CN101365133B - DCT domain interposing and rounding error compensation process - Google Patents

Abstract

The invention discloses a rounding error compensation method for DCT domain interpolation. The method comprises the following steps of (1) inputting interpolation blocks after the DCT domain interpolation; (2) judging the pointing directions of motion vectors in the horizontal and the vertical directions of each pixel of the interpolation blocks, and dividing the interpolation blocks into horizontal interpolation blocks, vertical interpolation blocks and bilateral interpolation blocks; (3) judging the horizontal roughness and the vertical roughness of each interpolation block, and determining the types thereof; (4) determining a DCT direct-current coefficient compensation values corresponding to each type according to the types of the interpolation blocks; (5) respectively conducting DCT direct-current coefficient compensation to the types of the interpolation blocks, and adding the compensated interpolation blocks to residual errors, so as to finish the five steps of the re-establishment of DCT domain blocks. The invention introduces a smoothness judgment of the DCT domain interpolation blocks and gives different compensation values of smooth and rough blocks of the DCT domain, thereby avoiding the phenomenon of error diffusion in image-rebuilding of the DCT domain, obtaining a better rounding compensation effect of bilinear half-pixel interpolation of the DCT domain and improving the image quality of reconstruction of the DCT domain.

Description

A kind of DCT domain interposing and rounding error compensation process

Technical field

The invention belongs to the DCT area image reconstruction technique field in the DCT domain code conversion, relate in particular to a kind of DCT domain interposing and rounding error compensation process.

Background technology

Discrete cosine transform (Discrete Cosine Transform is called for short dct transform) is a kind of mathematical operation that is closely related with Fourier transform, and IDCT is an inverse discrete cosine transform.The DCT domain code conversion does not need DCT and IDCT computing, and computational complexity is lower, and can keep picture quality preferably, thereby comes into one's own.Carrying out the DCT domain code conversion at first need be at DCT territory reconstructed image.The DCT area image is rebuild structure as shown in Figure 2, and basic principle is to utilize the orthogonality of dct transform, and the pixel domain computing that all are linear is motion compensation, interpolation etc. for example, transfers to the DCT territory and carries out.But nonlinear pixel domain computing, for example behind the motion compensated interpolation round off and IDCT after round off, can't find corresponding computing in the DCT territory, cause between DCT territory reconstructed image and the standard reconstruction image to have error.Even more serious is, owing to adopt the video coding algorithm of motion compensation to have recursive nature, the rounding error of each P frame all can be delivered to subsequent frame, and error constantly spreads accumulation along the P frame, causes reconstructed image quality constantly to descend.

Existing DCT territory bilinearity half-pix interposing and rounding error compensation process is by Ping-Hao Wu, Chen Chen and Homer H.Chen are at paper Rounding Mismatch BetweenSpatial-Domain and Transform-Domain Video Codecs.IEEE Trans.CircuitsSyst.Video Technol., vol.16, no.10, pp.1286-1293, Oct.2006. propose in, its basic principle is according to the probability distribution of the rounding error average at DCT domain compensation rounding error.The least unit of motion compensation is a piece, and each piece has the motion vector of oneself, and promptly all pixel motion vector are the same in the piece.The DCT area image of band interposing and rounding error compensation is rebuild structure as shown in Figure 3.

When motion vector points was put in order location of pixels, not needing to carry out interpolation arithmetic did not therefore have rounding error.When motion vector only pointed to half-pixel position on a level or a vertical direction, as the E among Fig. 4, F point, interpolation formula was E=(A+B)/2, and A, B are whole pixel, and then the fractional part of E only may be 0 or 0.5.When E is rounded off computing, the different operations of rounding off will produce different rounding errors, the rounding error sample space of operating that for example rounds up is { 0,0.5}, the rounding error sample space that rounds (getting the maximum integer that is not more than this value) operation downwards is { 0,-0.5} is hereinafter all to be rounded to the example explanation.Two kinds of rounding errors occur with certain probability, suppose that 0 probability of occurrence is q ₀, 0.5 probability of occurrence is q ₁, then the mathematical expectation of rounding error is 0.5q ₁In the DCT territory to each pixel compensation 0.5q ₁Can approach the round-off result of pixel domain better, implementation is very simple simultaneously, promptly is the DC coefficient compensation of the DCT piece of M * N to size

