Patent Application
20140044181
The present invention relates to a method and a system for video signal encoding and decoding with motion estimation. It relates to the technical field of video compression.
One of the recent developments in video compression is High Efficiency Video Coding (HEVC), which utilizes intra-pictures prediction with motion compensation and advanced methods of motion vectors prediction. HEVC has been described in details in: “High Efficiency Video Coding (HEVC) text specification draft 10 (for FDIS & Consent)” by B. Bross et al (JCTVC-L1003, Geneva, January 2013).
In HEVC, a picture is coded in blocks. The basic block, for which motion compensation prediction is performed, is called a Prediction Unit (PU). When coding a PU with motion compensation, the PU is assigned information related to its motion and a residual signal.
The residual signal describes a difference between values of pixels of the coded block and the values of pixels of the block resulting from motion compensation. In HEVC, a block may be coded in a SKIP mode, wherein no residual signal is sent for the block or not in a SKIP mode, wherein there is some residual signal sent for the block.
Motion information comprises motion vectors (MV) and reference image indexes (RefIdx). The process of searching the motion vector is called motion estimation (ME). In that process, a reference block is searched for a particular block PUx and indicated by a motion vector MVx such as to minimize the cost function defined by HEVC as:
J
pred,SATD=SATD+λprep*Bpred (1)
wherein SATD (a sum of absolute transformed differences) represents the measure of non-matching of the points of the reference block to the values of points of the coded block, λpred represents the Lagrange's coefficient and Bpred represents the number of bits necessary to code motion information.
The motion vector MVx itself is prediction-coded, using motion vectors assigned to PU blocks which have been coded previously and are neighboring to the coded block PUx. The motion vectors of the neighboring blocks PU form a list of motion vector predictions LMVXpred for the block PUx. Before the motion vectors of the neighboring blocks PUY are inserted to LMVXpred, they are scaled according to the temporal distance between current frame and reference frame pointed by neighboring PUY reference frame index and current PUX reference index. After scaling all predictions in LMVXpred represent the same temporal distance. In HEVC LMVXpred has a fixed length and if not enough predictions can be found at this point, then generated motion vectors are inserted to fill it. For example, a zero motion vector can be added as an prediction. In the bitstream, the motion vector MVx is represented an index LIdx of motion vector prediction on the list of predictions LMVXpred and a possible correction dMVx for the vector resulting from the prediction. When the prediction block is coded in a MERGE mode, the correction dMVx for the vector resulting from the prediction equals zero and is not transmitted. The motion vector predictions may have higher or lower quality. The quality of motion vector predictions is understood as generating the lowest coding cost J for the block PU for the particular prediction.
When a block is coded with motion compensation, when keeping a constant quality of the coded block, the number of bits B decreases as the quality of motion vectors predictions is increased.
US2011176611 discloses a method for decoder-side motion vector derivation (DMVD) that includes: checking a block size of a current block to be encoded and accordingly generating a checking result; and utilizing a DMVD module to refer to the checking result to control conveyance of first DMVD control information which is utilized for indicating whether a DMVD coding operation is employed to encode the current block. When the checking result indicates a predetermined criterion is satisfied, the first DMVD control information is sent in a bitstream; otherwise, the first DMVD control information is not sent.
US2012134416 discloses a method and apparatus for deriving a temporal motion vector predictor (MVP). The predictor is derived for a current block of a current picture in Inter, or Merge, or Skip mode based on co-located reference blocks of a co-located block. The co-located reference blocks comprise the above-left reference block of the bottom-right neighboring block of the co-located block. The reference motion vectors associated with the co-located reference blocks are received and used to derive the temporal MVP.
US2012128060 discloses an apparatus and method for deriving a motion vector predictor or a motion vector predictor candidate or a motion vector or a motion vector candidate for a current block.
US2011002390 discloses a method and apparatus for deriving a motion vector at a video decoder. A block-based motion vector may be produced at the video decoder by utilizing motion estimation among available pixels relative to blocks in one or more reference frames. A motion vector MVPRED which is found for a neighboring block Y is used to determine the motion vector for block X. The procedure is carried out for all blocks, which imposes significant load to the decoder apparatus.
