CN108810534A

CN108810534A - Method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things

Info

Publication number: CN108810534A
Application number: CN201810596841.1A
Authority: CN
Inventors: 石翠萍; 靳展; 何鹏; 朱恒军; 李静辉; 那与晶; 潘悦
Original assignee: Qiqihar University
Current assignee: Qiqihar University
Priority date: 2018-06-11
Filing date: 2018-06-11
Publication date: 2018-11-13
Anticipated expiration: 2038-06-11
Also published as: CN108810534B

Abstract

Method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things, the present invention relates to method for compressing image.The purpose of the present invention is to solve existing SPIHT methods seldom to consider edge blurry or ringing effect caused by the missing by high-frequency information, can not retain more details in image, lead to the problem that code efficiency is low.Process is：One, the image block after being divided；Two, the optimum prediction direction of image block after being divided；Three, directional interpolation filter coefficient is weighted by calculating, directional interpolation is weighted to fractional samples value, obtains interpolation image block；Four, using optimum prediction direction, the wavelet transformation promoted based on direction is carried out to interpolation image block respectively, obtains the image block after each transformation；Five, whole picture changing image is constituted by the image block after all transformation；Six, the changing image obtained using improved SPIHT methods pair five is encoded, and obtains encoded images.The present invention is used for compression of images field.

Description

Method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things

Technical field

The present invention relates to method for compressing image.

Background technology

Due to the huge advance of computing technique and sensor technology in recent years, Internet of Things (Internet of things, IoT) also enter fast development period [1] (Sezer OB, DogduE, Ozbayoglu AM (2018) Context-Aware Computing,Learning,and Big Data in Internet of Things:A Survey.IEEE Internet of Things Journal5(1):1-27.http://dx.doi.org/10.1109/JIOT.2017.2773600).In IoT Under meaning, " object " refers to relatively broad equipment, such as heart monitoring device, temperature measurement equipment and autonomous driving vehicle etc. [2-3]([2]XuL D.,HeW,LiS(2014)Internet of things in industries:a survey.IEEE Transactions on Industrial Informatics 10(4):2233-2243.http://dx.doi.org/ 10.1109/TII.2014.2300753[3]Iqbal M M,Farhan M,Jabbar S,et al(2018)Multimedia based IoT-centric smart framework for eLearning paradigm.Multimed Tools Appl 1-20.https://doi.org/10.1007/s11042-018-5636-y).IoT allows these equipment that can be set by network Progress remotely perception or remote control are applied, and the node energy of this kind of network, memory space and network bandwidth much smaller than pass System network.Moreover, with the development of multimedia technology, the data volume transmitted is needed also rapidly to increase, user also tends to more matchmakers Body signal (such as image or video) quality is put forward higher requirements.Therefore, multimedia letter how is efficiently transmitted under IoT environment Number, it is a problem that urgently needs to be solved.The basic result of IoT systems is as shown in Figure 1.In IoT, different equipment Be often used for different applications, this make these equipment have different data-handling capacity and transmission demand [4] (Khan R, Khan S U,Zaheer R,et al(2013)Future Internet:The Internet of Things Architecture,Possible Applications and Key Challenges[C].International Conference on Frontiers of Information Technology.IEEE,257-260.http:// dx.doi.org/10.1109/FIT.2012.53).In this case, there is low complex degree and can support multiple bit rate The compression method of transmission is more applicable.As the key technology in multimedia communication, compression of images is in our life can not Or lack.A kind of effective method for compressing image, should be able to make full use of the statistic correlation of signal, first be carried out to signal abundant Expression, then again to after expression signal carry out efficient coding.In order to improve the compression performance of image, domestic and foreign scholars are scheming Many work have been done as indicating and improving coding efficiency method.In graphical representation, the method based on transformation is the most commonly used.It is discrete Cosine transform (Discrete cosine transform, DCT) is the basis of Joint Photographic Experts Group.JPEG performances under low compression ratio Preferably, and when compressing relatively high, will occur blocking artifact in reconstruction image.Wavelet transform (Discrete Wavelet transform, DWT) solve the problems, such as this, and be always that image analysis and coding are led in the past twenty years Most important tool [5] (Liu S, Fu W, He L, et al (2017) Distribution of primary in domain additional errors in fractal encoding method.Multimed Tools Appl76(4):5787- 5802.http://dx.doi.org/10.1007/s11042-014-2408-1).Many famous method for compressing image or mark Standard, such as EZW [6] (J.M.Shapiro (1993) Embedded image coding using zerotrees of wavelet coefficients.IEEE Trans Signal Process41(12):3445–3462.http://dx.doi.org/ 10.1109/78.258085)、SPIHT[7](Said A,Pearlman W A(1996)A new,fast,and efficient image codec based on set partitioning in hierarchical trees.IEEE Trans Circuits SystVideo Technol6(3):243–250.http://dx.doi.org/10.1109/76.499834)、 SPECK[8](Pearlman W A,Islam A,NagarajN,Said A(2004)Efficient low complexity image coding with a set-partitioning embedded block coder.IEEE Trans Circuits Syst Video Technol,14(3):1219–1235.http://dx.doi.org/10.1109/ ) and JEPG2000 [9] (JPEG2000Image Coding System, ISO/IEC Std.15 TCSVT.2004.835150 444-1, (2000)), it is all based on DWT's.Although DWT can carry out effective table to the both horizontally and vertically information of image Show, isotropic characteristic prevents it preferably to be indicated from the direction character to image, such as edge and texture [10] (Shi C,Zhang J,Chen H,Zhang Y(2015)A Novel Hybrid Method for Remote Sensing Image Approximation Using the Tetrolet Transform.IEEE JSel Topics Appl Earth Observ 7(12):4949-4959.http://dx.doi.org/10.1109/JSTARS.2014.2319304).Therefore, it carries Some direction wavelet basis are gone out, such as curvelet [11] (Candes E J, Donoho D L (2004) New tight frames of curvelets and optimal representations of objects with piecewise C2 singularities.Commun Pure Appl Math57(2):219–266.http://dx.doi.org/10.1002/ cpa.10116)、contourlet[12](Do M N,Martin V(2005)Thecontourlet transform:an efficient directional multiresolution image representation,IEEE Trans Image Process 14(2):2091-2106.http://dx.doi.org/10.1109/TIP.2005.859376)、 directionlet[13](V.Velisavljevic,B.Beferull-Lozano,M.Vetterli,P.L.Dragotti (2006)Directionlets:Anisotropic multidirectional representation with separable filtering.IEEE Trans Image Process15(7):1916–1933.http:// ) and shearlet [14] (Kutyniok G, Lim WQ (2011) dx.doi.org/10.1109/TIP.2006.877076 Full length article:Compactly supported shearlets are optimally sparse.Journal of Approximation Theory163:1564-1589.http://dx.doi.org/ 10.1016/j.jat.2011.06.005) etc..These wavelet basis are more sensitive to certain specific directions, therefore can retain The more specific direction features of image.Some adaptive direction wavelet basis, such as bandelet [15] (Erwan L P, St é phane M(2005)Sparse geometric image representations with bandelets.IEEE Trans Signal Process 14(4):423-438.http://dx.doi.org/10.1109/TIP.2005.843753)、 wedgelet[16](Donoho D L(1999)Wedgelets:nearly minimax estimation of edges.Annals of Statistics27(3):859-897.http://dx.doi.org/10.1214/aos/ 1018031261)、grouplet[17](Mallat S(2009)Geometrical grouplets.Appl.Comput Harmon Anal26(2):161-180.http://dx.doi.org/10.1016/j.acha.2008.03.004) and EPWT (Easy path wavelet transform), can carry out image more flexible expression.However, these wavelet basis are usual Design with complexity, some wavelet basis even redundancy, this makes it not be widely used in compression of images.