When motion vector all pointed to half-pixel position on the horizontal vertical direction, as the point of the G among Fig. 4, interpolation formula was G=(A+B+C+D)/4.The fractional part of G be 0,0.25,0.5,0.75}, round up the operation rounding error for 0 ,-0.25,0.5,0.25}.Suppose that 0 probability of occurrence is p ₀, 0.25 probability of occurrence is p ₁, 0.5 probability of occurrence is p ₂,-0.25 probability of occurrence is p ₃, then the pixel compensation value is 0.25 (p ₁-p ₃)+0.5p ₂, corresponding D CT DC coefficient offset is $[0.25(p_1-p_3)+0.5p_2]\sqrt{\mathrm{MN}}$ 。

Existing method is thought that rounding error is obeyed and is discretely evenly distributed q ₀=q ₁=0.5, p ₀=p ₁=p ₂=p ₃=0.25, obtain unidirectional interpolation block DC coefficient offset thus and be

, two-way interpolation block DC coefficient offset is

The precondition that the above-mentioned rounding error regularity of distribution is set up is that the adjacent pixel values parity is an independent random.Have smooth region in the video scene, the pixel value in the zone changes mild, and the possibility that adjacent pixel values equates is very high, so rounding error is that 0 probability is bigger.If the smooth region proportion is higher in the video sequence, the prerequisite of above-mentioned distribution hypothesis just can not be set up.If adopt this moment existing method that interpolation block DCT DC coefficient is compensated, will the situation of overcompensation take place.

Summary of the invention

The invention provides a kind of DCT domain interposing and rounding error compensation process, this method is introduced the flatness judgement of DCT territory interpolation block, gives DCT territory smooth block and coarse different offset.

A kind of DCT domain interposing and rounding error compensation process comprises five steps: the interpolation block after the interpolation of (1) input DCT territory; (2) level of each pixel of the interpolation block after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector are divided into horizontal interpolation block, vertical interpolation block and two-way interpolation block with interpolation block; (3) judge the horizontal roughness and vertical roughness of interpolation block, determine the type of interpolation block; (4) determine the DCT DC coefficient offset of each type correspondence according to the type of the interpolation block of being judged; (5) respectively the interpolation block of each type is carried out the compensation of DCT DC coefficient, interpolation block after the compensation and residual error addition are finished the piece in DCT territory and are rebuild.

The level of each pixel of the interpolation block of image and the sensing of movement in vertical direction vector after the interpolation of judgement DCT of the present invention territory, the method that interpolation block is divided into horizontal interpolation block, vertical interpolation block and two-way interpolation block is as follows:

The level of each pixel of the interpolation block after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector, if horizontal motion vectors directed half-pixel position, the movement in vertical direction vectors directed is put in order location of pixels, and the interpolation block that is obtained by two neighbor interpolation of horizontal direction is horizontal interpolation block; If movement in vertical direction vectors directed half-pixel position, the horizontal motion vectors directed is put in order location of pixels, and the interpolation block that is obtained by two neighbor interpolation of vertical direction is vertical interpolation block; If level and movement in vertical direction vector all point to half-pixel position, the interpolation block that is obtained by four neighbor interpolation is two-way interpolation block.

The horizontal roughness of judgement interpolation block of the present invention and vertical roughness, determine that the method for type of interpolation block is as follows:

In order to judge the flatness of interpolation block, this method is introduced horizontal roughness HR and vertical roughness VR, represent the degree of roughness on an interpolation block region horizontal direction and the vertical direction respectively, horizontal roughness HR can be defined by different modes with vertical roughness VR, simultaneously different definition modes is selected different level and smooth threshold value HST of level and vertical level and smooth threshold value VST, and the definite of level and smooth threshold value can adopt existing multiple technologies scheme.

