CN101594533A - A kind of compression method that is applicable to sequence images of unmanned aerial vehicle - Google Patents

Abstract

The invention discloses a kind of compression method that is applicable to sequence images of unmanned aerial vehicle, be specially:, sequence images of unmanned aerial vehicle is made Intelligent Dynamic coded frame group divide according to airborne auxiliary data; In the frame group, respectively first frame and tail frame are made arrive separately the high accuracy fast motion estimation of present frame, obtain severally to match point, according to corresponding matching relationship, try to achieve the projective transformation matrix that first frame and tail frame arrive present frame separately, and then acquisition predicted picture and residual error; After the first frame of same frame group and the residual image of all the other frames are merged into single image, use the JPEG2000 compression, utilize its global optimization to block algorithm and obtain the high-performance compression result.The present invention takes all factors into consideration the sequence images of unmanned aerial vehicle characteristics, makes full use of the characteristic of airborne supplementary and JPEG2000 compression algorithm, has that algorithm complex is low, good compression property, compression bit rate can strict controls, hardware is easy to characteristics such as realization.

Description

A kind of compression method that is applicable to sequence images of unmanned aerial vehicle

Technical field

The present invention relates to image data processing technique, particularly relate to a kind of compression method that is applicable to sequence images of unmanned aerial vehicle.

Background technology

Development along with imaging technique, the image resolution ratio that unmanned plane obtains improves constantly, image data amount sharply increases, and the bandwidth of communication channel is very limited, therefore must carry out effective Real Time Compression significantly to reduce its data volume to the view data of obtaining.On the other hand, unmanned plane during flying speed is fast, and flying height is higher, relative normal image, and unmanned plane image resolution ratio height, spatial coherence is relatively poor, and temporal correlation is also obviously different with ordinary video.Conventional images and video compression algorithm and system more or less are short of to some extent at aspects such as algorithm complex, compression performance, compression speed, response time, code stream controls, can't satisfy the demand of engineering practice fully, MJPEG2000 algorithm (list of references: ISO/IEC 15444-3Motion-JPEG2000 (JPEG2000Part 3) as current use, 2002.), it uses the JPEG2000 algorithm to do compression in the frame, can obtain higher recovery effect, but, cause effectively to improve compression ratio owing to do not remove the interframe redundancy.And H.264 compression algorithm (referring to document: ITU Draft Recommendation H.264, ISO/IEC Draft InternationalStandard 14496-10.JVT-G050r1) methods such as many macro-block partition mode estimations, the estimation of multi-parameter frame have been adopted in, though make it obtain good compression effectiveness, but its algorithm complex is too high, deficiency is arranged, to the image and the incompatibility of unmanned plane one class simultaneously on Rate Control.Therefore, special-purpose sequence images of unmanned aerial vehicle compression method and the real-time implementation device of development has the important engineering meaning.

Summary of the invention

The object of the present invention is to provide that a kind of complexity is low, compression signal to noise ratio height, the sequence images of unmanned aerial vehicle compression method of strict control compression bit rate and the realtime graphic compression set of realizing this method.

A kind of compression method that is applicable to sequence images of unmanned aerial vehicle, carry out according to following steps:

Step S1 utilizes airborne auxiliary data to sequence frame image division frame group, and first frame and other frame in each frame group have doubling of the image zone;

Each frame group of step S2 is compressed in the following manner:

S201 calculates the overlapping region of present frame and first frame and tail frame respectively;

S202 chooses at least five match block and selects the territory respectively in the overlapping region of first frame and present frame in the overlapping region of tail frame and present frame, and selects to choose match block in the territory at these;

S203 uses the match block of first frame and tail frame correspondence to do matching operation respectively in present frame, further tries to achieve the projective transformation matrix that first frame and tail frame arrive present frame separately, as prediction matrix;

S204 utilizes first frame, tail frame and each self-corresponding prediction matrix thereof to estimate to obtain the predictive frame of present frame, asks for the residual error of present frame and predictive frame;

S205 asks for the residual error of all frames beyond the first frame in the frame group according to step S201～S204;

S206 is spliced into a big figure of single width with the residual image of first frame and other frame, adopts the JPEG2000 standard that the big figure of single width is done reduced overall.

