CN101192408A - Method and device for selecting conductivity coefficient vector quantization - Google Patents

Abstract

The invention discloses a method and a device for quantifying the vectors of the induced spectrum frequency (ISF) coefficients, which includes the steps as follows: the ISF coefficients are divided into at least two groups of sub-vectors; the inter-frame strong- and weak-correlation vectors of the induced spectrum are quantified for each group of sub-vectors; the residual signals are calculated after the strong- and weak-correlation vectors are quantified; the quantification errors of the above residual signals are also calculated; the quantification errors of the strong- and weak-correlation residual signals are judged for each group of sub-vectors; and the ISF coefficient vectors of the correlation manner with less quantification error in each group are quantified. The invention is based on the short-time correlation and the long-time correlation of the induced spectrum frequency coefficients, and different forecast coefficients are hereby applied to different dimension of the induced spectrum frequency coefficients, thus increasing forecast gains, narrowing the variation range of the residual signals, enhancing the accuracy for the ISF coefficients quantification, achieving clear quantification, and improving the quality for broadband phonetic coding.

Description

Select the method and the device of conductivity coefficient vector quantization

Technical field

The present invention relates to the digital speech code technical field, particularly select to lead the method and the device of spectral frequency (ISF) vector quantization of coefficient.

Background technology

The linear predictor coefficient of voice signal (LPC) can characterize voice signal amplitude spectrum envelope in short-term more accurately, in current, in the low rate voice coding, be applied: as Code Excited Linear Prediction (CELP, Code Excitation Linear Prediction), Multi-Band Excitation (MBE, MultibandExcitation Coding), waveform interpolation coding (WI, Waveform Interpolation) etc., but the LPC coefficient has the dynamic range of broad and the instability in composite filter, thereby the LPC coefficient is not suitable for direct quantification.At the problems referred to above, the representation that needs LPC is converted to other in the quantitative research of speech manual envelope quantizes the stable of back all-pole filter to guarantee linear predictor coefficient, reduces spectrum distortion.In wideband speech coding, usually the LPC coefficient is converted to and leads spectral frequency (ISF) and quantize as AMR-WB, AMR-WB+ etc., because ISF index variation scope is easy to quantize for a short time, and an error of leading the spectral frequency parameter only influences the speech manual that is close to this parameter respective frequencies place in the all-pole modeling, and do not influence other place, can satisfy in high-quality voice coding and the speech synthesis system requirement to LPC coefficient requirement " transparent quantification ".

For convenience of description, simply introduce earlier several basic conceptions:

Linear predictor coefficient (LPC, Linear Predictive Coefficient), one group of succinct voice signal model parameter, this group parameter has characterized the spectrum amplitude of voice signal more accurately.

Lead spectral frequency to (ISP, Immittance Spectral Pair), a kind of form of expression of linear predictor coefficient.

Lead spectral frequency (ISF, Immittance Spectrum Frequency), a kind of form of expression of linear predictor coefficient.

Vector quantization (VQ, Vector Quantization) constitutes a vector with several scalar datas, then vector space is given whole the quantification.

Below be example with ISF coefficient quantization scheme in the AMR-WB+ coding standard [Dec 2004 for 3GPP TS 26.290, " Extended AdaptiveMulti-Rate-Wideband (AMR-WB+) codec "], specify existing ISF coefficient vector method.

1. coding side

Each frame (256 point) is asked for the linear predictor coefficient a on one group of 16 rank ₁, a ₂..., a ₁₆, promptly the LPC coefficient is converted to 16 rank ISP coefficients with above-mentioned LPC coefficient, then the ISP coefficient is converted to 16 rank ISF coefficient isf of frequency domain ₁, isf ₂..., isf ₁₆As current quantification input, suppose isf ' ₁, isf ' ₂..., isf ' ₁₆Be the ISF coefficient on previous frame 16 rank, mean[i], i=1,2,3 ...., 16 is the average of each n dimensional vector n.