Need judge its flatness in the horizontal direction for horizontal interpolation block, if HR≤HST is judged to be horizontal interpolation smooth block, if HR〉HST, be judged to be coarse of horizontal interpolation; Need judge its flatness in vertical direction for vertical interpolation block, if VR≤VST is judged to be vertical interpolation smooth block, if VR〉VST, be judged to be coarse of vertical interpolation; Need judge its flatness on level and vertical direction simultaneously to two-way interpolation block, if HR≤HST and VR≤VST, be judged to be the two-way smooth block of two-way interpolation, if HR≤HST and VR〉VST, be judged to be two-way interpolation horizontal slide block, if HR〉HST and VR≤VST, be judged to be the vertical smooth block of two-way interpolation, if HR〉HST and VR〉VST, be judged to be coarse of two-way interpolation.

Along with HST and VST value increase, a piece is judged as the probability increase of smooth block.If HST and VST value are too big, can be smooth block with coarse erroneous judgement then; If HST and VST value are too little, then can produce and fail to judge.Therefore the value of HST and VST must limit within the specific limits, could guarantee the reasonability that block type is judged.Coarse rounding error with coarse of vertical interpolation of horizontal interpolation is tending towards being evenly distributed, and utilizes this character can determine reasonably level and smooth threshold value.

Type according to the interpolation block of being judged of the present invention determines that the DCT DC coefficient compensation value method of each type correspondence is as follows:

After determining the sliding threshold value of this group exemplary code levelling, mark is above-mentioned all types of in DCT area image reconstruction structure, and establishing block size is M * N.The ratio of all types of rounding errors of statistics is determined offset according to the ratio of rounding error in the pixel domain decoder.For horizontal interpolation smooth block and vertical interpolation smooth block, establish q ₀Be rounding error 0 proportion, q ₁Be rounding error 0.5 proportion, the pixel compensation value is 0.5q ₁, DCT DC coefficient offset is

For the two-way interpolation smooth block of three classes, suppose p ₀, p ₁, p ₂, p ₃Be respectively 0 ,-0.25,0.5,0.25 proportion, the pixel compensation value is 0.25 (p ₁-p ₃)+0.5p ₂, DCT DC coefficient offset is $[0.25(p_1-p_3)+0.5p_2]\sqrt{\mathrm{MN}}$ Coarse of horizontal interpolation, coarse of vertical interpolation, coarse of two-way interpolation still compensate by existing method, and the pixel compensation value of coarse of horizontal interpolation, coarse of vertical interpolation is 0.25, and DCT DC coefficient offset is

The pixel compensation value that two-way interpolation is coarse is 0.125, and the pixel compensation value is

The inventive method is introduced the flatness judgement of DCT territory interpolation block, give DCT territory smooth block and coarse different offset, suppress the error diffusion phenomenon of DCT area image in rebuilding, obtained the better DCT territory bilinearity half-pix interpolation compensation effect that rounds off, improve DCT territory reconstructed image quality.

Description of drawings

Fig. 1 is a technical scheme flow chart of the present invention;

Fig. 2 is not for rebuilding structural representation with the DCT area image of interposing and rounding error compensation;

Fig. 3 rebuilds structural representation for the DCT area image that adopts existing interposing and rounding error compensation process;

Fig. 4 is a half-pix interpolation schematic diagram;

Fig. 5 rebuilds structural representation for the DCT area image that adopts DCT domain interposing and rounding error compensation process of the present invention;

Fig. 6 is roughness definition schematic diagram;

Fig. 7 is the pixel domain decoder, adopts the DCT area image of DCT domain interposing and rounding error compensation of the present invention to rebuild structure, adopts the DCT area image of existing DCT domain interposing and rounding error compensation to rebuild structure, do not rebuild the reconstruction effect comparison diagram of structure to the mother_daughter sequence with the DCT area image of interposing and rounding error compensation.

