CN103596004A - Intra-frame prediction method and device based on mathematical statistics and classification training in HEVC - Google Patents

The invention discloses an intra-frame prediction method and device based on mathematical statistics and classification training in HEVC. The method comprises the following steps that 1, coding parameters are set; 2, training sample data are obtained; 3, PU blocks of a DC mode are classified; 4, reference pixel sample sets and original pixel sample sets are obtained from original frames and serve as training samples; 5, sample training is carried out. Compared with the prior art, the intra-frame prediction method and device based on mathematical statistics and classification training in the HEVC have the advantages that the change distribution of image textures are paid more attention to, the prediction precision of the DC mode and a PLANAR mode is improved, a better subjective effect is achieved at the detailed position of an image background, and at the same time, the proportion of the DC mode and the PLANAR mode is increased.

Intra-frame prediction method based on mathematical statistics and classification based training in HEVC and device

Technical field

The present invention relates to high-performance video coding (HEVC, High Efficiency Video Coding) field, more specifically, relate to the intra-frame prediction method based on mathematical statistics and classification based training and device in high-performance video coding.

Background technology

Due to H.264/MPEG-4AVC issue in 03 year, through 9 years, so along with the improving constantly of network technology and terminal processing capacity, people to now widely used MPEG-2, MPEG-4, H.264 etc., new requirement has been proposed.Wishing to provide: 1) high definition, 2) 3D, 3) mobile wireless, to meet the application of the frontiers such as new home theater, remote monitoring, digital broadcasting, mobile flow medium, mobile imaging, medical imaging.In addition, H.264/AVC, after issue, the accumulation (development of the technology such as Novel sports compensation, conversion, interpolation and entropy coding) through in a few years, has possessed the technical foundation of releasing video encoding standard of new generation.

Held the JCT-VT first session at Dresden, Germany in April, 2010, determines video encoding standard title of new generation: HEVC(High Efficiency Video Coding), and set up test model.HEVC is combined the video items of establishment by ITU-T VCEG and ISO/IEC mpeg standard tissue.The HEVC standard of front page is published in January, 2013.At present, working group is planning extra work existing HEVC standard is being expanded, and comprises professional purpose, higher accuracy and the support of color format, scalable coding, 3D/ solid/multi-vision-point encoding etc.HEVC has been designed to solve all existing application demand H.264/MPEG-4AVC in essence, and ensuing focus concentrates on two key issues especially: improve the resolution of video and increase the parallel processing framework using.

The design of HEVC is exactly in order to process all application demands of H.264/MPEG-4AVC producing.HEVC mainly pays close attention to two crucial subjects under discussion: promote video resolution and carry out parallel processing framework.The grammer of HEVC has versatility, makes it can be suitable for various application.It is standardized that HEVC only has bit stream structure and grammer, and the mapping relations to the image of the restriction of bit stream and bit stream and the generation of decoding.This mapping is to realize by the semanteme of grammer and the definition of decode procedure, makes so long as meet the decoder of standard-required, and under the given bit stream condition that meets standard constraint, decoded result is all the same.The restriction of this critical field allows the freedom of maximum degree to make to be applicable to specific application (balance compression quality, implementation cost etc.) with optimization implementation.The core objective of a new generation's video compression standard is that compression efficiency doubles on the basis of high profile H.264/AVC.Guaranteeing that, under the prerequisite of same video picture quality, the code check of video flowing reduces 50%.When improving compression efficiency, can allow coding side suitably to improve complexity.

Infra-frame prediction is an important coding tools for I frame and P/B frame, but makes to be used less in P/B frame.Institute of the present invention extracting method is directed to the improvement of I frame infra-frame prediction in HEVC.I frame is a kind of independently frame, and is the first frame of each GOP.It is intended for the starting point of new focus or re-synchronization point, and can be used for realizing the ability of the random access of F.F., rewinding etc., can not produce yet significantly fuzzy, and visual quality that can augmented video.Along with the fast development of hardware, I frame is conducive to the parallel processing of multi-core multithreading.I frame is a kind of frame very important in Video coding.