(1) calculates one group of residual signals r _i:

$r_i=\mathrm{is}{f}_i-\mathrm{mean}\left[i\right]-\frac{1}{3}\×\mathrm{is}{f}_i^{\′},i=1,.....,16$

(2) above-mentioned 16 rank residual signals are split into 2 sub-vector V Q1 and VQ2, VQ1 comprises r ₁, r ₂..., r ₉, VQ2 comprises r ₁₀, r ₁₁..., r ₁₆

(3) at above-mentioned sub-vector VQ1, use 8 bits to quantize in this example, the signal after supposing to quantize is

Calculate the residual signals e of one group of 9 dimension _i:

$e_i=r_i-{\hat{r}}_i,i=\mathrm{1,2},...,9$

(4) at above-mentioned residual signals e _i, it is split into VQ3, these 3 sub-vectors of VQ4 and VQ5.Wherein VQ3 comprises e ₁, e ₂, e ₃VQ4 comprises e ₄, e ₅, e ₆VQ5 comprises e ₇, e ₈, e ₉To VQ3, VQ4 and VQ5 carry out vector quantization respectively, and used bit number is 6,7 and 7 in this example.

(5) at sub-vector VQ2, use 8 bits to quantize, calculate the residual signals e of one group of 7 dimension _i':

${e_i}^{\′}=r_i-{\hat{r}}_i,i=\mathrm{10,11},...,16$

(6) at above-mentioned residual signals e _i', it is split into VQ6 and VQ7.Wherein VQ6 comprises e ₁₀, e ₁₁, e ₁₂Wherein VQ6 comprises e ₁₃, e ₁₄, e ₁₅, e ₁₆VQ6 and VQ7 are carried out vector quantization respectively, and used bit number is 5 and 5 in this example.

2. decoding end

In decoding end, according to code index value search code book, rebuild the ISF vector after VQ1 and VQ2 quantize, and then the inverse process of carrying out coding side (2) and (1) finally obtains the ISF coefficient rebuild

As everyone knows, the ISF coefficient self has some unique characteristics: lead spectral frequency (ISF) coefficient and be the ascending order arrangement, for strong voiced speech section, because such voice segments duration is longer, exist stronger relevant redundancy between the spectral frequency coefficient leading of adjacent speech frame, the predicted value set of reaction strong correlation selected for use in this section voice is a kind of inevitable choice; For voiceless sound and weak voiced speech section, adjacent speech frame lead between the spectral frequency coefficient correlativity a little less than, at this moment need to select to reflect weak relevant predicted value set.

And prior art is not considered spectral frequency (ISF) coefficient own characteristic, does not distinguish situation strong, weak correlativity, and quantization method is fairly simple, causes quantization error bigger, has damaged the wideband speech coding quality.

Summary of the invention

In order to improve ISF coefficient quantization precision, and then improve the wideband speech coding quality, the invention provides a kind of method and device of the ISF of selection vector quantization of coefficient.

The method of spectral frequency ISF vector quantization of coefficient is led in a kind of selection, comprising:

The ISF coefficient vector is split at least two group sub-vectors;

Carry out interframe respectively at each group sub-vector and lead spectrum strong correlation, weak dependent vector and quantize, calculate the residual signals after strong, the weak dependent vectorization, and calculate the quantization error of described strong, weak related residual signal;

Judge the size of the quantization error of the quantization error of strong correlation residual signals of each group sub-vector and weak related residual signal respectively, carry out the ISF vector quantization of coefficient at the relevant mode that this group selection quantization error is little.

The device of spectral frequency ISF vector quantization of coefficient is led in a kind of selection, comprises ISF coefficient vector division unit, and the quantization error computing unit of residual signals is differentiated selected cell and ISF vector quantization of coefficient index output unit, wherein,

ISF coefficient vector division unit is used for the ISF coefficient vector is split at least two group sub-vectors;

The quantization error computing unit of residual signals is used for that each group sub-vector is carried out interframe respectively and leads spectrum strong correlation, weak dependent vector and quantize, and calculates the residual signals after strong, the weak dependent vectorization, calculates the quantization error of described strong, weak related residual signal afterwards;

Differentiate selected cell, be used to judge the size of quantization error of the quantization error of strong correlation residual signals of each group sub-vector and weak related residual signal after, carry out the ISF vector quantization of coefficient at the relevant mode that this group selection quantization error is little;

ISF vector quantization of coefficient index output unit is used for exporting respectively every group of ISF vector quantization of coefficient index.

