CN103327337A - Classification quantization coding method based on bi-orthogonal lapped transform - Google Patents

Abstract

The invention provides a classification quantization coding method based on bi-orthogonal lapped transform, and belongs to the technical field of remote sensing image data transmission. According to the classification quantization coding method, the step of coding parameter designing is completed after image sequences are tested, and then the step of realizing an image coding algorithm to input images is conducted. The images are classified into different types according to coding characteristics through the method of classification training, an appropriate quantization method is selected for each type, therefore, the problem that image coding performances are different due to the fact that on the premise of quality-fixed coding, different types of remote sensing images have different requirements for coding performance parameters is solved, and the purpose of objective-quality-fixed coding of different types of images is achieved. The classification quantization coding method is applied to remote sensing satellites and greatly improves the downloading efficiency of remote sensing satellite image data.

Description

A kind of classification quantitative coding method based on the biorthogonal lapped transform

Technical field

The invention belongs to the remote sensing image data transmission technique field, be specifically related to a kind of classification quantitative coding method based on the biorthogonal lapped transform.

Background technology

Because the needs of observation, satellite sensor need to be transmitted back to ground by the image that observation system is captured.Along with user's demand, the resolution of image will be more and more higher, and this will cause the sharp increase of image data amount, and the data link channel capacity of at present space communication is limited.For so that ground can receive high-quality image, solve the contradiction of the transmission of remote sensing images big data quantity and limited interchannel by the Remote Sensing Image Coding method, be necessary.

Focus in the present Remote Sensing Image Coding method based on the embedded encoded method of discrete cosine transform and wavelet transformation.But in conventional method, the lifting of coding efficiency tends to be accompanied by the increase of complexity, and this brings more requirement with regard to the hardware designs of giving coded system.And in satellite system, hardware device all has a lot of restrictions on operational capability, internal memory and power consumption.Therefore, conventional method is difficult to satisfy the application demand of hardware system Real-time Collection transmission.Image coding technique based on the biorthogonal lapped transform provides a kind of effective method for realizing low complex degree, high-performance image coding, and it is adopted by JPEG XR standard.

In the remote sensing earth observation, widely different between the atural object in traditional coding method, carries out coding transmission with all cartographic features according to identical mode often.Therefore, the image of texture-rich tends to have more distortion than the simple image of texture.And for the user, tend to wish that various types of images all can satisfy quality requirement, and namely having similar coding distortion, this has just proposed new demand to method for encoding images.

The people such as Li have set up a kind of forecast model for describing the JPEG2000 coding quality by analysing in depth contacting between image liveness and the Image Coding performance.(referring to document: Ling Li and Zhen-Song Wang, Compression Quality Prediction Model for JPEG2000, IEEE Transactions on Image Processing, 2010) further can find, dissimilar image has different coding characteristics, and this feature can be weighed by the image liveness.

Summary of the invention

Problem for the background technology existence, the present invention proposes a kind of classification quantitative coding method based on the biorthogonal lapped transform, but efficient solution remote sensing images never of the same type are in the Image Coding performance difference problem of deciding under the quality coded prerequisite coding efficiency parameter request difference to be caused, thereby have reached the purpose that dissimilar image is decided quality coded.

For solving the problems of the technologies described above, the present invention adopts following technical scheme:

A kind of classification quantitative coding method based on the biorthogonal lapped transform comprises the steps,

Step 1, finish the design of coding parameter by the test pattern sequence;

Step 2, the coding parameter that utilizes step 1 to obtain are encoded to input picture.

Described step 1 specifically may further comprise the steps,

Step 1.1, test pattern is carried out the biorthogonal lapped transform;

Transform method adopts the transform method of JPEG XR standard, is specially: at first test pattern is divided into size and is 16 * 16 macro block, 4 * 4 piece carries out the biorthogonal lapped transform first time as the unit in each macro block; After the conversion, upper left angular data is the DC coefficient in each 4 * 4, and remaining 15 is the HP coefficient; The piece of DC coefficient compositions 4 * 4 all in 16 * 16 macro blocks is carried out the biorthogonal lapped transform second time, finally obtain a DC coefficient and 15 LP coefficients; Data after the final conversion of each macro block of 16 * 16 comprise 1 DC coefficient, 15 LP coefficients and 240 HP coefficients;