Method based on Lifting Wavelet can carry out adaptive boosting in image local.Many work are promoted specific Method is fused in wavelet transformation frame, to improve compression performance, such as [19-23] ([19] Ding W, Wu F, Wu X, Li S, Li H(2007)Adaptive directional lifting-based wavelet transform for image coding.IEEE Trans Image Process 16(2):416-427.http://dx.doi.org/10.1109/ TIP.2005.843753[20]C.Chang and B.Girod(2007)Direction adaptive discrete wavelet transform for image compression.IEEE Trans Image Process16(5):1289– 1302.http://dx.doi.org/10.1109/TIP.2007.894242[21]Zhang L,Qiu B(2013)Fast orientation prediction-based discrete wavelet transform for remote sensing image compression.Remote Sensing Letters4(12):1156-1165.https://doi.org/ 10.1080/2150704X.2013.858838[22]Chen D,Li Y,Zhang H,Gao W(2017)Invertible update-then-predict integer lifting wavelet for lossless image compression.EURASIP JAdvSignal Process 1:1-9.http://dx.doi.org/10.1186/ s13634-016-0443-y[23]Hasan M M,Wahid K A(2017)Low-Cost Architecture of Modified Daubechies Lifting Wavelets Using Integer Polynomial Mapping.IEEE Trans Circuits Syst 64(5):585-589.http://doi.org/10.1109/tcsii.2016.2584091)。 These compression methods based on promotion usually exist with adaptivenon-uniform sampling, statistical model, direction prediction, or the wavelet basis contact of modification Together, and compression performance promotion mainly from based on rate-distortion optimization segmentation or opposite side information coding obtain.These sides In method, in view of being protected to image material particular seldom in compression process.And the problem can influence code efficiency into One step is promoted, especially to texture region.In addition, cannot fully indicate image detail, the subjectivity of reconstruction image can be also influenced Quality.Therefore, a kind of effective image representing method how is designed, is the major issue in compression of images.

Coding is another key link in compression of images.For wavelet image, in the identical sky of different high-frequency sub-bands Between position coefficient, have strong correlation.In addition, after carrying out effective graphical representation, it will usually occur in small echo high-frequency region A large amount of inessential " block ".If can these are unessential " block " encode in an appropriate manner, coding efficiency can be further It is promoted.To the compression of images based on wavelet transformation, based on embedded block coding (the embedded block most preferably blocked Coding with optimized truncation, EBCOT) it is famous coding method, and adopted by JPEG2000 standards With [9] (JPEG2000Image Coding System, ISO/IEC Std.15 444-1, (2000)).The basic think of of EBCOT Think to be by each sub-band division to be several pieces, such as 32 × 32 or 64 × 64, then these blocks are separately encoded, and in different bits Under rate, according to compression after rate anti-aliasing techniques (post compression rate distortion, PCRD), to these code streams into Row blocks.Although preferable coding efficiency can be obtained, one of JPEG2000 the disadvantage is that do not utilize intersubband same position system Correlation [24] (Christophe E, Mailhes C, Duhamel P (2008) Hyperspectral image between number compression:adapting SPIHT and EZW to anisotropic 3-D wavelet coding.IEEE Trans Image Process17(12):2334-2346.http://dx.doi.org/10.1109/ TIP.2008.2005824).According to [25] (D.S.Taubman and M.W.Marcellin (2002) JPEG2000Image Compression Fundamentals,Standards and Practice.Boston,MA:Kluwer analysis), The selection of point of cut-off in JPEG2000 compensates for the deficiency for not utilizing intersubband set membership.However, this is with higher calculating Complexity is cost.[26](Lewis A S,Knowles G(1992)Image Compression Using the 2-D Wavelet Transform.IEEE Trans Image Process 1(2):244-250.http://dx.doi.org/ 10.1109/83.136601) point out, tree be it is a kind of can indicate subband Relationship of Coefficients in Wavelet image have efficacious prescriptions Method.For tree data structure, SPIHT is a kind of most common coding method, can utilize the set membership of intersubband, from And provide preferable coding efficiency.In recent years, it is proposed that the method for compressing image based on improved SPIHIT, such as [27-29] ([27] Hamdi M,Rhouma R,Belghith S(2017)A selective compression-encryption of images based on SPIHT coding and Chirikov Standard Map 131:514-526.Signal Processing.http://dx.doi.org/10.1016/j.sigpro.2016.09.011[28]Song X,Huang Q, Chang S,He J,Wang H(2016)Three-dimensional separate descendant-based SPIHT algorithm for fast compression of high-resolution medical image sequences.IET Image Processing11(1):80-87.http://dx.doi.org/10.1049/iet-ipr.2016.0564[29] Zhang M,Tong X(2017)Joint image encryption and compression scheme based on IWT and SPIHT.Optics&Lasers in Engineering90:254-274.http://dx.doi.org/ 10.1016/j.optlaseng.2016.10.025), also by improved SPIHIT method be used for video image compression [30-32] ([30]Kim S,Jang JH,Lee HJ,Rhee CE(2017)Fine-scalable SPIHT Hardware Design for Frame Memory Compression in Video Codec.Journal of Semiconductor Technology Andence17(3):446-457.http://dx.doi.org/10.5573/JSTS.2017.17.3.446 [31]El-Bakery EM,El-Rabaie S,Zahran O,El-Samie FEA(2017)Chaotic Interleaving for the Transmission of Compressed Video Frames with Self-Embedded Digital Signatures.Wireless Personal Communications96(2):1635-1651.http://dx.doi.org/ 10.1007/s11277-017-4218-z[32]Sowmyayani S,Rani P A J(2016)

An Efficient Temporal Redundancy Transformation for Wavelet Based Video Compression.International Journal of Image&Graphics16(3): 1650015.http://dx.doi.org/10.1142/S0219467816500157).This illustrates SPIHT methods since its is low Complexity and flexibility become the popular technique in multimedia communication.Although having more research to SPIHT methods, this A little researchs are most of all to be concentrated on how being further reduced in redundancy bits or scanning redundancy, seldom in view of by high-frequency information Edge blurry or ringing effect caused by missing.

Invention content

The purpose of the present invention is to solve existing SPIHT methods seldom to consider side caused by the missing by high-frequency information Edge obscures or ringing effect, can not retain more details in image, lead to the problem that code efficiency is low, and proposes base under Internet of Things In direction Lifting Wavelet and the method for compressing image of improved SPIHIT.

Method for compressing image detailed process based on direction Lifting Wavelet and improved SPIHIT under Internet of Things is：

Step 1: carrying out image block segmentation, the image block after being divided to remote sensing image；

Step 2: calculating separately optimum prediction direction to the image block after segmentation, image block is best pre- after being divided Survey direction；

Step 3: weighting directional interpolation filter coefficient by calculating, directional interpolation is weighted to fractional samples value, is obtained To interpolation image block；

Step 4: the optimum prediction direction obtained using step 2, to interpolation image block promoted based on direction respectively Wavelet transformation, obtain the image block after each transformation, i.e., the code block after each transformation；

Step 5: constituting whole picture changing image by the image block after all transformation；

Step 6: being encoded to the changing image that step 5 obtains using improved SPIHT methods, scheme after being encoded Picture.