The method and the device of above-mentioned selection ISF vector quantization of coefficient, by the ISF coefficient vector being split at least two group sub-vectors, correlation properties during based on short-term correlation of leading the spectral frequency coefficient and length, two predictive coefficient set of the weak relevant and strong correlation that training is corresponding respectively, each dimension of leading the spectral frequency coefficient is used different predictive coefficients respectively, improved prediction gain, reduced the dynamic change scope of residual signals, thereby reduced quantization error, improved ISF coefficient quantization precision, obtain transparent quantification, thereby improved the wideband speech coding quality.

Description of drawings

Fig. 1 is a realization flow synoptic diagram according to an embodiment of the invention;

Fig. 2 is two according to an embodiment of the invention prediction division formula multi-stage vector quantization synoptic diagram;

Fig. 3 is the apparatus structure synoptic diagram of selecting to lead spectral frequency ISF vector quantization of coefficient according to an embodiment of the invention.

Embodiment

Coding side is split at least two group sub-vectors with the ISF coefficient vector; Carry out interframe respectively at each group sub-vector and lead spectrum strong correlation, weak dependent vector and quantize, calculate the residual signals after strong, the weak dependent vectorization, and calculate the quantization error of described strong, weak related residual signal; Whether the quantization error of strong correlation residual signals of judging each group sub-vector respectively is smaller or equal to the quantization error of weak related residual signal, if, then carry out the ISF vector quantization of coefficient at this group selection strong correlation mode, otherwise, carry out the ISF vector quantization of coefficient at relevant mode a little less than this group selection; Export ISF vector quantization of coefficient index respectively for every group.Decoding end is rebuild two groups of sub-vectors that coding side divided according to ISF vector quantization of coefficient indexed search code book, according to the strong or weak relevant mode of obtaining from coding side, carries out the inverse process of coding side, rebuilds the ISF coefficient.

Below in conjunction with drawings and the specific embodiments the present invention is described in further details.

Referring to Fig. 1 and Fig. 2, in the present embodiment, each frame (256 point) is asked for the linear predictor coefficient a on one group of 16 rank ₁, a ₂..., a ₁₆, above-mentioned LPC coefficient is converted to 16 rank ISP coefficients, then the ISP coefficient is converted to 16 rank ISF coefficient isf of frequency domain ₁, isf ₂..., isf ₁₆As current quantification input, suppose isf ' ₁, isf ' ₂..., isf ' ₁₆Be the ISF coefficient on previous frame 16 rank, mean[i], i=1,2,3 ...., 16 is the average of each n dimensional vector n.The coding side specific operation process is as follows:

Step 101 is split into 2 sub-vector V Q1 and VQ2 with above-mentioned 16 rank ISF coefficients, and VQ1 comprises isf ₁, isf ₂..., isf ₉, VQ2 comprises isf ₁₀, isf ₁₁..., isf ₁₆

Step 102～104 are carried out interframe respectively at VQ1 and are led spectrum strong correlation, weak dependent vector and quantize, and calculate the residual signals after strong, the weak dependent vectorization, and calculate the quantization error of described strong, weak related residual signal.Be specially:

I) at sub-vector VQ1, suppose that interframe leads spectral frequency coefficient strong correlation, select one group of optimum prediction coefficient p ₁, p ₂..., p ₉, calculate one group of 9 residual signals r that ties up _i:

r _i＝isf _i-mean[i]-p _i×isf′ _i i＝1，2，....，9

Ii) at above-mentioned residual signals r _i, utilize 8 bits to quantize, the signal after supposing to quantize is

Calculate the residual signals e of one group of 9 dimension _i:

$e_i=r_i-{\hat{r}}_i,i=\mathrm{1,2},...,9$

Iii) at above-mentioned residual signals e _i, it is split into VQ3, VQ4 and VQ5.Wherein VQ3 comprises e ₁, e ₂, e ₃VQ4 comprises e ₄, e ₅, e ₆VQ5 comprises e ₇, e ₈, e ₉To VQ3, VQ4 and VQ5 carry out vector quantization respectively, and used bit number is 6,7 and 7 in the present embodiment.Suppose that the signal after VQ3 quantizes is

Signal after VQ4 quantizes is Signal after VQ6 quantizes is

Calculating is to residual signals e _iQuantization error error1:

$\mathrm{error}1=\underset{i=1}{\overset{9}{\mathrm{\Σ}}}{(e_i-{\hat{e}}_i)}^2$

Iv), suppose to be correlated with a little less than interframe is led the spectral frequency coefficient, select another group optimum prediction coefficient p ' at sub-vector VQ1 ₁, p ' ₂..., p ' ₉, calculate one group of 9 residual signals r ' that ties up _i:

r′ _i＝isf _i-mean[i]-p _i×isf′ _i i＝1，2，....，9

V) repeating step (ii) and (iii) calculates and quantizes error e rror2.