The present invention is relative, and prior art has the following advantages:

1. utilize on-board data to divide the frame group, utilize the H matrix to carry out the predictive coding of bidirectional frame group, improved in the speed that improves estimation and the precision of motion compensation, improved code efficiency, reduced computation complexity simultaneously with respect to existing algorithm;

2. adopt the residual image of first two field picture and present frame to carry out the image splicing in the frame group and then use JPEG2000 and do reduced overall, utilize the PCRD-OPT algorithm in the JPEG2000 compression standard, optimize code stream and distribute, the strict compression bit rate of controlling;

Adopt the present invention, available low-down hardware cost is realized the Real Time Compression up to the investigation image of 1024 * 1024 * 100 frame/seconds, 32 times fixed than the condition of compression under Y-PSNR (PSNR) reach more than the 40dB, all be better than having now the unmanned plane image compression system at aspects such as compression performance, processing speed and cost performances.

Description of drawings

Fig. 1 is the compression set structured flowchart;

Fig. 2 is the compression algorithm flow chart;

Fig. 3 is that match block is selected area schematic, and Fig. 3 (a) is a kind of selection mode, and Fig. 3 (b) is another kind of selection mode.

Embodiment

For more well-known elaboration the purpose, technical solutions and advantages of the present invention, the present invention is further detailed explanation below in conjunction with drawings and the specific embodiments.

Fig. 1 is the structured flowchart of sequence images of unmanned aerial vehicle compression set.

102,104 is the input/output interface of device circuit, this device is set to the LVDS interface, by other interface equipments, as pci card with the LVDS interface, can realize and being connected of PC, sequence images of unmanned aerial vehicle and the corresponding airborne auxiliary data that collects is sent to compressive plate by 102, sequential conversion by field programmable gate array (FPGA) 103 realizes the mutual of data and digital signal processor (DSP) 107.Also can make the encoding code stream after the compression pass back to PC or other equipment by interface 104.

DSP107 obtains after the data, the order that beginning is set according to algorithm, progressively finish Intelligent Dynamic coded frame group divide, empty unite the quick high accuracy estimation frequently and multi-mode H matrix resolves, based on the operations such as bidirectional frame group predictive coding of H matrix, afterwards frame spelling map interlinking picture is delivered to ADV202 by FPGA103.

ADV202 is the video released of ADI (ADI) and the complete monolithic PJEG2000 compression and decompression solution of rest image, and its space ultra high efficiency recursive filtering (SURF) technology that has patent makes it to have low-power consumption and wavelet compression cheaply; Support is up to 6 grades 9/7 and 5/3 wavelet transformation.Utilize it that image is done the JPEG2000 compression, finish to optimize and block Rate Control.

After the frame group compression dateout is returned to DSP by FPGA, DSP with compression result and H matrix composite coding after, transmit FPGA, export by 104.

FPGA103 finishes the logic control of data interaction between DSP and ADV202, and corresponding sequential conversion.

The intermediate data storage that needs to carry out in the DSP operation is all finished by SDRAM (106).FLASH (108) is the required configurator of saved system then, comprises the DSP program, FPGA configuration file, ADV202 firmware etc.

Fig. 2 is the flow chart of sequence image compression algorithm of the present invention:

S1: divide the frame group, first frame and other frame in each frame group have doubling of the image zone.

The frame group is divided and can be searched by content frame, exist the frame of local same number of frames content to represent to have the overlapping region, these frames that comprise the overlapping region constitute a framing group, but this content search method amount of calculation is big, efficient is low, therefore the invention provides another kind of division methods fast, this method is mainly according to the airborne auxiliary data search, and is specific as follows:

Sequence image and airborne auxiliary data are input to after the compressibility, and DSP calculates the series of frames that the overlapping region is arranged with first two field picture at first according to airborne auxiliary data information, constitute a frame group, and concrete grammar is:

Suppose to be learnt by airborne auxiliary data that the unmanned plane during flying height is h, the angle of visual field is θ _x* θ _y, the sequence image frame frequency is f, the course instantaneous velocity when taking the i two field picture is v _i, consider that the velocity amplitude error that auxiliary data provides is δ _v, then can calculate the displacement s of aircraft between i frame and the i+j frame _{I, j}For:

$s_{i,j}=(\frac{v_{i+1}+v_i}{2}-3{\mathrm{\δ}}_v)\·1/f+(\frac{v_{i+2}+v_{i+1}}{2}-3{\mathrm{\δ}}_v)\·1/f+...+(\frac{v_{i+j}+v_{i+j-1}}{2}-3{\mathrm{\δ}}_v)\·1/f$

$=(\frac{v_{i+j}+v_i}{2}+v_{i+1}+v_{i+2}+...+v_{i+j-1}-3j{\mathrm{\δ}}_v)\·1/f$ (1.1)

And the shooting area of single-frame images is: $L_x=2\·h\·\tan \frac{{\mathrm{\θ}}_x}{2},$ $L_y=2\·h\·\tan \frac{{\mathrm{\θ}}_y}{2}$

If s _{I, j}＜L _xSet up, show that image will have the overlapping region.

So satisfy:

$\left\{\begin{array}{c}(\frac{v_{i+j}+v_i}{2}+v_{i+1}+v_{i+2}+...+v_{i+j-1}-3j{\mathrm{\&delta;}}_v)\&CenterDot;1/f<2\&CenterDot;h\&CenterDot;\tan \frac{{\mathrm{\&theta;}}_x}{2}\\ (\frac{v_{i+j+1}+v_i}{2}+v_{i+1}+v_{i+2}+...+v_{i+j-1}+v_{i+j}-3(j+1){\mathrm{\&delta;}}_v)\&CenterDot;1/f\&GreaterEqual;2\&CenterDot;h\&CenterDot;\tan \frac{{\mathrm{\&theta;}}_x}{2}\end{array}\right.---\left(1.2\right)$

I frame to the i+j frame between all frames the overlapping region is arranged in the course, in like manner can obtain the frame that the overlapping region is upwards arranged on the side is i frame to the i+k frame.Get n=min (j, k).Then there is i frame to the i+n frame to constitute a frame group, frame headed by the i frame wherein, the i+n frame is the tail frame.Again with frame headed by the i+n frame, a framing that satisfies following formula will constitute a new frame group afterwards.A burst of group tail frame has guaranteed to have only some residual frame and a whole frame in the frame group compression as the method for the first frame of next frame group before this, has improved compression ratio.With current unmanned plane typical case flight parameter is example, and setting flying height h is 1000m, course speed v=250m/s, other is 0m/s to speed, velocity error 5m/s, frame f=25fps frequently, look shooting under the scenery of video camera to ground, angle of visual field θ is 3 °, and can try to achieve n is 4.Suppose the unmanned plane smooth flight, it is a frame group that the 0-4 frame is then arranged, and the 4-8 frame is a frame group, by that analogy.

These have the frame in doubling of the image zone to constitute a frame group.Frame headed by first frame, last frame are the tail frame, and DSP moves these frames and note to the SDRAM temporary simultaneously in auxiliary data, and S202 uses in order to next step operation.

If, n=0, promptly representing does not have the overlapping region between the two continuous frames image, and at this moment, separately as a frame group, promptly first frame is the tail frame with the i frame, and do to compress in the frame to frame and get final product this moment, needn't carry out estimation.

Each frame group of S2 is compressed in the following manner.

S201: in each frame group, calculate the overlapping region of present frame and first frame and tail frame respectively.

DSP at first utilizes:

$s_{i,j}=(\frac{v_{i+1}+v_i}{2}-3{\mathrm{\&delta;}}_v)\&CenterDot;1/f+(\frac{v_{i+2}+v_{i+1}}{2}-3{\mathrm{\&delta;}}_v)\&CenterDot;1/f+...+(\frac{v_{i+j}+v_{i+j-1}}{2}-3{\mathrm{\&delta;}}_v)\&CenterDot;1/f$

$=(\frac{v_{i+j}+v_i}{2}+v_{i+1}+v_{i+2}+...+v_{i+j-1}-3j{\mathrm{\&delta;}}_v)\&CenterDot;1/f$ (1.3)

(v wherein _iThe course speed of the i frame that provides for on-board data, δ _vBe velocity error)