Beneficial effects of the present invention are：

The present invention proposes a kind of new method for compressing image, and this method is by the adaptive boosting small echo based on directional interpolation Convert (directional interpolation-based adaptive lifting wavelet transform, DIAL- DWT it) is combined with improved SPIHT methods.Main innovation point includes two parts：First, the adaptive boosting small echo proposed becomes Change, can utilization orientation interpolation filter and optimal self-adaptive direction of improvement, image is adequately indicated；Second is that improving SPIHT coding, this method can retain detailed information important in image as possible, while be capable of providing preferable binary encoding Energy.The compression method of proposition is asymmetrical, has lower complexity in decoding end, this so that this method is highly suitable for pair Data transmission and real-time have the various IoT terminals of different requirements.The experimental results showed that the method for proposition is not only than traditional pressure Contracting method has higher compression performance, and can largely improve the subjective quality of reconstruction image.

The present invention devises a DIAL model, which can calculate separately the best direction of improvement of all image blocks, And directional interpolation is weighted to fractional samples in lifting process.Therefore, the wavelet transformation based on DIAL models carries in direction During rising, optimum orientation prediction and weighting directional interpolation are combined, is capable of providing more effective graphical representation, this helps to protect Stay the more direction characters of image.Due to the high concentration of image energy, the image representing method be capable of providing it is more " compared with Zero tree of length ", to improve code efficiency.It solves existing SPIHT methods seldom to consider to be caused by the missing of high-frequency information Edge blurry or ringing effect, more details in image can not be retained, lead to the problem that code efficiency is low.

The present invention devises a kind of improved SPIHT methods, and this method only changes inessential row in existing SPIHT methods The scanning sequency of table (List of insignificant sets, LIS), does not need to additional calculating, can be mutually year-on-year More significant coefficients are encoded under special rate.Improved SPIHT methods can retain more material particular information in image, can carry High binary encoding performance, does not need additional calculation amount, does not need additional bit as header file yet.

It is tested under different bit rates using the image in different images library, the experimental results showed that, PSNR highests carry 1.3dB is risen.

Description of the drawings

Fig. 1 is basic IoT system framework figures；

Fig. 2 a are that one-dimensional square lifting wavelet transform forward direction decomposes basic process figure, and x is original image, X_eFor in image Even samples set, X_oFor the odd samples set in image, DA_P_oThe direction-adaptive prediction used when being converted for the first order Operator, DA_U_oThe direction-adaptive update operator used when being converted for the first order, DA_P_kFor the side used when -1 grade of transformation of kth To adaptive prediction operator, DA_U_kFor the direction-adaptive update operator used when -1 grade of transformation of kth, K_eFor to changing image The weights that low frequency component is weighted, K_oWeights after being weighted for the high fdrequency component to changing image, a are finally obtained The low frequency component of changing image, b are the high fdrequency component of the changing image finally obtained；

Fig. 2 b are one-dimensional square lifting wavelet transform inverse composition basic process figure, x_eFor the even samples in reconstruction image Set, x_oFor the odd samples set in reconstruction image；

Fig. 3 a are the set of reference directions schematic diagram of the level wavelet transformation promoted based on direction, and m is the horizontal seat of tile location Mark, n are tile location ordinate；

Fig. 3 b are the set of reference directions schematic diagram of the vertical wavelet transform promoted based on direction；

Fig. 4 is the procedure chart in the optimum prediction direction for calculating given image block, and k is the serial number of reference direction；

Fig. 5 is the procedure chart of directional interpolation in horizontal variation；

Fig. 6 is the procedure chart for generating directional interpolation filter, a_-3、a_-2、a_-1、a₀、a₁、a₂For the parameter of interpolation filter；

Fig. 7 a are the level-one wavelet decomposition result figure of 9/7 wavelet filter；

Fig. 7 b are the level-one wavelet decomposition result figure of the wavelet filter based on ADL；

Fig. 7 c are the level-one wavelet decomposition result figure of the wavelet filter based on DIAL models；

Fig. 8 is the NLA result figures that different sparse representation methods obtain, and NLA is non-linear estimations, and The DIALmodel are DIAL (directional interpolation-based adaptive lifting wavelet transform, DIAL- DWT) model, ADL are that adaptive direction promotes (Adaptive direction lifting), and PSNR is Y-PSNR；

Fig. 9 a are that Europa3 tests remote sensing image collection schematic diagram；

Fig. 9 b are that bank tests remote sensing image collection schematic diagram；

Fig. 9 c are that aerial tests remote sensing image collection schematic diagram；

Fig. 9 d are that Lena tests remote sensing image collection schematic diagram；

Fig. 9 e are that Baboon tests remote sensing image collection schematic diagram；

Fig. 9 f are that pleiades_portdebouc_pan tests remote sensing image collection schematic diagram；

Figure 10 is under different bit rates, and the present invention proposes that the Kappa coefficient results of method and SPIHT methods compare figure；It is horizontal Coordinate is bit rate, unit bpp；Ordinate is Kappa coefficients；

Lena proposed are that the method for the present invention compresses test image Lena, and Lena SPIHT are SPIHT methods to surveying Attempt to compress as Lena；

Baboon proposed are that the method for the present invention compresses test image Baboon, and Baboon SPIHT are the side SPIHT Method compresses test image Baboon；

Bank proposed are that the method for the present invention compresses test image bank, and bank SPIHT are SPIHT methods to surveying Attempt to compress as bank；

Aerial proposed are that the method for the present invention compresses test image aerial, and aerial SPIHT are the side SPIHT Method compresses test image aerial；

Europa3proposed is that the method for the present invention compresses test image europa3, and europa3 SPIHT are SPIHT Method compresses test image europa3；

WoodlandHills proposed are that the method for the present invention compresses test image WoodlandHills, WoodlandHills SPIHT are that SPIHT methods compress test image WoodlandHills；

Figure 11 a are under 0.0625bpp bit rates, and the present invention proposes the reconstruction figure that compression method obtains；

Figure 11 b are the reconstruction figure that traditional SPIHT methods obtain under 0.0625bpp bit rates

Figure 11 c are under 0.125bpp bit rates, and the present invention proposes the reconstruction figure that compression method obtains；

Figure 11 d are the reconstruction figure that traditional SPIHT methods obtain under 0.125bpp bit rates；

Figure 11 e are under 0.25bpp bit rates, and the present invention proposes the reconstruction figure that compression method obtains；

Figure 11 f are the reconstruction figure that traditional SPIHT methods obtain under 0.25bpp bit rates；

Figure 11 g are under 0.5bpp bit rates, and the present invention proposes the reconstruction figure that compression method obtains；

Figure 11 h are the reconstruction figure that traditional SPIHT methods obtain under 0.5bpp bit rates；

Figure 11 i are under 1bpp bit rates, and the present invention proposes the reconstruction figure that compression method obtains；

Figure 11 j are the reconstruction figure that traditional SPIHT methods obtain under 1bpp bit rates.