Embodiment

In the process of DCT territory reconstructed image, the piece that does not carry out interpolation when not carrying out the piece of motion compensation and motion compensation does not need to compensate rounding error, and interpolation block needs the compensation rounding error after motion compensated interpolation is finished in the DCT territory.

Embodiment to a kind of DCT domain interposing and rounding error compensation process of the present invention is elaborated below in conjunction with accompanying drawing.

As shown in Figure 1, a kind of DCT domain interposing and rounding error compensation process comprises five steps: the interpolation block after the interpolation of (1) input DCT territory; (2) level of each pixel of the interpolation block after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector are divided into horizontal interpolation block, vertical interpolation block and two-way interpolation block with interpolation block; (3) judge the horizontal roughness and vertical roughness of interpolation block, determine the type of interpolation block; (4) determine the DCT DC coefficient offset of each type correspondence according to the type of the interpolation block of being judged; (5) respectively the interpolation block of each type is carried out the compensation of DCT DC coefficient, interpolation block after the compensation and residual error addition are finished the piece in DCT territory and are rebuild.

Now specifically introduce five steps of this method:

1) interpolation block after the interpolation of input DCT territory

2) level of each pixel of the interpolation block after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector are divided into horizontal interpolation block, vertical interpolation block and two-way interpolation block with interpolation block

The level of each pixel of the interpolation block of the image after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector, if horizontal motion vectors directed half-pixel position, the movement in vertical direction vectors directed is put in order location of pixels, and the interpolation block that is obtained by two neighbor interpolation of horizontal direction is horizontal interpolation block; If movement in vertical direction vectors directed half-pixel position, the horizontal motion vectors directed is put in order location of pixels, and the interpolation block that is obtained by two neighbor interpolation of vertical direction is vertical interpolation block; If level and movement in vertical direction vector all point to half-pixel position, the interpolation block that is obtained by four neighbor interpolation is two-way interpolation block.As Fig. 4 is half-pix interpolation schematic diagram, and E is horizontal interpolation point among the figure, and F is vertical interpolation point, and G is two-way interpolation point.

3) judge the horizontal roughness and vertical roughness of interpolation block, determine the type of interpolation block

In order to judge the flatness of interpolation block, this method is introduced horizontal roughness HR and vertical roughness VR, represent the degree of roughness on an interpolation block region horizontal direction and the vertical direction respectively, horizontal roughness HR can be defined by different modes with vertical roughness VR, and simultaneously different definition modes is selected different level and smooth threshold value HST of level and vertical level and smooth threshold value VST.

Need judge its flatness in the horizontal direction for horizontal interpolation block, if HR≤HST is judged to be horizontal interpolation smooth block, if HR〉HST, be judged to be coarse of horizontal interpolation; Need judge its flatness in vertical direction for vertical interpolation block, if VR≤VST is judged to be vertical interpolation smooth block, if VR〉VST, be judged to be coarse of vertical interpolation; Need judge its flatness on level and vertical direction simultaneously to two-way interpolation block, if HR≤HST and VR≤VST, be judged to be the two-way smooth block of two-way interpolation, if HR≤HST and VR〉VST, be judged to be two-way interpolation horizontal slide block, if HR〉HST and VR≤VST, be judged to be the vertical smooth block of two-way interpolation, if HR〉HST and VR〉VST, be judged to be coarse of two-way interpolation.

Horizontal roughness HR and the method for vertical roughness VR definition have multiplely in this method, will set forth three kinds of define methods among the embodiment.

Getting an interpolation block after the interpolation of DCT territory is:

$Z=\left(\begin{array}{ccc}{z}_{\mathrm{0,0}}& K& z_{0,n}\\ M& O& M\\ z_{m,0}& L& z_{m,n}\end{array}\right)$