Calculate the course displacement Lx of the first relatively frame of present frame (i+j frame) (i frame) _{I, j}=s _{I, j}In like manner can be in the hope of the lateral displacement Ly of the first relatively frame of present frame (i+j frame) (i frame) _{I, j}, it is (Lx that the present frame (i+j frame) and the overlapping region of first frame (i frame) are then arranged _{I, j}, Ly _{I, j}) and (L _x, L _y) between the imaging region of rectangular area on first frame.Wherein $L_x=2\&CenterDot;h\&CenterDot;\tan \frac{{\mathrm{\&theta;}}_x}{2},$ $L_y=2\&CenterDot;h\&CenterDot;\tan \frac{{\mathrm{\&theta;}}_y}{2}.$ The imaging resolution of supposing unmanned plane is N * M (may be 256 * 256,512 * 512,1024 * 1024), the then top left corner apex (Lx of overlapping region _{I, j}, Ly _{I, j}) coordinate on first frame is $X_{i,j}=\frac{x_{i,j}}{L_x}\&times;N,$ $Y_{i,j}=\frac{y_{i,j}}{L_y}\&times;M,$ And then (X on the frame headed by the overlapping region _{I, j}, Y _{I, j}) and (N, M) zone between.According to same way as, obtain the overlapping region of present frame and tail frame.

After obtaining the overlapping region, enter step S202.

S202: the match block of choosing at least five respectively in the overlapping region of the overlapping region of first frame and present frame and tail frame and present frame is selected the territory, selects to choose match block in territories at these;

Existing match block choosing method is mainly done full search to frame, and the zone that obtains having maximum local variance is as match block, and the computation complexity height when many match block, many match block clustering phenomena occurs in addition easily, can not guarantee that match block is dispersed representative.The present invention is in the overlapping region, and is best representative for matching effect is had, the tangible match block of use characteristic.At first according to selected center up and down, overlapping region totally 5 match block select territory A ₁, A ₂, A ₃, A ₄, A ₅, match block size N _a* M _aFolded area size of apparent weight and shape and decide generally should have: N/12≤N _a≤ N/3, M/12≤M _a≤ M/3 is made as 30 * 80 here, shown in Fig. 3 (a).

Because the overlapping region size of image is indefinite, under different overlapping situations, match block selects the distribution in territory overlapping situation can occur, and concrete condition is shown in Fig. 3 (b), under the less situation in overlapping region, can make up and down in the middle of five zones have overlapping distribution.

Select to select in the territory a slice to have the N ' * M ' fritter of maximum local variance in match block as match block, the size of match block selects to need to consider overlapping region size and shape and picture material, General N ' and M ' all can be taken as 4,8,16,64, be set to 8 * 64 size here.

In first frame, the top left corner apex coordinate be (a, the local variance of N ' b) * M ' sized images is defined as:

${\mathrm{Var}}_{(a,b)}=\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{(x_{(a+i,b+j)}-{\stackrel{\&OverBar;}{x}}_{(a,b)})}^2---\left(1.4\right)$

X wherein _{(a+i, b+j)}Denotation coordination be (a+i, the pixel value of some b+j),

Headed by in the frame top left corner apex coordinate be (a, the mean value of the pixel value of N ' b) * M ' sized images.If match block is selected territory A ₁The left upper apex coordinate be (a ₀, b ₀), use a=a ₀, a ₀+ 1 ..., a ₀+ N _a-N '; B=b ₀, b ₀+ 1 ..., b ₀+ M _a-M ' travels through, and just can try to achieve and select territory A ₁In have the match block of maximum local variance.But directly calculate, the double counting amount is very big, and is too big to computational resource and requirement computing time, for this reason, this traversal calculating provided improvement.

${\stackrel{\&OverBar;}{x}}_{(a,b)}=\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+i,b+j)}\right)$

$\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{(x_{(a+i,b+j)}-{\stackrel{\&OverBar;}{x}}_{(a,b)})}^2=\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+i,b+j)}^2-N^{\&prime;}{M}^{\&prime;}{\stackrel{\&OverBar;}{x}}_{(a,b)}^2$

$\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+1+i,b+j)}^2=\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+i,b+j)}^2-\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a,b+j)}^2+\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+N^{\&prime;}+1,b+j)}^2$

${\stackrel{\&OverBar;}{x}}_{(a+1,b)}=\frac{1}{N^{\&prime;}{M}^{\&prime;}}[\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+i,b+j)}\right)-\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a,b+j)}\right)+\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+N^{\&prime;}+1,b+j)}\right)]$