Illustrate that a bit it is still weak relevant to lead spectral frequency coefficient strong correlation for first hypothesis interframe, the not influence of integral body to calculating that is to say above-mentioned steps i)～iii) and step I do not have strict time sequencing between v)～v).

Whether step 105, the quantization error of judging sub-vector strong correlation residual signals are whether error1 is smaller or equal to error2, if then execution in step 106, otherwise execution in step 107 smaller or equal to the quantization error of weak related residual signal.

Step 106 selects the strong correlation mode to carry out the ISF vector quantization of coefficient at VQ1, finishes the vector quantization to VQ1.

Step 107 selects weak relevant mode to carry out the ISF vector quantization of coefficient at VQ1, finishes the vector quantization to VQ1.

Step 108～110 are carried out interframe respectively at VQ2 and are led spectrum strong correlation, weak dependent vector and quantize, and calculate the residual signals after strong, the weak dependent vectorization, and calculate the quantization error of described strong, weak related residual signal.Be specially:

Vi) at sub-vector VQ2, suppose that interframe leads spectral frequency parameter strong correlation, select one group of optimum prediction coefficient p ₁₀, p ₁₁..., p ₁₆, calculate one group of 7 residual signals r that ties up " _i:

r″ _i＝isf _i-mean[i]-p _i×isf′ _i i＝10，11，....，16

Vii) at above-mentioned residual signals r " _i, utilize 8 bits to quantize in this example, the signal after supposing to quantize is Calculate the residual signals e ' of one group of 7 dimension _i:

$e_i^{\′}=r_i^{\′\′}-{\hat{r}}_i^{\′},i=\mathrm{10,11},...,16$

Viii) at above-mentioned residual signals e ' _i, it is split into VQ6 and VQ7.Wherein VQ6 comprises e ' ₁₀, e ' ₁₁, e ' ₁₂VQ7 comprises e ' ₁₃, e ' ₁₄, e ' ₁₅, e ' ₁₆VQ6 and VQ7 are carried out vector quantization respectively, and used bit number is 5 and 5.Suppose that the signal after VQ3 quantizes is

Signal after VQ7 quantizes is

Calculating is to residual signals e ' _iQuantization error error3:

$\mathrm{error}3=\underset{i=10}{\overset{16}{\mathrm{\Σ}}}{(e_i-{\hat{e}}_i)}^2$

Ix) at sub-vector VQ2, suppose to be correlated with a little less than interframe is led the spectral frequency coefficient, select another group optimum prediction coefficient p ' ₁₀, p ' ₁₁..., p ' ₁₆, calculate one group of 7 residual signals of tieing up:

$r_i^{\′\′\′}={\mathrm{isf}}_i-\mathrm{mean}\left[i\right]-p_i\×{\mathrm{isf}}_i^{\′},i=\mathrm{10,11},...,16$

X) repeating step vii) and viii) calculates and quantizes error e rror4.

Illustrate that a bit it is still weak relevant to lead spectral frequency coefficient strong correlation for first hypothesis interframe, to the not influence of integral body of calculating, that is to say above-mentioned steps vi)～viii) and step I x)～do not have strict time sequencing between ix).

Whether step 110, the quantization error of judging sub-vector strong correlation residual signals are whether error3 is smaller or equal to error4, if then execution in step 112, otherwise execution in step 113 smaller or equal to the quantization error of weak related residual signal.

Step 112 selects the strong correlation mode to carry out the ISF vector quantization of coefficient at VQ2, finishes the vector quantization to VQ2.

Step 113 selects weak relevant mode to carry out the ISF vector quantization of coefficient at VQ2, finishes the vector quantization to VQ2.

After finishing, the ISF vector quantization exports ISF vector quantization of coefficient index respectively for every group.