Step 1.2, the DC coefficient to behind each macroblock transform, LP coefficient and HP coefficient quantize, and quantitative formula is:

$B_{\mathrm{ij}}=\mathrm{round}\left(\frac{A_{\mathrm{ij}}}{Q_{\mathrm{step}}}\right)$

Wherein, A _IjCoefficient behind the representation transformation, B _IjCoefficient after representative quantizes, Q _StepRepresent quantization step, initial quantization step is set to 1, round () and represents to round up computing;

Step 1.3, to B _IjCarry out inverse quantization, inverse quantization formula is:

Ci _j＝Bi _j×Q _step

Wherein, C _IjRepresent the dequantized coefficients of macro block, i, j represent respectively C _IjThe row, column coordinate;

Step 1.4, according to step 1.1 couple C _IjCarry out the inverse transformation of biorthogonal lapped transform, obtain reconstructed image;

The Y-PSNR (Peak Signal to Noise Ratio, PSNR) of step 1.5, calculating test pattern and reconstructed image, computing formula is:

$\mathrm{PSNR}=201g\sqrt{\frac{{255}^2}{\mathrm{MSE}}},$ $\mathrm{MSE}=\frac{1}{M\×N}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{[I(i,j)-\tilde{I}(i,j)]}^2$

Wherein, M, N be length and the width of representative image respectively, I (i, j) and

Represent respectively the pixel size of original image and reconstructed image, i, j represent respectively the row, column coordinate of original image and reconstructed image;

Step 1.6, to establish expection image objective quality be T _QIf, PSNR-T _Q＞0.5, Q then _Step=Q _Step+ 1, then repeating step 1.2 is to step 1.6; If PSNR-T _Q＜-0.5, Q then _Step=Q _Step-1, then repeating step 1.2 is to step 1.6; Other situations are finishing iteration computational process then;

Step 1.7, determine quantization step by iterative process, according to quantization step image is divided into dissimilarly, and set the quantization step of each type;

Step 1.8, computed image liveness comprise IAMD1, IAMD2, IAME1 and IAME2, and with it characteristic of division as image, the classification type in construction feature vector and the integrating step 1.7 is finished the training of SVMs;

Specific formula for calculation is as follows:

$\mathrm{IAMD}1=\frac{1}{(M-1)\×N}\underset{i=0}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}|x(i,j)-x(i+1,j)|$

$+\frac{1}{M\×(N-1)}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j)-x(i,j+1)|$

$\mathrm{IAMD}2=\frac{1}{(M-1)\×(N-1)}\underset{i=1}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j)-x(i-1,j+1)|$

$+\frac{1}{(M-1)\×(N-1)}\underset{i=0}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j)-x(i+1,j+1)|$

$\mathrm{IAME}1=\frac{1}{(M-2)\×N}\underset{i=1}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}|x(i-1,j)-x(i+1,j)|$

$+\frac{1}{M\×(N-2)}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=1}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j-1)-x(i,j+1)|$

$\mathrm{IAME}2=\frac{1}{(M-2)\×(N-2)}\underset{i=1}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=1}{\overset{N-2}{\mathrm{\Σ}}}|x(i-1,j-1)-x(i+1,j+1)|$

$+\frac{1}{(M-2)\×(N-2)}\underset{i=1}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=1}{\overset{N-2}{\mathrm{\Σ}}}|x(i+1,j-1)-x(i-1,j+1)|$

X (i, j) represents the pixel size of original image, and i, j represent respectively the row, column coordinate of original image.

Described step 2 specifically may further comprise the steps,

Step 2.1, the utilization method identical with step 1.1 are carried out the biorthogonal lapped transform to input picture;

The liveness of the method computed image that step 2.2, utilization are identical with step 1.8;

Step 2.3, combining image liveness make up the characteristic vector that is used for classification, and the classification based training result of integrating step 1.8 determines the type of input picture by SVMs, further determine quantization step;

Step 2.4, quantize according to the coefficient of quantization step after to input picture biorthogonal lapped transform;

Step 2.5, the coefficient after quantizing is carried out the entropy coding.