And then have:

${\mathrm{Var}}_{(a,b)}=\stackrel{\&OverBar;}{x_{(a,b)}^2}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}{\stackrel{\&OverBar;}{x_{(a,b)}}}^2---\left(1.5\right)$

${\mathrm{Var}}_{(a+1,b)}=\stackrel{\&OverBar;}{x_{(a,b)}^2}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a,b+j)}^2+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+N^{\&prime;}+1,b+j)}^2$

$-\frac{1}{N^{\&prime;}{M}^{\&prime;}}{({\stackrel{\&OverBar;}{x}}_{(a,b)}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a,b+j)}\right)+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+N^{\&prime;}+1,b+j)}\right))}^2$ (1.6)

${\mathrm{Var}}_{(a,b+1)}=\stackrel{\&OverBar;}{x_{(a,b)}^2}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+i,b)}^2+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+i,b+M^{\&prime;}+1)}^2$

$-\frac{1}{N^{\&prime;}{M}^{\&prime;}}{({\stackrel{\&OverBar;}{x}}_{(a,b)}-\frac{1}{N^{\&prime;}{m}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+i,b)}\right)+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+i,b+M^{\&prime;}+1)}\right))}^2$ (1.7)

Owing in the process of traversal, calculate Var _{(a+1, b)}And Var _{(a, b+1)}Shi Fuyong

With

, compare common computational methods, this method can reduce computing time greatly, can constitute the fast algorithm of asking for maximum local variance, obtains having in five zones the N ' * M ' image fritter of maximum local variance.Can be made as B _S1, B _S2, B _S3, B _S4, B _S5In like manner, can obtain five match block B of tail frame _E1, B _E2, B _E3, B _E4, B _E5After obtaining match block, then enter step S204.

S203: use the match block of first frame and tail frame in present frame, to do matching operation respectively, obtain two prediction matrixs of present frame.

With five match block B in the first frame _S1, B _S2, B _S3, B _S4, B _S5, in present frame, make coupling respectively and calculate.With B _S1Be example, in the overlapping region of first frame and present frame, use full search method, calculate present frame and B _S1The normalized crosscorrelation coefficient.The segment that obtains minimum normalized crosscorrelation coefficient is B _S1Match map.

If B _S1Top left corner apex in first frame is (a _S1, b _S1), for ease of statement, (a on it _S1+ i, b _S1+ j) pixel value is designated as x _{I, j}(a+i, pixel value b+j) still is designated as y in the present frame _{A+i, b+j}, use in addition

The expression present frame in (a is the pixel average in the N ' * M ' sized images zone of top left corner apex b), then goes average normalized crosscorrelation coefficient to be defined as:

$\mathrm{NPROD}(a,b)=\frac{\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}(y_{a+i,b+j}-{\stackrel{\&OverBar;}{y}}_{a,b})\&times;(x_{i,j}-\stackrel{\&OverBar;}{x})}{{\left[\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{(y_{a+i,b+j}-{\stackrel{\&OverBar;}{y}}_{a,b})}^2\right]}^{\frac{1}{2}}{\left[\underset{j=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{k=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{(x_{i,j}-\stackrel{\&OverBar;}{x})}^2\right]}^{\frac{1}{2}}}---\left(1.8\right)$

Wherein: $\stackrel{\&OverBar;}{x}=\frac{1}{N^{\&prime;}{M}^{\&prime;}}\&CenterDot;\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{i,j},$ So have: $\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}(x_{i,j}-\stackrel{\&OverBar;}{x})=0$

$\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}(y_{a+i,b+j}-{\stackrel{\&OverBar;}{y}}_{a,b})\&times;(x_{i,j}-\stackrel{\&OverBar;}{x})=\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{y}_{a+i,b+j}(x_{i,j}-\stackrel{\&OverBar;}{x})-{\stackrel{\&OverBar;}{y}}_{a,b}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}(x_{i,j}-\stackrel{\&OverBar;}{x})$

$=\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{y}_{a+i,b+i}(x_{i,j}-\stackrel{\&OverBar;}{x})$ (1.9)

So have $\mathrm{NPROD}(a,b)=\frac{\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{y}_{a+i,b+j}\&times;(x_{i,j}-\stackrel{\&OverBar;}{x})}{{\left[\underset{i=1}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{(y_{a+i,b+j}-{\stackrel{\&OverBar;}{y}}_{a,b})}^2\right]}^{\frac{1}{2}}{\left[\underset{j=1}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{(x_{i,j}-\stackrel{\&OverBar;}{x})}^2\right]}^{\frac{1}{2}}}$ (1.10)

It is very big directly to use related algorithm to carry out the images match amount of calculation.Employing can effectively improve calculated performance based on the fast correlation algorithm of FFT.