Wherein Z represents a matrix, Z _{I, j}An element in the representing matrix, i represents horizontal level, is 0,1,2 from left to right successively, 3..., j represents the upright position, is 0,1,2 from top to bottom successively, 3....First kind of define method is $\mathrm{HR}=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\left|z_{0,j}\right|,$ Ac coefficient absolute value sum in expression first row, $\mathrm{VR}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\left|z_{i,0}\right|,$ Ac coefficient absolute value sum in expression first row; Second kind of define method is $\mathrm{HR}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\left|z_{i,j}\right|,$ The DCT coefficient absolute value sum of expression except that first row, first row, $\mathrm{VR}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\left|z_{i,j}\right|,$ The DCT coefficient absolute value sum of expression except that first row, first row; The third define method is $\mathrm{HR}=\underset{i=0}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\left|z_{i,j}\right|,$ The DCT coefficient absolute value sum of expression except that first row, $\mathrm{VR}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=0}{\overset{n}{\mathrm{\Σ}}}\left|z_{i,j}\right|,$ The DCT coefficient absolute value sum of expression except that first row.Wherein Yi Shang m, n are positive integer, and schematic diagram as shown in Figure 6.

Along with HST and VST value increase, a piece is judged as the probability increase of smooth block.If HST and VST value are too big, can be smooth block with coarse erroneous judgement then; If HST and VST value are too little, then can produce and fail to judge.Therefore the value of HST and VST must limit within the specific limits, could guarantee the reasonability that block type is judged.Coarse rounding error with coarse of vertical interpolation of horizontal interpolation is tending towards being evenly distributed, and utilizes this character can determine reasonably level and smooth threshold value.

Adopt the system of selection of the level and smooth threshold value of following level (vertically) in this embodiment:

(1) selects one group of typical code stream, utilize the method for adding up to determine the level and smooth threshold value of level (vertically) of single code stream.

Set one group of candidate's threshold value, since 0, with 0.1 be step increments (0,0.1,0.2,0.3 ...).To each threshold value, rebuild in the structure mark coarse of the level of threshold determination (vertically) interpolation thus at the DCT area image, add up in the pixel domain decoder then that rounding error is 0.5 proportion in these pieces.The corresponding ratio of each candidate's threshold value, ratio increases along with the increase of candidate's threshold value, levels off to 50% at last.Since 0 statistics, if the ratio that obtains is less than 48% then attempt next threshold value, if ratio is more than or equal to 48% then determine that current candidate's threshold value is a smoothly threshold value of level (vertically).

(2) select the value of maximum of the level and smooth threshold value of level (vertically) of the single code stream in this typical case's code stream as the level and smooth threshold value of the level in the method (vertically).

4) type according to the interpolation block of being judged of the present invention determines that the DCT DC coefficient compensation value method of each type correspondence is as follows:

Type according to the interpolation block of being judged of the present invention determines that the DCT DC coefficient compensation value method of each type correspondence is as follows:

After determining the sliding threshold value of this group exemplary code levelling, mark is above-mentioned all types of in DCT area image reconstruction structure, and establishing block size is M * N.The ratio of all types of rounding errors of statistics is determined offset according to the ratio of rounding error in the pixel domain decoder.For horizontal interpolation smooth block and vertical interpolation smooth block, establish q ₀Be rounding error 0 proportion, q ₁Be rounding error 0.5 proportion, the pixel compensation value is 0.5q ₁, DCT DC coefficient offset is

For the two-way interpolation smooth block of three classes, suppose p ₀, p ₁, p ₂, p ₃Be respectively 0 ,-0.25,0.5,0.25 proportion, the pixel compensation value is 0.25 (p ₁-p ₃)+0.5p ₂, DCT DC coefficient offset is $[0.25(p_1-p_3)+0.5p_2]\sqrt{\mathrm{MN}};$ Coarse of horizontal interpolation, coarse of vertical interpolation, coarse of two-way interpolation still compensate by existing method, and the pixel compensation value of coarse of horizontal interpolation, coarse of vertical interpolation is 0.25, and DCT DC coefficient offset is

The pixel compensation value that two-way interpolation is coarse is 0.125, and the pixel compensation value is $0.125\sqrt{\mathrm{MN}}.$

5) respectively the interpolation block of each type is carried out the compensation of DCT DC coefficient, interpolation block after the compensation and residual error addition are finished DCT territory piece and are rebuild.