The aim of the invention is to improve the methods for deriving a motion vector in order to decrease the bandwidth necessary to transmit the stream coded with motion estimation.
The object of the invention is a computer-implemented method for video signal encoding with motion estimation and frames division into prediction units. The method comprises the steps of: determining the current prediction unit (PUX) to be encoded, creating a list (LMVXpred) comprising motion vector predictions (MVPUY) in neighboring units (PUY), selecting from the list (LMVXpred) the motion vector prediction which is the best according to a predetermined cost function for encoding the current prediction unit (PUX), using the selected prediction number to encode the current prediction unit (PUX). The method further comprises the steps of: for each neighboring unit (PUY), checking whether the neighboring unit (PUY) has been encoded in the MERGE and not SKIP mode and if so, determining a reconstruction motion vector (MVPUYrec) for that unit (PUY) as a motion vector which minimizes difference between reconstructed prediction block and a block in a reference frame pointed by this motion vector, and assigning the reconstruction motion vector (MVPUYrec) as a motion vector (MVPUY) to that unit (PUY) to be used in creating the list (LMVXpred).
The object of the invention is also a computer readable medium comprising program code means for performing all the steps of the method for video signal encoding according to the invention when said program is run on a computer.
Another object of the invention is a computer-implemented method for video signal decoding with motion estimation and working with the video signal comprising frames divided into prediction units. The method comprises the steps of: determining the current prediction unit (PUX) to be decoded, determining from the encoded stream information a neighboring unit (PUY) whose motion vector shall be used to decode the current prediction unit (PUX), using the motion vector (MVPUY) of the neighboring unit (PUY) to decode motion information in the current prediction unit (PUX) The method further comprises the steps of: checking whether the neighboring unit (PUY) has been encoded in a MERGE and not SKIP mode and if so, determining the reconstruction motion vector (MVPUYrec) for that unit (PUY) as a motion vector which minimizes difference between reconstructed prediction block and a block in a reference frame pointed by this motion vector, and assigning the reconstruction motion vector (MVPUYrec) as a motion vector (MVPUY) for that unit (PUY) to be used in decoding the current prediction unit (PUX) motion information.
The object of the invention is also a computer readable medium comprising program code means for performing all the steps of the method for video signal decoding according to the invention when said program is run on a computer.
A further object of the invention is a computer-implemented method for transmitting a video signal, comprising the steps of encoding the video signal according to the method of the invention and decoding the video signal according to the method of the invention.
The object of the invention is also video signal encoder utilizing motion estimation and frames division into prediction units. The encoder comprises: a motion estimation unit configured to: determine the current prediction unit (PUX) to be encoded, create a list (LMVXpred) comprising motion vector predictions (MVPUY) of neighboring units (PUY), select from the list (LMVXpred) the motion vector prediction which is the best according to a predetermined cost function for encoding the current prediction unit (PUX), use the selected prediction number to encode the current prediction unit (PUX). The encoder further comprises a prediction correction unit. The prediction correction unit is configured to: for each neighboring unit (PUY), check whether the neighboring unit (PUY) has been encoded in the MERGE mode and not in a SKIP mode and if so, determine the reconstruction motion vector (MVPUYrec) for that unit (PUY) as a motion vector which minimizes difference between reconstructed prediction block and a block in a reference frame pointed by this motion vector. The reconstruction motion vector (MVPUYrec) is used as a motion vector (MVPUY) to that unit (PUY) to be used in creating the list (LMVXpred).
Another object of the invention is a video signal decoder utilizing motion estimation and working with the video signal comprising frames divided into prediction units. The decoder comprises: a motion estimation unit configured to: determine the current prediction unit (PUX) to be decoded, determine from the encoded stream information a neighboring unit (PUY) whose motion vector shall be used to decode the current prediction unit (PUX), use the motion vector (MVPUY) of the neighboring unit (PUY) to decode motion information in the current prediction unit (PUX). The decoder further comprises: a prediction correction unit configured to: check whether the neighboring unit (PUY) has been encoded in a MERGE and not SKIP mode and if so, determine the reconstruction motion vector (MVPUYrec) for that unit (PUY) as a motion vector which minimizes difference between reconstructed prediction block and a block in a reference frame pointed by this motion vector. The reconstruction motion vector (MVPUYrec) is used to get a motion vector (MVPUY) for that unit to be used in decoding the current prediction unit (PUX).