If NPROD (a _M1, b _M1(NPROD (a, b)), establish this correspondence segment is B to)=min ₀, get B ₀Center point coordinate be (x _C1, y _C1), B _S1Center point coordinate be (x _S1, y _S1), to other match block, try to achieve the match point coordinate in the corresponding first two field picture of center point coordinate of match block in the present frame respectively $(x_{c2},y_{c2})\&LeftRightArrow;(x_{s2},y_{s2}),$ $(x_{\mathrm{<figure-callout\; id="3"\; label="c"\; filenames="A2009100629290019H1.png"\; state="\{\{state\}\}">c</figure-callout>}3},y_{c3})\&LeftRightArrow;({\mathrm{<figure-callout\; id="3"\; label="x\; s"\; filenames="A2009100629290019H1.png"\; state="\{\{state\}\}">x</figure-callout>}}_s,y_{s3}),(x_{c4},y_{c4})\&LeftRightArrow;(x_{s4},y_{s4}),(x_{c5},y_{c5})\&LeftRightArrow;(x_{s5},y_{s5}),$ Each center match point to ask for difference vector in the overlapping region, is established (x _{C_su}, y _{C_su}), (x _{S_su}, y _{S_su}) be respectively the top left corner apex of the overlapping region of present frame and first frame, then each difference vector is:

(i＝1，2，3，4，5)。Consider the generation of mistake coupling, adopt following method to determine:

Get

Be the average of 5 difference vector, promptly ${\stackrel{\&OverBar;}{V}}_s=\underset{i=1}{\overset{5}{\mathrm{\&Sigma;}}}{V}_{\mathrm{si}}/5;$ In 5 difference vector, find out with the difference maximum of vector average, promptly

, k=1 ..., 5.Setting fault-tolerant threshold value is Δ, can be made as 3 pixels, thinks difference vector V _SiWith average

Difference greater than the expression matching error of Δ, belong to the mistake coupling, otherwise think that coupling is correct.

Promptly have if

, k=1 ..., 5, (Δ is fault-tolerant threshold value, can be made as 3 pixels), then think 5 match block all correctly the coupling; Otherwise if having , then remove V _Sk, in 4 remaining vectors, ask for new average $\stackrel{\&OverBar;}{V_s^{\&prime;}}=(\underset{i=1}{\overset{5}{\mathrm{\&Sigma;}}}{V}_{\mathrm{si}}-V_{\mathrm{sk}})/4,$ Carry out once missing match check again.Duplicate test is mated vector until determining all correct couplings with mistake.If the quantity of correct coupling vector is less than 4, then should be in first frame choose reasonable match block again, such as between the match block that obtains correct coupling, reselecting match block, do new matching operation, until the correct matching result that obtains (containing 4) more than 4.

Select four correct couplings right, the coordinate matching relationship of establishing match point is: $(x_{\mathrm{cu}},y_{\mathrm{cu}})\&LeftRightArrow;(x_{\mathrm{su}},y_{\mathrm{su}}),u=1,...,4,$ With its substitution projective transformation equation: $z_{\mathrm{cs}}\left[\begin{array}{c}{x}_{\mathrm{cu}}\\ y_{\mathrm{cu}}\\ 1\end{array}\right]=H_s\left[\begin{array}{c}{x}_{\mathrm{su}}\\ y_{\mathrm{su}}\\ 1\end{array}\right],$ U=1 ..., in 4, and then try to achieve z _Cs, H _s, in the solution procedure, make z _Cs=1, in like manner, can obtain the z of the homograph of the relative tail frame of present frame _Se, H _eH _sAnd H _eFrame and tail frame be to the projective transformation matrix of present frame respectively, that is first frame and tail frame are to the prediction matrix of present frame.