Adopt the DCT area image of block sort interposing and rounding error compensation process to rebuild structure as shown in Figure 5.The present invention can adopt different software platforms and cycle tests to implement in the concrete process of using, for the effect of this method of simple declaration, selected the representative video sequence of akiyo, carphone, coastguard, flower, foreman, mobile, mother_daughter, news, silence, the tempete etc. of several English names below as exemplary enforcement.

Now be software platform with Xvid1.1.3, realized adopting the DCT area image of the DCT domain interposing and rounding error compensation process of the inventive method to rebuild, adopt the DCT area image of existing interposing and rounding error compensation process to rebuild and do not rebuild respectively with the DCT area image of compensation.In order to estimate DCT territory quality of reconstructed images, it is transformed into pixel domain by the IDCT budget.The video sequence code stream is generated by the Xvid1.1.3 encoder, and original series is that resolution is 352 * 288 CIF sequence, and code check is set to 4Mb/s, and picture group GOP length is 100 frames, does not have the B frame.Original series adopts the half-pixel accuracy motion compensation, and does not adopt global motion compensation.Adopt the horizontal roughness HR in the DCT domain interposing and rounding error compensation process of the present invention all to adopt above-mentioned first kind of definition mode with vertical roughness VR, setting the interpolation block size is 8 * 8, utilizes the method for statistics to determine coarse threshold value HST of level and vertical coarse threshold value VST.Set one group of candidate's threshold value, since 0, with 0.1 be step increments (0,0.1,0.2,0.3 ...).To each threshold value, rebuild in the structure mark coarse of the level of threshold determination (vertically) interpolation thus at the DCT area image, in the pixel domain decoder, add up rounding error in these pieces then and be 0.5 proportion.The corresponding ratio of each candidate's threshold value, ratio increases along with the increase of candidate's threshold value, levels off to 50% at last.Since 0 statistics, if the ratio that obtains is less than 48% then attempt next threshold value, if ratio is more than or equal to 48% then determine that current candidate's threshold value is a smoothly threshold value of level (vertically).The value of maximum of the level and smooth threshold value of level (vertically) of selecting the single code stream in this typical case's code stream is as the level and smooth threshold value of the level in the method (vertically).Obtain the coarse threshold value HST of level at last and be 5 with vertical coarse threshold value VST.Judge the horizontal roughness and vertical roughness of interpolation block, determine the type of interpolation block, rebuild all types of smooth blocks of mark in the structure, the rounding error ratio of all types of smooth blocks of statistics in the pixel domain decoder at the DCT area image.The rounding error ratio of all types of smooth blocks and the offset of determining thus are as shown in table 1:

Table 1 pixel compensation value and DCT DC coefficient offset

The pixel compensation value of coarse of horizontal interpolation, coarse of vertical interpolation is 0.25, and DCT DC coefficient offset is 2; The pixel compensation value that two-way interpolation is coarse is 0.125, and the pixel compensation value is 1.Experimental result is as shown in table 2, and PSNR is a Y-PSNR in the table.

Table 2 reconstruction video sequence Y-PSNR

From table 2 data as can be seen, for scene information relatively abundanter coastguard, foreman, sequences such as mobile, tempete, existing method can compensate certain rounding error, the block sort compensation method that the present invention proposes obtains than the better reconstructed image quality of existing method by the segmentation to block type.For sequences such as the akiyo that has a large amount of smooth regions in the scene, mother_daughter, news, silence, flower, existing method is owing to exist the overcompensation problem, reconstructed image after the compensation is compared with uncompensated reconstructed image, and quality improves limited even descends to some extent.The block sort compensation method correction that the present invention proposes the overcompensation problem of existing method, still can effectively compensate rounding error, obtain better reconstructed image quality.All in all, existing relatively compensation method, the block sort compensation method can make the reconstructed image Y-PSNR on average promote about 0.2dB.

As shown in Figure 7, because error is along with the P frame constantly spreads accumulation in the reconstructed image of DCT territory, adopt the inventive method DCT domain interposing and rounding error compensation process more effective departure to spread accumulating rate, obtain better reconstructed image quality at the GOP latter end.