I frame is the key frame in Video coding, and infra-frame prediction is very important to the coding of I frame, and prediction has directly affected the performance of quantification, conversion and entropy coding.HEVC still adopts traditional hybrid coding structure, obtains very unitary Item gain and mainly ascribes the technology that HEVC has adopted a lot of novelties to.Wherein in frame, part is the infra-frame prediction based on spatial domain, HEVC is 33 directions within the scope of 180 ° by prediction angle spread, and adopt piece flexibly to divide, basic unit block is divided into CU (coding unit, coding unit), TU (converter unit, transform units) and PU (predicting unit, prediction unit), the magnitude range of piece expands to 64 * 64 to 4 * 4.Yet the method for reference pixel mean value is asked in the utilization of existing DC predictive mode in HEVC, can not there is to all pieces good prediction effect.In addition, also some is coarse for the method that PLANAR predictive mode utilizes gradient interpolation, has affected estimated performance.Therefore,, for the highest predictive mode of these two kinds of average probability of use (DC predictive mode and PLANAR predictive mode), still there is improved space.

Summary of the invention

Present inventor considers the above-mentioned situation of prior art and has made the present invention.Method of the present invention is absorbed in the improvement of part Forecasting Methodology in HEVC I frame frame.The present invention proposes the intra-frame prediction method based on mathematical statistics and classification based training, in order to the above-mentioned two kinds of predictive mode DC in training and improvement HEVC coding and the prediction weight coefficient of PLANAR.The present invention has improved the usage rate of DC and two kinds of patterns of PLANAR, has reduced code check in the situation that of identical PSNR, and the subjective quality of video image is had to certain improvement.

Therefore the technical problem to be solved in the present invention: in HEVC, DC and the existing Forecasting Methodology of PLANAR pattern are fairly simple, with respect to angle predictive mode, these two kinds of patterns all cause predicting the outcome and have a larger predicated error, especially in the position, the lower right corner of PU piece.

According to embodiments of the invention, provide a kind of for carrying out the infra-frame prediction device of the intra-frame prediction method as described in of claim 2 to 10, comprise HEVC coding unit module, HEVC predicting unit module and sample training module, wherein, described HEVC coding unit module is carried out quad-tree partition to the maximum coding unit of frame of video, maximum coding unit is divided into a plurality of coding unit pieces, described HEVC coding unit module is proceeded quad-tree partition to coding unit piece, each coding unit piece is divided into a predicting unit piece in current division layer, at next, divide in layer and be divided into a plurality of predicting unit pieces, described HEVC predicting unit module is passed through infra-frame prediction in predicting unit piece, quantize, conversion, pseudo-entropy encoding operation, obtain code check and rate distortion costs under the current predictive mode of current predicting unit piece, more current code check and rate distortion costs of dividing layer and next division layer, according to comparative result, judge and adopt current division layer or the predicting unit piece division of next division layer, described sample training module comprises sample acquisition unit and sample training unit, described sample acquisition unit is used for: when carrying out predictive coding under DC and PLANAR predictive mode, from video sequence data, obtain the brightness value of each pixel in each predicting unit piece, the brightness value of the reference pixel of each predicting unit piece, and provide it to sample training unit, described sample training unit is used for carrying out described step 3 to 6.

Advantage beneficial effect major embodiment of the present invention is in the following areas: by adopting methods and apparatus according to embodiments of the present invention, compared with prior art, more pay attention to the change profile of image texture, improved the precision of prediction of DC and PLANAR pattern, in image background details position, there is better subjective effect, increased the ratio of DC and PLANAR predictive mode simultaneously.

Accompanying drawing explanation

Fig. 1 is the location diagram of pixel and reference pixel in 4 * 4PU piece according to an embodiment of the invention.

Fig. 3 is the training flow chart of intra-frame prediction method according to an embodiment of the invention.

Embodiment

Below in conjunction with accompanying drawing, the enforcement of technical scheme is described in further detail.

It should be noted that, in the following description, take DC or PLANAR predictive mode is example, and the principle of embodiments of the invention is described.Yet, those skilled in the art will appreciate that the content of the relevant DC predictive mode of describing with reference to the accompanying drawings can be applied to PLANAR predictive mode similarly because PLANAR predictive mode and DC predictive mode are similarly in principle, vice versa.