Specific implementation mode

Specific implementation mode one：Image based on direction Lifting Wavelet and improved SPIHIT under the Internet of Things of present embodiment Compression method detailed process is：

Step 3: directional interpolation filter coefficient is weighted by calculating, to needing the score used in the lifting process of direction Sample value is weighted directional interpolation, obtains interpolation image block；

Specific implementation mode two：The present embodiment is different from the first embodiment in that：To remote sensing in the step 1 Image carries out image block segmentation, the image block after being divided；Detailed process is：

In order to keep direction of improvement consistent with the local grain direction of image, image segmentation is first carried out.In document [19], Use a kind of rate-distortion optimization dividing method based on quaternary tree.However, the efficiency and picture material of this dividing method It is closely related.To some image types, such as remote sensing images, complicated landforms are usually reflected, therefore detailed information is typically more It enriches, the flat site of few large area, self-adapting division method is difficult to show its advantage at this time.The reason is that, right Image with complex contents, using the result of self-adapting division method, it is more likely that almost all of piece is all to allow to divide Minimum block, this result is almost equal with the block segmentation result of directly progress same size, but with higher calculating Complexity is cost.In addition, being a large amount of side information using another expense of self-adapting division method.To being based on rate distortion The method of optimization, to different bit rates, corresponding " cut tree " is different.In order to be correctly decoded, these " cut trees " Also it to be sent as side information to decoding end.Picture material is more complicated, and the branch of " cut tree " is more, resulting side information It is more.Therefore, the rate-distortion optimization dividing method based on quaternary tree is not appropriate for all images.

The block partitioning scheme of same size is employed herein in order to make dividing method that there is generality based on above-mentioned analysis. To the image I that a width size is M × N, if block size is 16 × 16.Therefore, initial pictures block is represented by B_i,_j, i=1, 2 ..., M/16, j=1,2 ..., N/16.Any two image block is all unduplicated, and all image blocks constitute whole picture figure As I.After transformation, block size depends on Decomposition order.It is assumed that total Decomposition order of directional wavelet transform is J, it is right to decomposition layer k The block size answered is L_k×L_k.That is

L_k=16/2^k-1, k=1,2 ..., J

Compared with the Adaptive quadtree partition method based on rate-distortion optimization, the partitioning scheme of this same size is big The earth reduces complexity, and need not transmit side information.

Next, calculating the optimum prediction direction of each block.Assuming that set of reference directions is θ_ref=[- 7, -6, -5, -4, - 3, -2, -1,0,1,2,3,4,5,6,7], to each piece of B_l, l=0,1 ... MN/256-1, corresponding optimum prediction directionFor

D () indicates the measure of image fault.Herein, D () is defined as | |.That is, to every A block B_l, optimum prediction direction is the direction of corresponding minimum prediction error.Find process such as Fig. 4 institutes of optimical block prediction direction Show.

It can be with output, for given image, first along all reference direction θ from Fig. 4_ref,i(i=1,2 ..., 15), point Directional wavelet transform is not carried out.Then, in these changing images, to the block B of same position_l(l=0,1 ..., MN/256- 1) prediction error, direction, that is, optimum prediction direction of corresponding minimum prediction error are calculated separatelyEach sample in image block Prediction and renewal process are shown in Fig. 2 a.

Compared with self-adapting division method, the Directional Lifting Wavelet Transform of proposition need not transmit under each bit rate " cut tree ", it is only necessary to transmit the optimum prediction direction of each block.It is therefore proposed that the side information very little needed for method.

By Remote Sensing Image Segmentation at identical piece of size, the image block after being divided, block here divides size, should be with The block size of coding stage is consistent below.

Adaptive boosting wavelet transformation (DIAL-DWT) based on directional interpolation

Traditional two-dimentional lifting wavelet transform merely with horizontal or vertical direction adjacent sample.However, it is most of from Right image includes many different directional informations, and such as edge, profile and texture etc., this makes traditional two-dimentional Lifting Wavelet Transformation can not indicate these directional informations well.How a kind of effective image representing method is provided, is to improve image The key of compression performance.Here, it is proposed that a kind of new DIAL-DWT methods.This method now divides an image into several pieces, so The best direction of improvement of each block is calculated afterwards.Next, utilization orientation interpolation filter to fractional samples into row interpolation, to Retain more direction characteristic in interpolation image.The detailed design process of DIAL-DWT method methods is as follows.

The structure of Directional Lifting Wavelet Transform

Typical lifting wavelet transform includes four steps：Division, prediction, update, and standardization [33] (Sweldens W(1995)The lifting scheme:a construction of second generation wavelets.SIAM J Math Anal29(2):511-546.http://dx.doi.org/10.1137/S0036141095289051).It does not lose general Property, basic Directional Lifting Wavelet Transform is also based on this four steps.The frame of one-dimensional square lifting wavelet transform and inverse transformation It is as shown in Figure 2 a and 2 b respectively.

For a width two dimensional image x (m, n)_m,n∈Z, first, all samples are divided into two parts：Even samples set x_eWith Odd samples set x_o。

In forecast period, odd samples are predicted by adjacent even samples, and prediction direction is by a certain What decision criteria obtained.Assuming that direction-adaptive predictive operator is DA_P, then predict that process is represented by

D [m, n]=x_o[m,n]+DA_P_e[m,n] (2)

In the more new stage, even samples are updated by the prediction error of adjacent sample, more new direction and prediction side To identical.Assuming that direction-adaptive update operator is DA_U, then renewal process is represented by

C [m, n]=x_e[m,n]+DA_U_d[m,n] (3)

Here, direction prediction operator DA_P is

Direction update operator DA_U is

Here, p_iAnd u_jThe coefficient of high-pass filter and low-pass filter is indicated respectively.θ_vIndicate prediction and newer side To.

Finally, COEFFICIENT K is used in the output after promotion respectively_eAnd K_oIt is weighted.

After the above process, the low pass subband L and a high pass subband H of horizontal direction can be obtained.Next, In a like fashion, one-dimensional column direction transformation is carried out.

The selection of direction of improvement θ is extremely important.In order to carry out preferable graphical representation, image is first divided into several images Block, and calculate separately direction of improvement to each piece.For given block, all sample standard deviations are promoted by identical direction in block.Reason By upper, more with reference to direction of improvement, the expression of image block is better, but needs the side information transmitted also more.If on the contrary, only A few refers to direction of improvement, then cannot indicate image well.Here, one-dimensional horizontal transformation and vertical transitions are selected 15 refer to direction of improvement, as shown in Figure 3a and Figure 3b shows respectively.Anisotropic filter can be along direction d=(d_x,d_y)^T, It is indicated.Here, some adjacent integers and fractional samples are utilized in 15 reference directions.These reference directions are as follows：d_-7 =(3, -1)^T, d_-6=(2, -1)^T,d_-5=(1, -1)^T,d_-4=(3/4, -1)^T,d_-3=(1/2, -1)^T,d_-2=(1, -3)^T,d_-1 =(1/4, -1)^T,d₀=(0, -1)^T,d₁=(- 1/4, -1)^T, d₂=(- 1, -3)^T, d₃=(- 1/2, -1)^T, d₄=(- 3/4, -1 )^T,d₅=(- 1, -1)^T,d₆=(- 2, -1)^T,d₇=(- 3, -1)^T.Set of reference directions is as shown in Figure 3.

Other steps and parameter are same as the specific embodiment one.

Specific implementation mode three：The present embodiment is different from the first and the second embodiment in that：It is right in the step 2 Image block after segmentation calculates separately optimum prediction direction, the optimum prediction direction of image block after being divided；Detailed process is：

To the wavelet transformation promoted based on direction, predict that error and high-frequency sub-band are closely related.Predict that error is bigger, Information in high-frequency sub-band is more, and coding efficiency is lower.For an image block, optimum prediction direction should can make height The direction of frequency subband residual risk minimum.