In decoding end, at first decoding obtains present frame VQ1 and the used predictive coefficient of VQ2 vector, then according to quantization index search code book, rebuild the ISF vector after VQ1 and VQ2 quantize, and indicate, carry out the inverse process of coding side according to strong or weak relevant mode from coding side, rebuild the ISF coefficient

For example,, use the strong correlation mode to carry out vector quantization, use weak relevant mode to carry out vector quantization VQ2 to VQ1 if at coding side; Then, equally the VQ1 that rebuilds is carried out the inverse process of strong correlation mode, the VQ2 that rebuilds is carried out the inverse process of weak relevant mode, thereby finally rebuild the ISF coefficient in decoding end.

It is VQ1 and VQ2 that above-mentioned embodiment only is split into the ISF coefficient 2 sub-vectors, it is that example describes that each sub-vector is split into three sub-vectors again, according to actual needs, the ISF coefficient can be split into a plurality of sub-vectors such as 3,4, each sub-vector is as long as be split into two or more sub-vectors again, its specific implementation process repeats no more as hereinbefore.

As seen, embodiment of the present invention is by being split into the ISF coefficient vector at least two group sub-vectors, correlation properties during based on short-term correlation of leading the spectral frequency coefficient and length, two predictive coefficient set of the weak relevant and strong correlation that training is corresponding respectively, each dimension of leading the spectral frequency coefficient is used different predictive coefficients respectively, thereby raising prediction gain, reduced the dynamic change scope of residual signals, improved ISF coefficient quantization precision, obtain transparent quantification, thereby improved the wideband speech coding quality.

The invention also discloses a kind of selection and lead the device of spectral frequency ISF vector quantization of coefficient,, comprise ISF coefficient vector division unit 310 referring to Fig. 3, the quantization error computing unit 320 of residual signals, differentiate selected cell 330 and ISF vector quantization of coefficient index output unit 340, wherein

ISF coefficient vector division unit 310 is used for the ISF coefficient vector is split at least two group sub-vectors;

The quantization error computing unit 320 of residual signals is used for that each group sub-vector is carried out interframe respectively and leads spectrum strong correlation, the quantification of weak dependent vector, calculate the residual signals after strong, the weak dependent vectorization, calculate the quantization error of described strong, weak related residual signal afterwards;

After differentiating the size of quantization error that selected cell 330 is used to judge the quantization error of strong correlation residual signals of each group sub-vector and weak related residual signal, carry out the ISF vector quantization of coefficient at the relevant mode that this group selection quantization error is little;

ISF vector quantization of coefficient index output unit 340 is used for exporting respectively every group of ISF vector quantization of coefficient index.

Device of the present invention may further include: rebuild the ISF coefficient elements, be used for according to ISF vector quantization of coefficient indexed search code book, rebuild two groups of sub-vectors that coding side divided, according to strong or weak relevant mode, carry out the inverse process of vector quantization, rebuild the ISF coefficient.

The above is preferred embodiment of the present invention only, is not to be used to limit protection scope of the present invention.All any modifications of being done within the spirit and principles in the present invention, be equal to replacement, improvement etc., all be included in protection scope of the present invention.

Claims (8)

1. the method for spectral frequency ISF vector quantization of coefficient is led in a selection, it is characterized in that, comprising:

The ISF coefficient vector is split at least two group sub-vectors;

Carry out interframe respectively at each group sub-vector and lead spectrum strong correlation, weak dependent vector and quantize, calculate the residual signals after strong, the weak dependent vectorization, and calculate the quantization error of described strong, weak related residual signal;

Judge the size of the quantization error of the quantization error of strong correlation residual signals of each group sub-vector and weak related residual signal respectively, carry out the ISF vector quantization of coefficient at the relevant mode that this group selection quantization error is little.

2. method according to claim 1, it is characterized in that, the process that the relevant mode that described selection quantization error is little is carried out the ISF vector quantization of coefficient comprises: if the quantization error of strong correlation residual signals is then carried out the ISF vector quantization of coefficient at this group selection strong correlation mode smaller or equal to the quantization error of weak related residual signal; If the quantization error of strong correlation residual signals greater than the quantization error of weak related residual signal, is then carried out the ISF vector quantization of coefficient at relevant mode a little less than this group selection.