Compared with prior art, but the present invention's efficient solution remote sensing images never of the same type in the Image Coding performance difference problem of deciding under the quality coded prerequisite coding efficiency parameter request difference to be caused, thereby reached the purpose that dissimilar image is decided objective quality coding.Be used for remote sensing satellite and use, can greatly improve the lower transfer efficiency of remote sensing satellite image data.

Description of drawings

Fig. 1 is flow chart of the present invention;

Fig. 2 (a) is the typical test pattern in the category-A after the classification;

Fig. 2 (b) is the typical test pattern in the category-B after the classification;

Fig. 2 (c) is the typical test pattern in the C class after the classification;

Fig. 3 (a) is input picture Road;

Fig. 3 (b) is input picture Farmland;

Fig. 3 (c) is input picture Meadow;

Fig. 3 (d) is input picture City;

Fig. 4 (a) is decompressed image Road;

Fig. 4 (b) is decompressed image Farmland;

Fig. 4 (c) is decompressed image Meadow;

Fig. 4 (d) is decompressed image City;

Embodiment

Embodiment 1:

(1) set up the test pattern storehouse, picture number is 15 width of cloth;

(2) setting the objective image quality PSNR that expects is 35dB;

(3) calculate the image liveness value of test pattern, and by behind the biorthogonal lapped transform through after quantizing inverse transformation, near the quantization parameter value in the time of 35dB, its result is as shown in table 1 respectively;

Table 1

(4) resemble 1,2,3,4,5,11,12,13,14,15 according to large young pathbreaker's resolution chart of quantization parameter and be defined as category-A, test image 6,15 is defined as category-B, and test image 7,8,9,10 is divided into the C class;

(5) the image liveness value of different test patterns is used for the training of SVMs as characteristic value;

(6) quantization step of category-A test image being got average is 47, and it is that to get average be 66 for the quantization step of 59, C class testing image that the quantization step of category-B test image is got average;

(7) input picture comprises dissimilar image Road, Farmland, Meadow, City four width of cloth, and its liveness calculated value is as shown in table 2 respectively;

Table 2

Image	IAMD1	IAMD2	IAME1	IAME2
					Road	21.7111	27.7085	33.7382	38.3088
Farmland	7.9463	13.0992	12.4334	19.5948
					Meadow	8.8276	12.1177	13.4544	16.9056
City	36.7683	48.1544	52.3366	62.4131

(8) according to the training result of SVMs in image liveness value and the integrating step 5, by SVMs input picture is classified.Classification results is: Road and City are divided into category-A; Meadow is divided into category-B; Farmland is divided into the C class;

(9) input picture is carried out biorthogonal lapped transform and quantification one by one, quantization step selects respectively 47,66,59 and 47.

(10) data after quantizing are finished the entropy coding.Entropy coding method has adopted the adaptive arithmetic code method.

(11) assess for the validity to coding method, the decoding data to behind the entropy coding obtains decoded reconstructed image.Weigh the difference of input picture and reconstructed image by PSNR, its result is respectively 35.45,35.06,35.02 and 34.53, can find out that worst error is no more than 1dB, has reached the purpose that the objective quality coding is decided in expection.

Claims (3)

1. classification quantitative coding method based on the biorthogonal lapped transform is characterized in that: comprises the steps,

Step 1, finish the design of coding parameter by the test pattern sequence;

Step 2, the coding parameter that utilizes step 1 to obtain are encoded to input picture.

2. a kind of classification quantitative coding method based on the biorthogonal lapped transform as claimed in claim 1, it is characterized in that: described step 1 specifically may further comprise the steps,

Step 1.1, test pattern is carried out the biorthogonal lapped transform;

Transform method adopts the transform method of JPEG XR standard, is specially: at first test pattern is divided into size and is 16 * 16 macro block, 4 * 4 piece carries out the biorthogonal lapped transform first time as the unit in each macro block; After the conversion, upper left angular data is the DC coefficient in each 4 * 4, and remaining 15 is the HP coefficient; The piece of DC coefficient compositions 4 * 4 all in 16 * 16 macro blocks is carried out the biorthogonal lapped transform second time, finally obtain a DC coefficient and 15 LP coefficients; Data after the final conversion of each macro block of 16 * 16 comprise 1 DC coefficient, 15 LP coefficients and 240 HP coefficients;