The process for calculating image block optimum prediction direction is：As shown in figure 4,

Assuming that set of reference directions is θ_ref, set of reference directions θ_refIncluding 15 reference directions, these directions are denoted as -7, - 6,-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7}；If the image block sum after segmentation is N_a, each image block is B_l, l= 0,1,...,N_a-1；

Image block B after segmentation_lRespectively along all reference direction θ_ref,i(i=1,2 ..., 15) carry out direction prediction, Obtain the prognostic chart picture block under all reference directions；

Under mean-square error criteria, prognostic chart picture block pixel under all reference directions respectively with step 1 remote sensing image picture Element compares, when error minimum corresponding reference direction, the as optimum prediction direction of the prognostic chart picture block

The optimum prediction direction of prognostic chart picture blockIt calculates as follows

In formula, D () is image fault function, and x (m, n) is image block B_lThe corresponding sample value of middle position (m, n), DA_Pⁱ For the predictive operator of i-th of reference direction, m is the abscissa of corresponding position, and n is the ordinate of corresponding position；Enable D ()= |·|；

Sample：Pixel is in original image, and coefficient is in changing image.That is, before first order wavelet transformation, Here it is pixel.But wavelet transformation is typically multistage, is all just coefficient here since the second level.In order to facilitate statement, Collectively referred to here in as sample.

It repeats the above process, the optimum prediction direction of image block after determining all segmentations；

Compared with self-adapting division method, the small wave converting method of proposition promoted based on direction does not need additional transmissions All " cut trees " under bit rates, it is only necessary to the corresponding optimum prediction direction of transmission block.It is therefore proposed that needed for method The auxiliary information of transmission is seldom.

Other steps and parameter are the same as one or two specific embodiments.

Specific implementation mode four：Unlike one of present embodiment and specific implementation mode one to three：The step 3 In by calculating weight directional interpolation filter coefficient, to needing the fractional samples value used to be weighted in the lifting process of direction Directional interpolation obtains interpolation image block；Detailed process is：

Directional interpolation

To the wavelet transformation promoted based on direction, some direction of improvement need to use the sample value of fractional position.Namely It says, the tangent tan θ of direction of improvement not always integer.Therefore, it is necessary to the sample of fractional position into row interpolation.Interpolation mistake Journey is represented by

Here, k indicates the integer position used in Interpolation Process；a_kIndicate the parameter of interpolation filter.In itself, sub- The process of picture element interpolation is exactly the design process of optimal interpolation filter.Most of wavelet transformations promoted based on direction, are all adopted With Sinc interpolation methods.However, similar with some other interpolation method, Sinc interpolation methods are also only with along horizontally or vertically side To sample come to fractional samples into row interpolation, this can make the directional information in image thicken.For texture or details compared with More images, according to Sinc interpolation methods, then direction prediction error will increase.Herein, it is slotting to use a kind of direction Value method, this method utilize adjacent integral sample, along local grain direction to fractional position into row interpolation.With horizontal transformation For, the process of directional interpolation is as shown in Figure 5.

To different fractional samples positions, the integral sample for interpolation is also different, and the characteristic of this and local signal is mutually fitted It answers.Due to different integral samples fractional samples position is contributed it is different, interpolation filter also should different [34] (Liu Y, Ngan K N(2008)Weighted adaptive lifting-based wavelet transform for image coding.IEEE Trans.Image Process17(4):500-511.http://dx.doi.org/10.1109/ TIP.2008.917104).The interpolation filter of use is as shown in Figure 6.As seen from Figure 6, directional interpolation filter is final Coefficient is determined by three kinds of filters, is bi-linear filter, Telenor 4-tap filters and 2-tap filtering respectively Device.The coefficient of these filters is shown in Table 1.

The interpolation filter coefficients that table 1 uses

In figure 6, some different integral samples are used for the interpolation of fractional samples, interpolation direction and the letter for being used for interpolation Number local characteristics be adapted.For example, in order to, into row interpolation, not only use integer position to the sample of a quarter position Sample { a_-2,a_-1,a₀,a₁, also to use the sample { a along prediction direction_-3,a₂}.These samples { a_-3,a_-2,a_-1,a₀,a₁, a₂Can be used for building directional interpolation filter, then the sample of fractional position is predicted.It will be appreciated from fig. 6 that { a_-3,a₂Be Bi-linear filter volume inputs, { a_-2,a_-1,a₀,a₁Be Telenor 4-tap filters input, bi-linear filter and The output of Telenor 4-tap filters together constitutes the input of 2-tap filters.Therefore, the output of 2-tap filters is just It is the coefficient of directional interpolation filter.The correspondence of directional interpolation filter coefficient and different fractional position samples is shown in Table 2.

2 directional interpolation filter coefficient of table

The final output of directional interpolation filter is determined by three filters, is respectively：Bi-linear filter, Telenor 4-tap filters and 2-tap filters；

Below current sample is expert in two rows, two in the row arranged with the sample column interval two are taken respectively Sample, the input as bi-linear filter；In row, the lastrow where current sample, and in lower two rows, the sample is taken respectively Four samples of the next column of this column, as the input of Telenor 4-tap filters, bi-linear filter and The output of Telenor 4-tap filters constitutes the input of 2-tap filters, and the output of 2-tap filters is exactly that direction is inserted The weighting coefficient of value filter；

As seen from Figure 6, { c_-3,c₂Integral sample be bi-linear filter input, { c_-2,c_-1,c₀,c₁Integer sample Originally it is the input of Telenor 4-tap filters, the output of bi-linear filter and Telenor 4-tap filters constitutes 2- The input of tap filters, the output of 2-tap filters are exactly the weighting coefficient of directional interpolation filter；

Pass through integer position sample { c_-3,c_-2,c_-1,c₀,c₁,c₂And weighting coefficient structure directional interpolation filter, direction Interpolation filter is weighted directional interpolation to the sample value of fractional position, obtains interpolation image block.

Other steps and parameter are identical as one of specific implementation mode one to three.

Specific implementation mode five：Unlike one of present embodiment and specific implementation mode one to four：The step 4 The middle optimum prediction direction obtained using step 2, is carried out the wavelet transformation promoted based on direction to interpolation image block respectively, obtained Image block to after each transformation, i.e., the code block after each transformation；Detailed process is：

The optimum prediction direction obtained according to step 2Formula (2) and (3) are utilized respectively to insert to what step 3 obtained Value image block carries out the wavelet transformation promoted based on direction：

Direction prediction operator DA_P is

In formula, x_e[m, n] is the even samples set for the interpolation image block that step 3 obtains, DA_P_e[m, n] is even number sample The corresponding direction prediction operator of this set；I indicates the serial number of high-pass filter coefficient, p_iIndicate high-pass filter coefficient；

Interpolation image block is divided into two parts：Even samples set x_e[m, n] and odd samples set x_o[m,n]；

Direction update operator DA_U is

In formula, j indicates the serial number of low-pass filter coefficients, u_jIndicate low-pass filter coefficients, DA_U_d[m, n] is odd number sample The corresponding direction update operator of this set；D [m, n] is to be expressed as by the odd samples after neighbouring even-numbered sample predictions

D [m, n]=x_o[m,n]+DA_P_e[m,n]

x_o[m, n] is the odd samples set for the interpolation image block that step 3 obtains；

Utilization orientation predictive operator and direction update operator obtain the code block after each transformation.

Other steps and parameter are identical as one of specific implementation mode one to four.

