selection utilization once algorithm, analyze the difference between acquired results and desired effects, and then debugging algorithm program, the performance of gradual perfection algorithm, to realize better treatment effect, reaches our desired value.

This patent is under MATLAB programmed environment, carries out artificial debugging checking to the segmentation compression algorithm based on area-of-interest newly carried.In experimentation, select unmanned plane reconnaissance image as experimental image.But because original unmanned plane reconnaissance image is larger, display is not easy in screen, easy to use then in order in experimentation, we only have chosen one piece of region wherein to carry out experiment simulation process, and the resolution therefore processing picture in experiment is not little.

Maxshift method in JPEG2000 standard is after a wavelet transform process, the wavelet coefficient relevant to interesting target is scaled up or on move, make the bit-planes at these wavelet coefficient places higher than the wavelet coefficient of image background part.General SPIHT is Embedded Wavelet position coded system, the encoding stream after interesting target wavelet coefficient compression process is placed on the front end of whole compressed file, prior to the encoding stream after background area wavelet coefficient compression process.They move the wavelet coefficient relevant to region of interest on being all, make bit-planes at them higher than the bit-planes at the coefficient of background area, then follow-up embedded encoded in, the wavelet coefficient of ROI is allowed to be positioned at before the wavelet conversion coefficient of BG, so, ROI will be encoded and refinement prior to BG.

The segmentation compression algorithm based on area-of-interest selected by this patent is the contradiction in order to solve between the Quality of recovery of target area and image compression rate, also reaches the real-time Transmission of image simultaneously.

In experiment simulation process, select different bit rates to decompress to two parts of image, namely correspond to different compression ratios.In order to ensure the Quality of recovery of the Postprocessing technique quality in spot small area higher than background area, the bit rate that this patent have selected area-of-interest is greater than the bit rate of background area.

After this patent also can analyze use said method from quantitative angle, the spot region interested of unmanned plane reconnaissance image and the Quality of recovery of background area, repeat no more here.

In the figure 7, figure (a) is the gray-scale map of original military vehicle image, figure (b) is the result of generic scaling based method simulation process, figure (c) is the simulation process result of spiht algorithm, and figure (d) and figure (e) is the result selecting the process of this patent segmentation compression method.Figure (b) data volume: the data volume of 15108, figure (c): 15062.

Figure (d) be ensure unmanned plane reconnaissance image spot region reconstruction quality and figure (b) and scheme (c) suitable while, figure (b) is then greater than to the decrement of background area and schemes the decrement of (c) middle background area.Like this, when the spot of equal extent reconstruct unmanned plane reconnaissance image, select the data volume after the process of this patent disjoining pressure compression method, less than the data volume selected after generic scaling based method and spiht algorithm process, namely decrease data volume.Figure (d) data volume: 5148.

Figure (e) allows the reconstruction quality in unmanned plane reconnaissance image spot region higher than the reconstruction quality of figure (b) and figure (c), namely the ROI region bit rate of figure (e) is selected to be greater than the ROI region bit rate of figure (d), meanwhile strengthen the decrement of background area again, the BG region bit rate of namely scheming (e) is less than the BG region bit rate of figure (d).Like this, clearly can see military surveillance target, make again the data volume of image be reduced.Figure (e) data volume: 4826.

Integrated comparative, this patent is a kind of disjoining pressure compression method, and before compression process, the spot region interested of unmanned plane reconnaissance image has fully achieved with background area and has been separated, be independent of each other between the two, therefore can the suitable compression method of unrestricted choice to these two parts.This patent separately selects different bit rates to area-of-interest and background area, ensures spot region interested small reduction ratio compression, and background area is large compression ratio compression then.Meanwhile, this disjoining pressure compression method also can the size of unrestricted choice compression ratio, achieves the picture quality freely adjusting image object and background.Also according to the requirement of device transmission bandwidth, can freely adjust the compression bit rate of spot and background area, meet the compression requirement of unmanned plane reconnaissance image.Better with time domain Quality of recovery, the Y-PSNR of its corresponding region is also larger.

The above is only the preferred embodiment of the present invention, it should be pointed out that for those skilled in the art, can also make some improvement under the premise without departing from the principles of the invention, and these improvement also should be considered as protection scope of the present invention.

Claims (2)

1. a unmanned plane reconnaissance image compression method, is characterized in that: comprise the steps

Step 1, initialization

Calculate the value of n, initialization significant coefficient list LSP is empty list, belongs to set R by all _rOOT∪ R _lLnode (i, j) deposit in inessential coefficient list LIP, then by all belong to set R _rOOTnode (i, j) deposit in inessential aggregate list LIS, and regard category-A element as;

Step 2, sequence

2.1, the important function S of all nodes (i, j) in calculations list LIP _n(i, j)

If S _n(i, j)=1, then add in list LSP by node (i, j), exports the sign bit of its corresponding wavelet coefficient C (i, j) simultaneously;

2.2, all nodes (i, j) in list LIS are analyzed

If A. node (i, j) is a category-A element, calculate its important function value S _n(D (i, j)); If S _n(D (i, j))=1, then:

A. the important function value S of all nodes (k, l) in set of computations O (i, j) _n(k, l)

If S _n(k, l)=1, then deposit node (k, l) in list LSP, exports the sign bit of its corresponding wavelet coefficient C (k, l) simultaneously,

If S _n(k, l)=0, then deposit the afterbody to list LIP by node (k, l);

If b. gathering L (i, j) is an empty set, i.e. L (i, j) ≠ φ, just node (i, j) is stored in the afterbody of list LIS, and regards category-B element as, then proceed step B process; Otherwise just node (i, j) is deleted from list LIS;

If B. node (i, j) is a category-B element, calculate its important function value S _n(L (i, j)), if S _n(L (i, j))=1, just deposits afterbody to LIS by all nodes (k, l) belonging to set O (i, j) with category-A element, node (i, j) is shifted out from list LIS simultaneously;

Step 3, progressive

The absolute value of the wavelet coefficient C (i, j) in calculations list LSP corresponding to all nodes (i, j) | C (i, j) |, and its bit value of n-th is exported;

Step 4, threshold value upgrade

If n=n-1, going back to step 2 again carries out next stage coded scanning simultaneously.

2. unmanned plane reconnaissance image compression method as claimed in claim 1, it is characterized in that: after completing compressed encoding, the co-ordinate position information of two-part compressed data information and area-of-interest is carried out volume frame according to certain data frame structure order, then these image information datas is sent to grounded receiving station by wireless channel.