The invention further relates to a video signal transmission system, comprising the video signal encoder according to the invention and a video signal decoder according to the invention.
The invention is shown by means of exemplary embodiments on a drawing, in which:
The description will now introduce a HEVC encoder and decoder as a background to the invention, and after that a detailed description of new elements added to the HEVC will be introduced. It is to be understood that HEVC is described here only as an exemplary embodiment of the invention. The invention can be carried out with other video compression techniques which utilize motion estimation.
Each frame of the original video sequence 11 is first divided into a grid of coding units (CU) during stage 101 and slices and tiles are defined. In general, s two methods define slice boundaries by either defining a given number of CUs per slices (entropy or coding slices) or a given number of bytes per slice. Tiles boundaries are defined by spatial division of an frame into rectangular sections. The subdivision of an LCU into CUs and the partitioning of a CU into TUs and PUs are determined based on a rate distortion criterion. Each PU of the CU being processed is predicted spatially by an “Intra” predictor 117, or temporally by an “Inter” predictor 118. Each predictor is a block of pixels issued from the same image or another image, from which a difference block (or “residual”) is derived.
The encoded images can be of two types: temporal predicted images which can be either predicted from one or more reference images in one direction are called P-frames or predicted from at least two reference frames in two directions (forward and backward) are called B-frames; and non-temporal predicted frames called Intra frames or I-frames. In I-frames, only Intra prediction is considered for coding CUs/PUs. In P-frames and B-frames, intra and Inter prediction are considered for coding CUs/PUs.
In the “Intra” prediction processing module 117, the current block is predicted by means of an “Intra” predictor which corresponds to a block of pixels constructed from the information of the current image already encoded. The module 102 determines an intra prediction mode that is used to predict pixels of a current PU to encode from neighboring PUs pixels. In HEVC, Planar, DC and up to 33 angular modes are available. A residual PU is obtained by computing the difference between the intra predicted PU and current PU of pixels. An intra-predicted PU therefore comprises a intra prediction mode with a residual. The coding of the intra prediction mode is partly inferred from the prediction mode of neighboring prediction units. This inferring process 103 of prediction mode enables the coding rate of the intra prediction direction mode to be reduced. The Intra prediction processing module 117 also uses the spatial dependencies of the frame for predicting the pixels and for inferring the intra prediction direction of the prediction unit.
With regard to the second processing module 118 that is “Inter” coding, two prediction types are possible. Mono-prediction (P-type) entails predicting the PU by referring to one reference area from one reference picture. Bi-prediction (B-type) entails predicting the PU by referring to two reference areas from one or two reference pictures. In HEVC, B-type frames have been generalized and replace P-type frames which now predict the PU by referring to two reference areas in one reference picture. An estimation of motion 104 between the current PU and reference images 115 is made in order to identify, in one or several of these reference images, one (for P-type) or several (for B-type) areas of pixels to use them as predictors of this current PU. In the case where several areas predictors are used (B-type), they are merged to generate one single prediction. The reference images are images in the video sequence that have already been coded and then reconstructed (by decoding).
The reference area is identified in the reference frame by a motion vector that is equal to the displacement between the PU in current frame and the reference area. The next stage 105 of the inter prediction process involves computing the difference between the prediction area and the current PU. This difference is the residual of the inter predicted PU. At the end of the inter prediction process the current PU is composed of one motion vector and a residual.
By virtue of spatial dependencies of movement between neighboring PUs, HEVC provides a method to predict the motion vectors of each PU. Several motion vector predictions are employed: typically, the motion vector of the PU localized on the top of, the left of or the top left corner of the current PU are a first set of spatial predictions. One temporal motion vector candidate is also used that is the one of the collocated PU (i.e. the PU at the same coordinate) in a reference frame. The coder then removes predictions that are equal within the set of candidates. It selects one of the predictions based on a criterion that minimizes the difference between the MV prediction and that of the current PU. In HEVC, this process is referred to as Advanced Motion Vector Prediction (AMVP). Finally, the current PU's motion vector is coded 106 with an index that identifies the prediction within the set of candidates and a MV difference MVD of PU's MV with the selected MV prediction candidate. The inter prediction processing module also relies on spatial dependencies between motion information of prediction units to increase the compression ratio of inter predicted coding units.