Specific implementation mode six：Unlike one of present embodiment and specific implementation mode one to five：The step 6 It is middle that the changing image that step 5 obtains is encoded using improved SPIHT methods, obtain encoded images；Detailed process For：

SPIHT coding methods are exactly to be encoded to changing image.The coefficient mentioned in coding method each means transformation Wavelet coefficient in image.

Recently increasingly extensive concern has been obtained come the coding method based on tree.In these coding methods based on tree, SPIHT methods are due to preferable distortion performance and moderate complexity, being most widely used.However, SPIHT methods Scan mode limits its coding efficiency.In the scanning process of SPIHT, the importance of coefficient only passes through the absolute value of its assignment To judge.In fact, human eye is more sensitive to the contour distortion of image.In high-frequency sub-band, the wavelet coefficient at image outline Often there is larger amplitude.The gray level variation of image is typically slow, therefore in high-frequency sub-band, is centered around important system The wavelet coefficient that several weeks enclose generally also has larger amplitude.From another angle, it is if be centered around around a coefficient Number is all important, then this coefficient also has prodigious probability to be important, even if the coefficient amplitude and not up to specified threshold Value.The significant coefficient being centered around around a coefficient is more, this coefficient is generally also more important.Therefore, if those are possessed very The coefficient also priority encoding of more important " neighbours " can encode more more important coefficient then under to bit rates, so as to improve weight Build the quality of image.

A kind of coding efficiency that good Image Coding Algorithms should be able to not only provide, will also there is faster operation speed.So And the two is often contradictory.The reason is that, the raising of coding efficiency is often to improve computation complexity as cost.Cause This reduces algorithm complexity how while the coding efficiency provided, is the problem of another needs research.

This paper presents a kind of improved SPIHT methods, this method energy priority scan has the coefficient of important " neighbours ", from And improve coding efficiency.In order to reduce algorithm complexity, proposition method only changes the partial scan sequence of SPIHT, and is not required to Want additional calculation amount.Another advantage of proposition method is that scanning sequency is adaptively determined by the significant coefficient being previously obtained , therefore any information need not be stored as header file.

To spiht algorithm, D collection is indicated with inessential aggregate list (list of insignificant sets, LIS) It closes and L gathers.LSP is first initialized as an empty table, LIP initializes lowest frequency sub-band coefficients location sets, and LIS is initialized as The root node coordinate set of each spatial orientation tree.To each bit-plane, by taking turns stream encryption LIP, LIS, the note in and LSP Record, to realize compression of images.

Step 6 one, initial threshold value T=2^n′, initial table LSP, LIS and LIP；N ' is the maximum value of bit-plane number；

Table LSP is initialized as an empty table, LIP is initialized as lowest frequency sub-band coefficients location sets, and LIS is initialized as The root node coordinate set of each spatial orientation tree；

Step 6 two encodes LIP according to table LSP, LIS and LIP, and process is：

Whether step 621 includes significant coefficient (the corresponding system in significant coefficient position according in threshold decision LIP set Number is significant coefficient), it is output 1 and sign bit, coefficient is just that sign bit 0, coefficient is negative, sign bit 1, by important system Numerical digit is set (i, j) and is deleted from LIP, and is added to the ends LSP；

It is no, then export 0；

Whether according to including significant coefficient in threshold decision LIP set, process is：

Coefficient is more than threshold value, is significant coefficient；Coefficient is less than or equal to threshold value, is not significant coefficient；

Step 6 two or two judges whether all coefficient positions for including in LIP set have been processed, if it is not, re-executing Step 621；If executing step 6 three；

Step 6 three, coding LIS, process are：

Step 631 judges that the current record of LIS is D (i, j) or L (i, j), if the current record of LIS be D (i, J), step 6 three or two is executed, if the current record of LIS is L (i, j), executes step 6 three or five；

D (i, j) is the coordinate set of all descendants of coefficient positions (i, j)；

L (i, j) is the coordinate set of all non-direct descendants of coefficient positions (i, j)；

Step 6 three or two, according in threshold decision D (i, j) whether include significant coefficient, be output be 1, otherwise output be 0；

If including significant coefficient in D (i, j), D (i, j) is decomposed into L (i, j) and O (i, j)；L (i, j) is made marks It is put into the tail portions LIS；

O (i, j) is the coordinate set of all children of coefficient positions (i, j)；

Whether according to including significant coefficient in threshold decision D (i, j), process is：

Quaternary tree is established with 4 coefficients of O (i, j) and encodes and (is encoded to output 0 or 1), if the tree root of quaternary tree (4 It is maximum in coefficient) it is more than or equal to threshold value, illustrate there is significant coefficient in this four coefficients, output 1；Otherwise, if the tree of quaternary tree Root is (maximum in 4 coefficients) to be less than threshold value, illustrates do not have significant coefficient in this four coefficients, output 0；Execute step 6 three Three；

Significant coefficient (significant coefficient in 4 coefficients) is put into LIP or LSP, and exports significant coefficient (4 by step 6 three or three Significant coefficient in a coefficient) symbol, significant coefficient be just, symbol 0, significant coefficient is negative, symbol 1；Execute step 6 Three or four；

Step 6 three or four judges whether L (i, j) is sky,

If so, deleting D (i, j) from LIS；Execute step step 6 three or eight；

If it is not, L (i, j) coefficient positions (i, j) move to the tail portions LIS；Execute step step 6 three or eight；

Step 6 three or five judges whether tape label (in step 6 three or two, obtained L (i, j) is decomposed to D (i, j) to L (i, j) It is marked (label has setting in a program)), if so, step 6 three or six is executed, if it is not, executing step 6 Radix Notoginseng；

Step 6 three or six, according in threshold decision L (i, j) whether include significant coefficient, be that L (i, j) is important, from LIS L (i, j) is deleted, D (2i, 2j), D (2i+1,2j), D (2i, 2j+1) and D (2i+1,2j+1) are added to the end of LIS, no Export any information；Execute step step 6 three or eight；

No, L (i, j) is inessential, executes step step 6 three or eight；

Whether according to including significant coefficient in threshold decision L (i, j), process is：

Step 6 Radix Notoginseng, according in threshold decision L (i, j) whether include significant coefficient, be that L (i, j) is important, then by L (i, j) is deleted from LIS, and D (2i, 2j), D (2i+1,2j), D (2i, 2j+1) and D (2i+1,2j+1) are added to LIS's End, and exports coding；Execute step step 6 three or eight；

No, L (i, j) is inessential, then does not export any information；

Step 6 three or eight；Judge whether the root node coordinate of all spatial orientation trees in LIS has been processed, if It is no, re-execute step 631；If executing step 6 four；

Step 6 four, the label for removing all L (i, j) check each (i, j) in LSP, if not in sequence scanning Newly plus (in current iteration), export the nth bit (101 third positions are exactly 1) of the position coefficient of correspondence, execute step 6 five； If newly adding in sequence scanning, then any information is not exported；

Step 6 five judges whether the length of compressed bit stream reaches specified length, if so, output compressed bit stream；If No T=T/2 executes step 6 two.

If can be seen that the importance for first judging L (i, j) from LIS cataloged procedures, then encode O (i, j) four are Number, then can save one.It is unessential since L (i, j) has very high probability, many positions can be saved in this way.It is worth note Meaning, this process does not increase additional bit or calculation amount, is only the sequence for changing judgement.Moreover, working as from D The L (i, j) that division obtains in (i, j) is important, then it includes significant coefficient that O (i, j), which has very high probability,.Therefore, Ke Yiyong A kind of effective mode encodes O (i, j).