Claims (4)

1. DCT domain interposing and rounding error compensation process, this method comprises following five steps:

(1) interpolation block after the interpolation of input DCT territory;

(2) level of each pixel of the interpolation block after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector are divided into horizontal interpolation block, vertical interpolation block and two-way interpolation block with interpolation block;

(3) judge the horizontal roughness and vertical roughness of interpolation block, determine the type of interpolation block;

(4) ratio of all types of rounding errors of statistics in the pixel domain decoder is determined offset according to the ratio of rounding error;

(5) respectively the interpolation block of each type is carried out the compensation of DCT DC coefficient, interpolation block after the compensation and residual error addition are finished the piece in DCT territory and are rebuild.

2. a kind of DCT domain interposing and rounding error compensation process according to claim 1, it is characterized in that: the level of each pixel of the interpolation block of the image after the interpolation of described judgement DCT territory and the sensing of movement in vertical direction vector, the method that interpolation block is divided into horizontal interpolation block, vertical interpolation block and two-way interpolation block is as follows:

The level of each pixel of the interpolation block after the interpolation of judgement DCT territory and the sensing of movement in vertical direction vector, if horizontal motion vectors directed half-pixel position, the movement in vertical direction vectors directed is put in order location of pixels, and the interpolation block that is obtained by two neighbor interpolation of horizontal direction is horizontal interpolation block; If movement in vertical direction vectors directed half-pixel position, the horizontal motion vectors directed is put in order location of pixels, and the interpolation block that is obtained by two neighbor interpolation of vertical direction is vertical interpolation block; If horizontal direction and movement in vertical direction vector all point to half-pixel position, the interpolation block that is obtained by four neighbor interpolation is two-way interpolation block.

3. a kind of DCT domain interposing and rounding error compensation process according to claim 1 is characterized in that: the horizontal roughness of described judgement interpolation block and vertical roughness, determine that the method for type of interpolation block is as follows:

Define horizontal roughness HR and vertical roughness VR, represent the degree of roughness on an interpolation block region horizontal direction and the vertical direction respectively, select level and smooth threshold value HST of corresponding level and vertical level and smooth threshold value VST simultaneously; If HR≤HST is judged to be horizontal interpolation smooth block, if HR〉HST, be judged to be coarse of horizontal interpolation; If VR≤VST is judged to be vertical interpolation smooth block, if VR〉VST, be judged to be coarse of vertical interpolation; If HR≤HST and VR≤VST are judged to be the two-way smooth block of two-way interpolation, if HR≤HST and VR〉VST, be judged to be two-way interpolation horizontal slide block, if HR〉HST and VR≤VST, be judged to be the vertical smooth block of two-way interpolation, if HR〉HST and VR〉VST, be judged to be coarse of two-way interpolation.

4. a kind of DCT domain interposing and rounding error compensation process according to claim 1 is characterized in that: described type according to the interpolation block of being judged determines that the DCT DC coefficient compensation value method of each type correspondence is as follows:

The type of the interpolation block after the interpolation of mark DCT territory is established the interpolation block size and is M * N; The ratio of all types of rounding errors of statistics is determined offset according to the ratio of rounding error in the pixel domain decoder; For horizontal interpolation smooth block and vertical interpolation smooth block, establish q ₀Be rounding error 0 proportion, q ₁Be rounding error 0.5 proportion, the pixel compensation value is 0.5q ₁, DCT DC coefficient offset is

For the two-way interpolation smooth block of three classes, suppose p ₀, p ₁, p ₂, p ₃Be respectively 0 ,-0.25,0.5,0.25 proportion, the pixel compensation value is 0.25 (p ₁-p ₃)+0.5p ₂, DCT DC coefficient offset is The pixel compensation value of coarse of horizontal interpolation, coarse of vertical interpolation is 0.25, and DCT DC coefficient offset is

The pixel compensation value that two-way interpolation is coarse is 0.125, and the pixel compensation value is

Images