S204: utilize first frame, tail frame and each self-corresponding prediction matrix thereof to obtain the predictive frame of present frame, ask for the residual error of present frame and predictive frame.

If (x _s, y _s) the overlapping region O of first frame and present frame in the first frame of expression _CsThe coordinate of pixel, (x _Cs, y _Cs) interior first frame of expression present frame and current frame image overlapping region O _CsThe coordinate of future position of pixel, thereby utilize the projective transformation equation: $z_{\mathrm{cs}}\left[\begin{array}{c}{x}_{\mathrm{cs}}\\ y_{\mathrm{cs}}\\ 1\end{array}\right]=H_s\left[\begin{array}{c}{x}_s\\ y_s\\ 1\end{array}\right],$ (x _s, y _s) ∈ O _Cs, can obtain the part predicted picture F that present frame is come by first frame prediction _CsObtain the part predicted picture F that present frame is come by the prediction of tail frame with quadrat method _CeTo F _CsAnd F _CeDo suitable amalgamation, promptly obtain the bidirectional predictive picture F ' of present frame _cUse present frame F _cWith F ' _cDo and ask the difference processing then to obtain residual image D _c=F _c-F ' _c

S205: the residual error of asking for all frames beyond the first frame in the frame group according to step S201～S204

S206: the residual image of first frame and other frame is spliced into a big figure of single width, adopts the JPEG2000 standard to do reduced overall output to splicing big figure.

In the frame group coding, the residual image of first frame and all the other frames is done the image splicing in the treatable scope of ADV202, make the entire frame group form the big figure of single width.The image size of ADV202 single treatment is restricted to the 1.048M pixel, and Breadth Maximum is restricted to 4096 pixels, if the frame stitching image is excessive, can do suitable cutting.Characteristics at JPEG2000 block encoding, piecemeal distribution code stream, determined the compression parameters of JPEG2000 compression, selected after coded block size, small echo type and the progression, in the splicing of image, also should do corresponding adjustment, methods such as merging are cut in employing, the content constraints of image that makes single frames avoids adjacent interframe the situation that the encoding block branch is striden two frames to occur in the integral multiple scope of coded block size.After obtaining single image, use the JPEG2000 compression, utilize the flow-optimized control technology of blocking of outstanding global title that the PCRT-OPT algorithm provides in the JPEG2000 standard, code stream is distributed according to the image information global optimization, compression bit rate is controlled in strictness simultaneously, obtains the compressed image (S402) of a frame group.The tail frame becomes the first frame of a next frame group in the next one processing cycle, so can not consider the tail frame in the compression of frame group.Frame group compression result is integrated with the H matrix information (S403) of each present frame, and the compression result that promptly can be used as this frame group is exported.

More than the present invention is had been described in detail, but the present invention is not limited to the above embodiments, in the scope that does not break away from technical conceive of the present invention, can certainly carry out various improvement and change.

Claims (5)

1, a kind of compression method that is applicable to sequence images of unmanned aerial vehicle, carry out according to following steps:

Step S1 utilizes airborne auxiliary data to sequence frame image division frame group, and first frame and other frame in each frame group have doubling of the image zone;

Each frame group of step S2 is compressed in the following manner:

S201 calculates the overlapping region of present frame and first frame and tail frame respectively;

S202 chooses at least five match block and selects the territory respectively in the overlapping region of first frame and present frame in the overlapping region of tail frame and present frame, and selects to choose match block in the territory at these;

S203 uses the match block of first frame and tail frame correspondence to do matching operation respectively in present frame, further tries to achieve the projective transformation matrix that first frame and tail frame arrive present frame separately, as prediction matrix;

S204 utilizes first frame, tail frame and each self-corresponding prediction matrix thereof to estimate to obtain the predictive frame of present frame, asks for the residual error of present frame and predictive frame;

S205 asks for the residual error of all frames beyond the first frame in the frame group according to step S201～S204;

S206 is spliced into a big figure of single width with the residual image of first frame and other frame, adopts the JPEG2000 standard that the big figure of single width is done reduced overall.