Step 1.2, the DC coefficient to behind each macroblock transform, LP coefficient and HP coefficient quantize, and quantitative formula is:

$B_{\mathrm{ij}}=\mathrm{round}\left(\frac{A_{\mathrm{ij}}}{Q_{\mathrm{step}}}\right)$

Wherein, A _IjCoefficient behind the representation transformation, B _IjCoefficient after representative quantizes, Q _StepRepresent quantization step, initial quantization step is set to 1, round () and represents to round up computing;

Step 1.3, to B _IjCarry out inverse quantization, inverse quantization formula is:

C _ij＝B _ij×Q _step

Wherein, C _IjRepresent the dequantized coefficients of macro block, i, j represent respectively C _IjThe row, column coordinate;

Step 1.4, according to step 1.1 couple C _IjCarry out the inverse transformation of biorthogonal lapped transform, obtain reconstructed image;

The Y-PSNR (Peak Signal to Noise Ratio, PSNR) of step 1.5, calculating test pattern and reconstructed image, computing formula is:

$\mathrm{PSNR}=201g\sqrt{\frac{{255}^2}{\mathrm{MSE}}},$ $\mathrm{MSE}=\frac{1}{M\×N}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{[I(i,j)-\tilde{I}(i,j)]}^2$

Wherein, M, N be length and the width of representative image respectively, I (i, j) and

Represent respectively the pixel size of original image and reconstructed image, i, j represent respectively the row, column coordinate of original image and reconstructed image;

Step 1.6, to establish expection image objective quality be T _QIf, PSNR-T _Q＞0.5, Q then _Step=Q _Step+ 1, then repeating step 1.2 is to step 1.6; If PSNR-T _Q＜-0.5, Q then _Step=Q _Step-1, then repeating step 1.2 is to step 1.6; Other situations are finishing iteration computational process then;

Step 1.7, determine quantization step by iterative process, according to quantization step image is divided into dissimilarly, and set the quantization step of each type;

Step 1.8, computed image liveness comprise IAMD1, IAMD2, IAME1 and IAME2, and with it characteristic of division as image, the classification type in construction feature vector and the integrating step 1.7 is finished the training of SVMs;

Specific formula for calculation is as follows:

$\mathrm{IAMD}1=\frac{1}{(M-1)\×N}\underset{i=0}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}|x(i,j)-x(i+1,j)|$

$+\frac{1}{M\×(N-1)}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j)-x(i,j+1)|$

$\mathrm{IAMD}2=\frac{1}{(M-1)\×(N-1)}\underset{i=1}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j)-x(i-1,j+1)|$

$+\frac{1}{(M-1)\×(N-1)}\underset{i=0}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j)-x(i+1,j+1)|$

$\mathrm{IAME}1=\frac{1}{(M-2)\×N}\underset{i=1}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}|x(i-1,j)-x(i+1,j)|$

$+\frac{1}{M\×(N-2)}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=1}{\overset{N-2}{\mathrm{\Σ}}}|x(i,j-1)-x(i,j+1)|$

$\mathrm{IAME}2=\frac{1}{(M-2)\×(N-2)}\underset{i=1}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=1}{\overset{N-2}{\mathrm{\Σ}}}|x(i-1,j-1)-x(i+1,j+1)|$

$+\frac{1}{(M-2)\×(N-2)}\underset{i=1}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=1}{\overset{N-2}{\mathrm{\Σ}}}|x(i+1,j-1)-x(i-1,j+1)|$

Wherein, x (i, j) represents the pixel size of original image, and i, j represent respectively the row, column coordinate of original image.

3. a kind of classification quantitative coding method based on the biorthogonal lapped transform as claimed in claim 1 or 2, it is characterized in that: described step 2 specifically may further comprise the steps,

Step 2.1, the utilization method identical with step 1.1 are carried out the biorthogonal lapped transform to input picture;

The liveness of the method computed image that step 2.2, utilization are identical with step 1.8;

Step 2.3, combining image liveness make up the characteristic vector that is used for classification, and the classification based training result of integrating step 1.8 determines the type of input picture by SVMs, further determine quantization step;

Step 2.4, quantize according to the coefficient of quantization step after to input picture biorthogonal lapped transform;

Step 2.5, the coefficient after quantizing is carried out the entropy coding.

Images