To spiht algorithm, D collection is indicated with inessential aggregate list (list of insignificant sets, LIS) It closes and L gathers.First by LSP to each bit-plane, by taking turns stream encryption LIP, LIS, the record in and LSP, to realize image Compression.

Detailed LIS scanning processes are shown in algorithm 1 in improved SPIHIT algorithm.

Other steps and parameter are identical as one of specific implementation mode one to five.

Beneficial effects of the present invention are verified using following embodiment：

Embodiment one：

Method for compressing image based on direction Lifting Wavelet and improved SPIHIT under the present embodiment Internet of Things be specifically according to Prepared by lower step：

First, experiment is devised to verify the validity of the DIAL models of proposition.Then, to improved spiht algorithm into Test is gone.Finally, under different bit rates, using different quality evaluation standards, by the method for proposition and common compression side Method is compared.

The DIAL models of proposition

In order to prove the validity of the DIAL models proposed, test image is used as using common " Barbara ".The image Size is 512 × 512.By the test image respectively with 9/7 biorthog-onal wavelet filter, based on the wavelet filter of ADL, and Wavelet filter based on DIAL models carries out level of decomposition, and obtained decomposition result is as shown in Fig. 7 a, 7b, 7c.It can be with from Fig. 8 Find out, compared with the decomposition result obtained using 9/7 wavelet filter, changing image that the wavelet filter based on ADL obtains High-frequency sub-band has smaller coefficient amplitude.To the wavelet filter based on DIAL models, changing image medium-high frequency subband is taken a fancy to It goes almost black, illustrates that the sparse result that this method obtains is optimal.The reason is that, in lifting process, DIAL models More directional informations are considered, this helps more energy in image all concentrating on low frequency sub-band.Moreover, with common Sinc interpolation methods are compared, and DIAL models use directional interpolation, can be along local grain direction to fractional pixel position Into row interpolation.Therefore, it is possible to retain the more direction information in image.All these DIAL models that each contribute to obtain preferably Sparse performance.

The average amplitude of the high frequency coefficient obtained using these three transform methods is set forth in table 3, and relative to biography It unites the percentage of 9/7 wavelet transformation high-frequency sub-band coefficient amplitude reduction (with the digital representation in bracket).From table 3 it can be seen that For each high-frequency sub-band, the coefficient average amplitude of DIAL models is minimum.

The mean coefficient amplitude of LH, HL, HH and the percentage of reduction under 3 three kinds of transform methods of table

DIAL models are not needed upon the adaptive decomposition of rate-distortion optimization.Moreover, entropy coding method can be used into one Step reduces side information.Therefore, the side information for needing to transmit is considerably reduced.The comparison result of side information is shown in Fig. 4.It can from Fig. 4 To find out, less side information is needed compared to the DIAL methods with ADL methods, proposition, this is extremely to have to improving compression efficiency Profit.

Table 4. gives the code check (bpp) of the following information of bit rates

Non-linear estimations (NLA) are a kind of effective ways [35] that can weigh given transformation rarefaction representation ability (Eslami R,Radha H(2007)A new family of nonredundant transforms using hybrid wavelets and directional filter banks.IEEE Trans Image Process 16(4):1152- 1167.http://dx.doi.org/10.1109/TIP.2007.891791).If having preferable NLA performances, the transformation Method is in some signal processing applications, as being all more potential in coding, denoising and feature extraction.Therefore, it devises several Group tests the NLA performances of the wavelet transformation based on DIAL models to test proposition.To test image " Barbara ", retaining Under the coefficient number of different number, the NLA Xi Eng of distinct methods can be as shown in Figure 8.As seen from Figure 8, based on the small of DIAL models Wave conversion is better than always 9/7 common wavelet transformation and the wavelet transformation based on ADL.Especially when retention factor number M is less When, propose that the NLA performances of method become apparent.For other test images, such as " Boats ", " Fingerprint ", " GoldHill " and some texture images are tested NAL performances with identical method, can be led to the same conclusion.

The performance of the compression method of proposition

The compression method of proposition is to be combined the wavelet transformation based on DIAL models with improved SPIHT methods.In order to demonstrate,prove The validity of bright proposition method has carried out some experiments and has compared.Here six width test images have been selected, respectively from different biographies Sensor, and reflect different scenes.Wherein, " bank ", " aerial ", " Lena ", " Baboon ", and " Woodland Hills " is selected from USC-SIPI databases [36] (USC-SIPI database. [Online]:http://sipi.usc.edu/ Database/), " Europa3 " is selected from CCSDS test charts image set [37] (Consultative committee for space data systems,CCSDS test images.[Online].Available:http://cwe.ccsds.org/sls/ docs/sls-dc/).These test image sizes are 512 × 512, as shown in Fig. 9 a, 9b, 9c, 9d, 9e, 9f.

In an experiment, the wavelet decomposition number of plies is set as five layers.The compression method of it is proposed and traditional is respectively adopted SPIHT methods compress above-mentioned test image.The PSNR obtained under different bit rates the results are shown in Table 5.

Table 5 proposes the PSNR results (dB) that compression method and tradition SPIHT methods obtain

As can be seen from Table 5, under all to bit rates, propose that the coding efficiency of compression method is better than SPIHT's Coding efficiency.This is because, the rarefaction representation that DIAL models have been capable of providing by the more energy of image as a result, focus on low frequency Subband.This is for the coding method based on zero tree, it is meant that zero tree of more " longer " can be generated under same bits face. Moreover, improved SPIHT methods can more effectively scan these zero trees.It is all these to each contribute to propose compression side Method obtains better coding efficiency.

In order to comprehensively assess the compression method of proposition, Kappa coefficients are additionally used as quality evaluation index.Kappa Coefficient is usually used in assessing nicety of grading [38] (Gaucherel C, Alleaume S, Hely C (2008) The Comparison Map Profile Method:A Strategy for Multiscale Comparison of Quantitative and Qualitative Images.IEEE TransGeosciRemote Sens,46(9):2708-2719.http:// dx.doi.org/10.1109/TIP.2007.891791).Document [39] (Cohen J (1960) A coefficient of agreement for nominal scales.Educational and Psychological Measurement20(1): It 37-46.) points out, Kappa coefficients are also used as the measurement of original image and reconstruction image consistency.To these test images, Under different bit rates, the Kappa coefficients that proposition method and general SPIHT compression methods obtain are as shown in Figure 10.According to Figure 10, As can be seen that it is all to bit rates under, propose that the Kappa coefficients of compression method are still better than obtaining using SPIHT methods Result.

In addition to PSNR and Kappa coefficients, the subjective quality of image is also an important indicator for assessing compression algorithm performance. By taking test image " bank " as an example, under different bit rates, the reconstruction image that different compression methods obtain is as shown in figure 12.From figure The compression method that 11a, 11b, 11c, 11d, 11e, 11f, 11g, 11h, 11i, 11j can be seen that proposition is capable of providing preferably Reconstruction image visual quality, the more region of texture information especially in the block.This demonstrate that the compression method proposed has Help retain the more main details of image.The present invention can also have other various embodiments, without departing substantially from spirit of that invention and its In the case of essence, those skilled in the art make various corresponding change and deformations in accordance with the present invention, but these are corresponding Change and distortion should all belong to the protection domain of appended claims of the invention.