In addition, it will be appreciated by those skilled in the art that, within not deviating from the scope of principle of the present invention and spirit, embodiments of the invention also can be applied to other various predictive modes except DC and PLANAR predictive mode existing and that may occur in the future.

First, summarize technical scheme according to an embodiment of the invention.

According to embodiments of the invention, the described infra-frame prediction device based on mathematical statistics and classification based training mainly comprises HEVC coding unit module, HEVC predicting unit module and sample training module.

In HEVC, one two field picture is divided into a plurality of independently LCU(largest coding unit), described HEVC coding unit module is by carrying out quad-tree partition to LCU, LCU is divided into a plurality of CU(coding unit), CU is the minimum unit of intraframe coding, on its basis the residual error after quantizing is carried out to entropy coding.For example, the size of CU can be 8 * 8,16 * 16,32 * 32,64 * 64.

Described HEVC coding unit module is carried out quad-tree partition to CU, be divided into one or more PU, for example, current C U size is 8 * 8, and the PU in CU divides and may have for situation: current a divisions layer of 8 * 8PU() or 44 * 4PU(next divide layer).Described HEVC predicting unit module obtains the code check of current PU prediction and the estimated value (rate distortion costs of distortion by steps such as infra-frame prediction, quantification, conversion, pseudo-entropy codings in predicting unit (PU), Rate Distortion cost), more current layer and the code check of next division layer and the estimated value of distortion of dividing, judges whether to divide this PU piece.By predicting unit (PU), divide, according to rate distortion costs optimization, select adaptively best predicting unit division result, and then obtain the division result of CU.

Wherein, the intra prediction mode of described HEVC predicting unit module comprises angle predictive mode, DC predictive mode and PLANAR predictive mode.

Described angle predictive mode is for selecting a kind of angle direction to predict image detail content compared with the region of horn of plenty, be included in the careful angle prediction that is divided into 33 directions in 180 degree, to being tending towards the image-region of direction distribution, there is good prediction effect, can effectively reduce prediction residual.

Described DC predictive mode is a kind of pattern of predicting by calculating the method for PU reference pixel mean value, is applicable to comparatively smooth image-region.

Described PLANAR predictive mode is a kind of pattern of predicting by calculating the method for the gradation zone interpolation evaluation of PU reference pixel, is applicable to the image-region of gradual change trend.

Wherein, described sample training module comprises sample acquisition unit and sample training unit.

Described sample acquisition unit is used for: when under DC and PLANAR predictive mode, normal use HEVC encoding software carries out predictive coding, from a plurality of training video sequences, obtain respectively different sizes (4 * 4, 8 * 8) brightness value of PU piece pixel in the original video frame of piece, the brightness value of the reference pixel of correspondence position in original video frame, the data that obtain are stored in respectively in different files under DC predictive mode and under PLANAR predictive mode, for the data that obtain under same pattern (, the brightness value of above-mentioned two groups of pixels (PU piece pixel and reference pixel)), size according to its corresponding piece is stored respectively, for sample training unit.

Described sample training unit, for the brightness value of the original video frame PU piece pixel of the acquisition of normal encoding under DC predictive mode is carried out to variance counting statistics, distributes to obtain whole sample variance.

Described sample training unit is also for carrying out in the horizontal and vertical directions Grads threshold counting statistics to the brightness value of the original video frame reference pixel of the acquisition of normal encoding under PLANAR predictive mode.

Also classify for the sample evidence primitive frame original pixels variance threshold values scope to obtaining under DC predictive mode in described sample training unit, in the sample group of difference classification, the corresponding reference pixel obtaining in the sample point use primitive frame of diverse location in PU piece is weighted to the training of coefficient, to obtain the optimum weight estimation coefficient for DC pattern of each location of pixels in PU piece.

Described sample training module distributes and carries out sample classification for the reference pixel Grads threshold that the sample evidence primitive frame obtaining under PLANAR predictive mode is obtained, in the sample group of difference classification, the corresponding reference pixel obtaining in the sample point use primitive frame of diverse location in PU piece is weighted to the training of coefficient, to obtain the optimum weight estimation coefficient for PLANAR pattern of each location of pixels in PU piece.