2, the compression method that is applicable to sequence images of unmanned aerial vehicle according to claim 1 is characterized in that, described step S1 judges in the following manner the overlapping region between two frames: the displacement of calculating aircraft between i frame and the i+j frame $s_{i,j}=(\frac{v_{i+j}+v_i}{2}+v_{i+1}+v_{i+2}+...+v_{i+j-1}-3j{\mathrm{\&delta;}}_v)\&CenterDot;1/f,$ If $s_{i,j}\&le;2\&CenterDot;h\&CenterDot;\tan \frac{{\mathrm{\&theta;}}_x}{2},$ Show that then i frame and i+j frame have the overlapping region; Wherein, v _pCourse instantaneous velocity when the p two field picture is taken in expression, p=i, i+1 ..., i+j, δ _vBe the velocity amplitude error, f is the picture frame frequency, and h is the unmanned plane during flying height, θ _xFor taking the course angle of visual field.

3, the compression method that is applicable to sequence images of unmanned aerial vehicle according to claim 1 and 2 is characterized in that, the tail frame of present frame group is the first frame of next frame group.

4, the compression method that is applicable to sequence images of unmanned aerial vehicle according to claim 1, it is characterized in that, described step S202 chooses match block in the following manner in selecting the territory: choose cycling among windows, with this window application in selecting the territory, select have maximum local variance video in window as match block, the computational methods of local variance are:

The top left corner apex coordinate is (a, the local variance of video in window b)

${\mathrm{Var}}_{(a,b)}=\stackrel{\&OverBar;}{x_{(a,b)}^2}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}{\stackrel{\&OverBar;}{x_{(a,b)}}}^2,$

The top left corner apex coordinate is (a+1, the local variance of video in window b)

${\mathrm{Var}}_{(a+1,b)}=\stackrel{\&OverBar;}{x_{(a,b)}^2}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a,b+j)}^2+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+N^{\&prime;}+1,b+j)}^2$

$-\frac{1}{N^{\&prime;}{M}^{\&prime;}}{({\stackrel{\&OverBar;}{x}}_{(a,b)}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a,b+j)}\right)+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{M^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+N^{\&prime;}+1,b+j)}\right))}^2$

The top left corner apex coordinate is (a, the local variance of video in window b+1)

${\mathrm{Var}}_{(a,b+1)}=\stackrel{\&OverBar;}{x_{(a,b)}^2}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+i,b)}^2+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{j=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}{x}_{(a+i,b+M^{\&prime;}+1)}^2$

$-\frac{1}{N^{\&prime;}{M}^{\&prime;}}{({\stackrel{\&OverBar;}{x}}_{(a,b)}-\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+i,b)}\right)+\frac{1}{N^{\&prime;}{M}^{\&prime;}}\underset{i=0}{\overset{N^{\&prime;}-1}{\mathrm{\&Sigma;}}}\left(x_{(a+i,b+M^{\&prime;}+1)}\right))}^2$

Wherein, N ' is the transverse width of video in window, and M ' is vertical width of video in window, the bottom right mark expression point coordinates of x (), and whole implication is represented this point coordinates corresponding pixel value,

For the top left corner apex coordinate be (a, the pixel value mean value of video in window b),

For the top left corner apex coordinate is (a, the mean value of video in window pixel value b) square.

5, the compression method that is applicable to sequence images of unmanned aerial vehicle according to claim 1 is characterized in that, described step S204 determines the projective transformation matrix of first frame to present frame in the following manner:

It is right to choose coupling $(x_{\mathrm{cu}},y_{\mathrm{cu}})\&LeftRightArrow;(x_{\mathrm{su}},y_{\mathrm{su}}),$ U=1 ..., 4, (x _Su, y _Su) headed by the centre coordinate of u match block of frame, (x _Cu, y _Cu) be the match map centre coordinate of u match block of corresponding first frame in the present frame, make up the projective transformation equation $z_{\mathrm{cs}}\left[\begin{array}{c}{x}_{\mathrm{cu}}\\ y_{\mathrm{cu}}\\ 1\end{array}\right]=H_s\left[\begin{array}{c}{x}_{\mathrm{su}}\\ y_{\mathrm{su}}\\ 1\end{array}\right],$ Find the solution and obtain coordinate homogeneous partial differential variable z _CsWith the projective transformation matrix H of first frame to present frame _s

Images