3. method according to claim 1 is characterized in that,

The division of described ISF coefficient vector, vector quantization also calculate quantization error and judge that selection operation carried out by coding side;

This method further comprises:

Decoding end is rebuild two groups of sub-vectors that coding side divided according to ISF vector quantization of coefficient indexed search code book, according to the strong or weak relevant mode of obtaining from coding side, carries out the inverse process of coding side, rebuilds the ISF coefficient.

4. according to claim 1 or 3 described methods, it is characterized in that, describedly carry out interframe respectively at each group sub-vector and lead spectrum strong correlation, weak dependent vector and quantize that the process of calculating the residual signals after strong, the weak dependent vectorization comprises:

I) select one group of optimum prediction FACTOR P ₁, P ₂... P _n, n is the dimension of described one group of sub-vector;

Described optimum prediction coefficient obtains according to sample training, and strong, weak described coefficient difference when relevant;

Calculate the residual signals r of this group n dimension _i:

r _i=isf _i-mean[i]-p _i* isf _i', mean[i] be the average of each n dimensional vector n, i=1,2......n;

Ii) to step I) described residual signals quantizes, and the signal after supposing to quantize is

Residual signals e after then quantizing _iFor:

$e_i=r_i-{\hat{r}}_i$ i＝1，2......n。

5. method according to claim 4 is characterized in that, described process at the quantization error of strong, weak related residual signal after each group sub-vector compute vectorsization comprises:

With described residual signals e _iBe split at least two group sub-vectors again, every group of sub-vector comprises two above residual signals at least, every group of sub-vector carried out strong, weak dependent vector more respectively quantize;

If to e _iThe signal that carries out behind the strong correlation vector quantization is

I=1,2......n is then to strong correlation residual signals e _iStrong correlation quantization error error be: $\mathrm{error}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(e_i-{\hat{e}}_i)}^2;$

If to e _iThe signal that carries out after weak dependent vector quantizes is

I=1,2......n is then to weak related residual signal e _iWeak dependent quantization error e rror ' be: ${\mathrm{error}}^{\′}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(e_i-{\hat{e}}_i^{\′})}^2;$

Described n is the dimension of vector.

6. method according to claim 5 is characterized in that,

Described coding side is split into two groups of sub-vector VQ1 and VQ2 with 16 dimension ISF coefficient vectors; And described VQ1 is 9 dimensions, and described VQ2 is 7 dimensions;

At described VQ1, described residual signals e _i, i=1,2......9 is split into three groups of sub-vector VQ3 again, VQ4 and VQ5, wherein, VQ3 comprises e ₁, e ₂, e ₃VQ4 comprises e ₄, e ₅, e ₆VQ5 comprises e ₇, e ₈, e ₉

At described VQ2, described residual signals e _i, i=10,11......16 is split into two groups of sub-vector VQ6 and VQ7 again, and wherein, VQ6 comprises e ₁₀', e ₁₁', e ₁₂'; VQ7 comprises e ₁₃', e ₁₄', e ₁₅', e ₁₆'.

7. the device of spectral frequency ISF vector quantization of coefficient is led in a selection, it is characterized in that, comprises ISF coefficient vector division unit, and the quantization error computing unit of residual signals is differentiated selected cell and ISF vector quantization of coefficient index output unit, wherein,

Described ISF coefficient vector division unit is used for the ISF coefficient vector is split at least two group sub-vectors;

The quantization error computing unit of described residual signals, be used for that each group sub-vector is carried out interframe respectively and lead spectrum strong correlation, the quantification of weak dependent vector, calculate the residual signals after strong, the weak dependent vectorization, calculate the quantization error of described strong, weak related residual signal afterwards;

Described differentiation selected cell, be used to judge the size of quantization error of the quantization error of strong correlation residual signals of each group sub-vector and weak related residual signal after, carry out the ISF vector quantization of coefficient at the relevant mode that this group selection quantization error is little;

Described ISF vector quantization of coefficient index output unit is used for exporting respectively every group of ISF vector quantization of coefficient index.

8. device according to claim 7 is characterized in that, further comprises:

Rebuild the ISF coefficient elements, be used for, rebuild two groups of sub-vectors that coding side divided,, carry out the inverse process of vector quantization, rebuild the ISF coefficient according to strong or weak relevant mode according to ISF vector quantization of coefficient indexed search code book.

Images