Claims

1. the method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things, it is characterised in that：The method has Body is crossed referred to as：

Step 2: optimum prediction direction is calculated separately to the image block after segmentation, the optimum prediction side of image block after being divided To；

Step 3: weighting directional interpolation filter coefficient by calculating, directional interpolation is weighted to fractional samples value, is inserted It is worth image block；

Step 4: the optimum prediction direction obtained using step 2, to interpolation image block promoted based on direction small respectively Wave conversion obtains the image block after each transformation, i.e., the code block after each transformation；

Step 6: being encoded to the changing image that step 5 obtains using improved SPIHT methods, encoded images are obtained.

2. the method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things according to claim 1, special Sign is：Image block segmentation, the image block after being divided are carried out to remote sensing image in the step 1；Detailed process is：

By Remote Sensing Image Segmentation at identical piece of size, the image block after being divided, block here divides size, should with below The block size of coding stage is consistent.

3. the method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things according to claim 2, special Sign is：Optimum prediction direction is calculated separately to the image block after segmentation in the step 2, image block is most after being divided Good prediction direction；Detailed process is：

Assuming that set of reference directions is θ_ref, set of reference directions θ_refIncluding 15 reference directions, be denoted as -7, -6, -5, -4, -3, - 2,-1,0,1,2,3,4,5,6,7}；If the image block sum after segmentation is N_a, each image block is B_l, l=0,1 ..., N_a-1；

Image block B after segmentation_lRespectively along all reference direction θ_ref,i(i=1,2 ..., 15) carry out direction prediction, obtain Prognostic chart picture block under all reference directions；

Under mean-square error criteria, prognostic chart picture block pixel under all reference directions respectively with step 1 remote sensing image pixel phase Compare, when error minimum corresponding reference direction, the as optimum prediction direction of the prognostic chart picture block

In formula, D () is image fault function, and x (m, n) is image block B_lThe corresponding sample value of middle position (m, n), DA_PⁱIt is The predictive operator of i reference direction, m are the abscissa of corresponding position, and n is the ordinate of corresponding position；Enable D ()=| |；

It repeats the above process, the optimum prediction direction of image block after determining all segmentations.

4. the method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things according to claim 3, special Sign is：Directional interpolation filter coefficient is weighted by calculating in the step 3, being weighted direction to fractional samples value inserts Value, obtains interpolation image block；Detailed process is：

The final output of directional interpolation filter is determined by three filters, is respectively：Bi-linear filter, Telenor4-tap filters and 2-tap filters；

Below current sample is expert in two rows, two samples in the row arranged with the sample column interval two are taken respectively This, the input as bi-linear filter；In row, the lastrow where current sample, and in lower two rows, the sample is taken respectively Four samples of the next column of column, as the input of Telenor 4-tap filters, bi-linear filter and Telenor The output of 4-tap filters constitutes the input of 2-tap filters, and the output of 2-tap filters is exactly directional interpolation filter Weighting coefficient；

Pass through the input sample of bi-linear filter, the input sample of Telenor 4-tap filters and weighting coefficient structure side To interpolation filter, directional interpolation filter is weighted directional interpolation to the sample value of fractional position, obtains interpolation image block.

5. the method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things according to claim 4, special Sign is：The optimum prediction direction obtained using step 2 in the step 4 carries out interpolation image block to be based on direction respectively The wavelet transformation of promotion obtains the image block after each transformation, i.e., the code block after each transformation；Detailed process is：

The optimum prediction direction obtained according to step 2It is utilized respectively formula (2) and interpolation image that (3) obtain step 3 Block carries out the wavelet transformation promoted based on direction：

Direction prediction operator DA_P is

In formula, x_e[m, n] is the even samples set for the interpolation image block that step 3 obtains, DA_P_e[m, n] is even samples collection Close corresponding direction prediction operator；I indicates the serial number of high-pass filter coefficient, p_iIndicate high-pass filter coefficient；

Direction update operator DA_U is

In formula, j indicates the serial number of low-pass filter coefficients, u_jIndicate low-pass filter coefficients, DA_U_d[m, n] is odd samples collection Close corresponding direction update operator；D [m, n] is to be expressed as by the odd samples after neighbouring even-numbered sample predictions

D [m, n]=x_o[m,n]+DA_P_e[m,n]

6. the method for compressing image based on direction Lifting Wavelet and improved SPIHIT under Internet of Things according to claim 5, special Sign is：The changing image that step 5 obtains is encoded using improved SPIHT methods in the step 6, is encoded Image afterwards；Detailed process is：

Step 6 one, initial threshold value T=2^n′, initial table LSP, LIS and LIP；

N ' is the maximum value of bit-plane number；

Table LSP is initialized as an empty table, LIP is initialized as lowest frequency sub-band coefficients location sets, and LIS is initialized as each The root node coordinate set of spatial orientation tree；

Step 6 two encodes LIP according to table LSP, LIS and LIP, and process is：

Step 621, according to whether including significant coefficient in threshold decision LIP set, be that output 1 and sign bit, coefficient are Just, sign bit 0, coefficient are negative, and sign bit 1 deletes significant coefficient position (i, j) from LIP, and are added to the ends LSP Tail；

It is no, then export 0；

Step 6 three, coding LIS, process are：

Step 631 judges that the current record of LIS is that D (i, j) or L (i, j) are held if the current record of LIS is D (i, j) Row step 6 three or two executes step 6 three or five if the current record of LIS is L (i, j)；

If including significant coefficient in D (i, j), D (i, j) is decomposed into L (i, j) and O (i, j)；L (i, j) is put into LIS tails Portion；

Quaternary tree is established with 4 coefficients of O (i, j) and is encoded, if the tree root of quaternary tree is more than or equal to threshold value, illustrates that this four are There are significant coefficient, output 1 in number；Otherwise, if the tree root of quaternary tree is less than threshold value, there is no significant coefficient in four coefficients, export 0；Execute step 6 three or three；

Significant coefficient is put into LIP or LSP by step 6 three or three, and exports the symbol of significant coefficient, and significant coefficient is just symbol It is 0, significant coefficient is negative, symbol 1；Execute step 6 three or four；

Step 6 three or four judges whether L (i, j) is sky,

If so, deleting D (i, j) from LIS；Execute step step 6 three or eight；

Step 6 three or five, judge L (i, j) whether tape label, if so, execute step 6 three or six, if it is not, execute step 6 Radix Notoginseng；

Step 6 three or six, according in threshold decision L (i, j) whether include significant coefficient, be that L (i, j) is important, is deleted from LIS D (2i, 2j), D (2i+1,2j), D (2i, 2j+1) and D (2i+1,2j+1) are added to the end of LIS, not exported by L (i, j) Any information；Execute step step 6 three or eight；

No, L (i, j) is inessential, executes step step 6 three or eight；

Step 6 Radix Notoginseng, according in threshold decision L (i, j) whether include significant coefficient, be that L (i, j) is important, then by L (i, j) It is deleted from LIS, D (2i, 2j), D (2i+1,2j), D (2i, 2j+1) and D (2i+1,2j+1) is added to the end of LIS, And exports coding；Execute step step 6 three or eight；

No, L (i, j) is inessential, then does not export any information；

Step 6 three or eight；Judge whether the root node coordinate of all spatial orientation trees in LIS has been processed, if it is not, Re-execute step 631；If executing step 6 four；

Step 6 four, the label for removing all L (i, j) check each (i, j) in LSP, if not newly adding in sequence scanning , the nth bit of the position coefficient of correspondence is exported, step 6 five is executed；If newly adding in sequence scanning, then do not export Any information；

Step 6 five judges whether the length of compressed bit stream reaches specified length, if so, output compressed bit stream；T=if not T/2 executes step 6 two.