In HEVC, the recurrence of CU is divided with the recurrence of PU and is divided and combine closely, first carry out the division of CU aspect, the division of CU is again to obtain according to pre-measured compressed and the rate-distortion optimization of PU, while predicting in PU unit conducting frame, by traveling through angle prediction, DC prediction and the PLANAR of above-mentioned 33 directions, predict, obtain best predictive mode.

Further, described infra-frame prediction device has following characteristics: off-line obtains training sample, reduces the algorithm complex likely increasing in actual coding process, and when training coefficient for improvement of DC pattern and PLANAR pattern, predicted time remains unchanged substantially.

Further, described infra-frame prediction device has following characteristics: by mathematical statistics, sample is analyzed to classification, on the basis of classification, carry out sample training and obtain optimum prediction weight coefficient, more careful picture material details and the variation tendency of catching, to obtain best prediction effect.

In order better to carry out the infra-frame prediction of DC and PLANAR predictive mode, first video image content is carried out to mathematical statistics analysis.Under DC predictive mode, the method for embodiments of the invention is only used left reference pixel r (9)～r (12) and upper reference pixel r (1)～r (4) as shown in Figure 1.Correlation analysis refers to that the variable element that two or more are possessed to correlation analyzes, thereby weighs the related intimate degree of two Variable Factors.By calculating the correlation between each location point pixel in each reference pixel and 4 * 4, can find that correlation and position relationship have closely contacts, be that the original pixels of diverse location and the reference pixel of diverse location have the degree of correlation apart from property, piece is interior along the direction from the upper left corner to the lower right corner, therefore correlation weakens, and predicated error increases.4 * 4 the prediction piece of take is example, the correlation maximum of r (1) and r (9) in pixel s (1,1) and 8 reference pixels in piece, it is 0.91 left and right, and pixel s (4,4) is more or less the same with the correlation of 8 reference pixels, float in 0.74 left and right.As can be seen here, the method for existing DC and PLANAR predictive mode can not be predicted all pixels in polylith well.

Due to the uncertainty of pixel distribution, there is some difference in the distribution of different sample block and reference pixel, and the training of directly all samples being carried out to original PU piece and its reference pixel can not obtain the optimum prediction coefficient changing according to picture material.Therefore need further video image pixel to be analyzed.

Variance and standard deviation are most important, the most frequently used indexs of the discrete trend of measuring and calculating.Variance is the average of each variable and its deviation from mean square, is the important method of weighing sample fluctuation size.Optimal mode is that the piece sample distribution of DC it seems it is more balanced substantially, but for the finer picture material distribution characteristics of catching, the standard that the method is used variance to distribute as further observed data.Pearson's coefficient correlation (Pearson correlation coefficient) also claims Pearson's product moment correlation coefficient (Pearson product-moment correlation coefficient), is a kind of linearly dependent coefficient.Pearson's coefficient correlation is for reflecting the statistic of two linear variable displacement degrees of correlation.The method of embodiments of the invention is weighed the variance of reference pixel and the degree of correlation of original pixels variance by Pearson's coefficient correlation.By the variance of computing reference pixel and Pearson's coefficient correlation of original pixels variance, can weigh this two variable fluctuation correlations by variance.For example, Pearson's phase relation numerical value of the variance yields of PU piece in a training sequence sample and the variance yields of its reference pixel is 0.84907, due to Pearson's phase relation numerical value of two variablees (variance yields of PU piece and the variance yields of its reference pixel), more to approach its correlation of 1 explanation larger, so, in the case, can estimate with the variance yields of reference pixel the fluctuation situation of the pixel value of the PU piece that it is corresponding.