These two types of codings thus supply several texture residuals (the difference between the current PU and the predictor), which are compared in a module 116 for selecting the best coding mode.
The residual obtained at the end of an inter or intra prediction process is then transformed in module 107. The transform applies to a Transform Unit TU that is included in a CU. A TU can be further split into smaller TUs using a so-called Residual QuadTree (RQT) decomposition 107. In HEVC, generally 2 or 1 levels of decompositions are used and authorized transform sizes are from 32*32, 16*16, 8*8 and 4*4. The transform basis is derived from a discrete cosine transform DCT.
The residual transformed coefficients are then quantized 108. The coefficients of the quantized transformed residual are then coded by means of an entropy coding process 109 and then inserted into the compressed bit stream 21. Coding syntax elements are also coded by entropy encoding 109. This processing module uses spatial dependencies between syntax elements to increase the coding efficiency.
In order to calculate the “Intra” predictors or to make an estimation of the motion for the “Inter” predictors, the encoder performs a decoding of the PUs already encoded by means of a so-called “decoding” loop 111, 112, 113, 114, 115. This decoding loop makes it possible to reconstruct the PUs and images from the quantized transformed residuals.
Thus the quantized transformed residual is dequantized 111 by applying the inverse quantization to that provided at quantization step 108 and reconstructed 112 by applying the inverse transform to that of the step 107.
If the residual comes from an “Intra” coding process 117, the used “Intra” predictor is added to this residual in order to recover a reconstructed PU corresponding to the original PU modified by the losses resulting from a transformation with loss, for example in this case the quantization operations.
If on the other hand the residual comes from an “Inter” coding 118, the areas pointed to by the current motion vectors (these areas belong to the reference images 115 referred by the current image indices) are merged then added to this decoded residual. In this way the original PU is modified by the losses resulting from the quantization operations.
A final loop filter processing module 119 is applied to the reconstructed signal in order to reduce the effects created by heavy quantization of the residuals obtained and to improve the signal quality. The loop filter processing module comprises two steps, a “deblocking” filter and a linear filtering. The deblocking filter 113 smoothes the borders between the PUs and TUs in order to visually attenuate the high frequencies created by the coding. Such a filter being known to a skilled person, it will not be described in any further detail here. The sample adaptive offset (SAO) 114 filter further improves the signal using filter coefficients adaptively determined. The coefficients of the filters are coded and transmitted in the bitstream. The filter 119 is thus applied to an image when all the PUs of pixels of the image have been decoded. The filtering process is performed on a frame by frame basis and uses several pixels around the pixel to be filtered. This processing module 119 also uses spatial dependencies between pixels of the frame.
The filtered images, also known as reconstructed images, are then stored as reference images 115 in order to allow the subsequent “Inter” predictions taking place during the compression of the subsequent images of the current video sequence.
In the context of HEVC, it is possible to use several reference images 115 for the estimation and motion compensation of the current image. In other words, the motion estimation is carried out on N images. Thus the best “Inter” predictors of the current PU, for the motion compensation, are selected in some of the multiple reference images. Consequently two adjoining PUs may have two predictor PUs that come from two distinct reference images. This is particularly the reason why, in the compressed bit stream, the index of the reference image (in addition to the motion vector) used for the predictor area is indicated.
The decoder receives as an input a bit stream 21 corresponding to a video sequence 11 compressed by an encoder of the HEVC type, such as the one in
The partition of the LCU is parsed and CU, PU and TU subdivision are identified. The decoder successively processes each CU by intra 207 or inter 206 processing modules, inverse quantization and inverse transform modules and finally loop filter processing module 219.
The “Inter” or “Intra” coding mode for the current block is parsed from the bit stream 21 with help of the parsing process module. Depending on the coding mode, either intra prediction processing module 207 or inter prediction processing module 206 is employed. If the coding mode of the current block is “Intra” type, the intra prediction mode is extracted from the bit stream and decoded with help of neighbors' prediction mode during stage 204 of intra prediction processing module 207. The intra predicted block is then computed 203 with the decoded intra prediction mode and the already decoded pixels at the boundaries of current PU. The residual associated with the current block is recovered from the bit stream and then entropy decoded 201.
If the coding mode of the current PU indicates that this PU is of “Inter” type, the motion information is extracted from the bit stream and decoded 204. AMVP process is performed during step 204. Motion information of neighbors PU already decoded are also used to compute the motion vector of current PU. This motion vector is used in the reverse motion compensation module 205 in order to determine the “Inter” predictor PU contained in the reference images 215 of the decoder. In a similar manner to the encoder, these reference images 215 are composed of images that precede in decoding order the image currently being decoded and that are reconstructed from the bit stream (and therefore decoded previously).
The next decoding step consists in decoding the residual block that has been transmitted in the bitstream. The parsing module 201 extracts the residual coefficients from the bitstream and performs successively the inverse quantization 211 and inverse transform 212 to obtain the residual PU. This residual PU is added to the predicted PU obtained at output of intra or inter processing module.
At the end of the decoding of all the PUs of the current image, the loop filter processing module 219 is used to eliminate the block effects and improve the signal quality in order to obtain the reference images 215. As done at the encoder, this processing module employs the deblocking filter 213 and then the SAO filter 214.
The images thus decoded constitute the output video signal 31 of the decoder, which can then be displayed and used.
In the present invention, the standard HEVC encoder and HEVC decoder have been improved by adding a prediction correction units 151, 251, which operate according to procedures shown in
The encoder motion estimation unit 104 and the encoder prediction correction unit 151 operate jointly as shown in
it must have a motion vector correction dMVY equal to zero (MERGE)
some residual signal must be sent for the block (not SKIP)
Blocks that satisfy the above criteria have been coded such because the cost function indicated that it is more efficient to send a residual signal for the block than to correct the motion vector. In order to improve the image quality, instead of using the motion vector of the neighboring unit PUY as an motion vector prediction for the current unit PUX, the reconstruction motion vector MVPUYrec is determined in step 404. The reconstruction motion vector MVPUYrec is a vector that minimizes difference between reconstructed prediction block PUY and a block in a reference frame pointed by this motion vector. It can be searched by motion estimation or modification of motion estimation, wherein the cost function Jpred, SATD is modified and does not involve the number of bits necessary to code the part or full information of motion. This can be performed by any adequate method known to persons skilled in video coding. The reconstruction motion vector MVPUYrec is assigned as the motion vector to the neighboring block PUY in step 405. Next MVPUY is scaled in step 406 and if the resulting MVPUY is not already on list LMVXpred of predictions then it is added as a candidate to the list LMVXpred of predictions in step 407. It is then checked in step 408 whether all neighboring prediction units PUY have been analyzed and if not, the procedure moves to the next neighboring prediction unit PUY in step 409. After all neighboring units are analyzed, additional generated MV can be included in list LMVXpred to fill all available list entries in step 410. Next in step 411 the best prediction for the currently analyzed unit PUX is selected from the list LMVXpred of candidate predictions generated in step 407 and the prediction number corresponding to that motion vector prediction is assigned for coding the current prediction unit PUX in step 412.
Therefore, the list of candidate predictions for the currently analyzed unit PUX includes standard motion vectors (i.e. determined according to HEVC rules) for blocks which do not satisfy the criteria (a MERGE mode and not a SKIP mode) and reconstruction motion vectors for blocks which are coded in the MERGE and not SKIP mode.
Since the actual motion vector values are not transmitted for the PUX unit, but a prediction number pointing to a particular neighboring unit PUY from which the motion vector shall be used together with optional adjustment, it is necessary that the same functionality of determining a reconstruction prediction is implemented in the decoder prediction unit 251.
The decoder motion information unit 204 and the decoder prediction correction unit 251 operate jointly as shown in
The procedures shown in
| Number | Date | Country | Kind |
|---|---|---|---|
| 400344 | Aug 2012 | PL | national |