As shown in Figure 2, the variance histogram of original pixels and reference pixel shows, distributes close to Gaussian Profile, and peak point left side is comparatively precipitous, and there is hangover characteristic on right side.As shown in the drawing, peak point drops in 1～2 scope, and this meets the applicable smooth feature of piece content of DC pattern.And for the sample on peak point right side, sample variance is larger, illustrates that sample place image-region content difference is large, needs to obtain more excellent prediction weight coefficient by training.According to the histogrammic hangover feature of variance, each PU piece of data sample is classified as follows: D1(flat site, variance yields (variance yields of PU piece reference pixel) is 0～1); The micro-fuctuation within a narrow range of D2(region, variance yields is 1～2.5); D3(fuctuation within a narrow range region, variance yields is 2.5～4); D4(larger difference region, variance yields is 4～10); The uneven region of D5(, variance yields is 10～40).For example, very smooth division D1 accounts for greatly about 25%-35% of all samples, remains different divisions and still occupies very large ratio, and therefore in not all PU piece that has adopted DC predictive mode, pixel is all suitable for the prediction mode of calculating mean value.

For PLANAR pattern, the object of primal algorithm is the prediction for gradient region.Calculate the difference between upper left point and top-right point (or the lower left corner and upper left point), during two tap interpolations, carry out the compensation of a gradual change degree.In order more effectively to train optimal mode, be the sample of PLANAR, embodiments of the invention carry out comparatively careful gradient calculation according to Gradient Features.Particularly, for example, by (calculating neighbor, reference pixel r (the 9)～r (12) on the left side, reference pixel r (the 1)～r (4) of top, the difference between as shown in Figure 1), obtains the gradient set of the horizontal and vertical direction of reference pixel.According to the calculating of the gradient in above horizontal direction and vertical direction (Grad in horizontal direction and vertical direction positive and negative), each PU piece of the data sample obtaining is divided into 5 types: P1(reference pixel vertical direction increases progressively horizontal direction and increases progressively); The P2(reference pixel vertical direction horizontal direction of successively decreasing increases progressively); P3(reference pixel vertical direction increases progressively horizontal direction and successively decreases); The P4(reference pixel vertical direction horizontal direction of successively decreasing increases progressively); P5(other).

The flow process of the method for the embodiment of the present invention is described below.Fig. 3 is the training flow chart of intra-frame prediction method according to an embodiment of the invention.It should be noted that, Fig. 3 only, for being used for illustrating the schematic diagram of the principle of embodiments of the invention, does not form the restriction to embodiments of the invention, and the operation in each frame in Fig. 3 must be not corresponding with each step of describing in this manual yet.

The training flow process of intra-frame prediction method mainly comprises the following steps according to an embodiment of the invention:

Other four groups of sample datas to 4 * 4PU piece under DC pattern repeat said process, finally obtain five groups of weight coefficient groups of classifying according to original PU piece variance threshold values of 4 * 4PU piece under DC pattern, and every group of weight coefficient group comprises 16 weight estimation coefficient matrixes.

In like manner, for 8 * 8PU piece under DC pattern, also carry out similar training process with 4 * 4PU piece under PLANAR pattern with 8 * 8PU piece.

In HEVC, the design of angle prediction has reached the interior intensive level of coverage of 180 degree of the statistical property that meets image pixel distribution of content substantially, and angle prediction has reached extraordinary effect.And statistical result showed DC and PLANAR pattern account for the percentage of all use patterns up to 25%～30%, be two kinds of predictive modes that frequency of utilization is very high, improve this predictive mode has very great meaning concerning whole predictive coding.

Advantage beneficial effect major embodiment of the present invention is in the following areas: by adopting methods and apparatus according to embodiments of the present invention, compared with prior art, more pay attention to the change profile of image texture, improved the precision of prediction of DC and PLANAR pattern, in image background details area, there is better subjective effect, meanwhile, the DC after improvement compares with the utilization rate of PLANAR predictive mode with original DC with PLANAR predictive mode, increases to some extent.Experimental results is as shown in following table 1 to 3, and wherein Δ R represents that code check changes percentage, and Δ PSNR represents that PSNR value changes.

Table 1: the increase of DC pattern and PLANAR pattern using rate while encoding after improving

Table 2: the test result of first group of QP

Table 3: the test result of second group of QP

In sum, those skilled in the art will appreciate that the above embodiment of the present invention can be made various modifications, modification and be replaced, it all falls into the protection scope of the present invention limiting as claims.