Moving picture coding apparatus
Tsurube - August, 1992 - 5136378
Motion compensated prediction interframe coding system
Tanka et al. - September, 1992 - 5144426
Apparatus for encoding/decoding video signals to improve quality of a specific region
Ueno et al. - September, 1992 - 5150432
Video signal coding method
Yonemitsu et al. - October, 1992 - 5155593
Encoding/decoding moving images with forward and backward keyframes for forward and reverse display
Normille et al. - November, 1993 - 5267334
| 5293229 | Apparatus and method for processing groups of fields in a video data compression system | March, 1994 | Iu | 375/240.15 |
| 5315326 | Efficient coding/decoding apparatuses for processing digital image signal | May, 1994 | Sugiyama | 375/240.12 |
| 5361105 | Noise reduction system using multi-frame motion estimation, outlier rejection and trajectory correction | November, 1994 | Iu | 348/699 |
| 5386234 | Interframe motion predicting method and picture signal coding/decoding apparatus | January, 1995 | Veltman et al. | 375/240.13 |
| 5412428 | Encoding method and decoding method of color signal component of picture signal having plurality resolutions | May, 1995 | Tahara | 348/396.1 |
| 5436665 | Motion picture coding apparatus | July, 1995 | Ueno et al. | 375/240.14 |
| 5467136 | Video decoder for determining a motion vector from a scaled vector and a difference vector | November, 1995 | Odaka et al. | 375/240.15 |
| 5481310 | Image encoding apparatus | January, 1996 | Hibi | 348/413.1 |
| 5905534 | Picture decoding and encoding method and apparatus for controlling processing speeds | May, 1999 | Sumihiro et al. | 375/240.15 |
| 6104753 | Device and method for decoding HDTV video | August, 2000 | Kim et al. | 375/240.16 |
| 6184935 | Upsampling filter and half-pixel generator for an HDTV downconversion system | February, 2001 | Iaquinto et al. | 348/441 |
| 6219383 | Method and apparatus for selectively detecting motion vectors of a wavelet transformed video signal | April, 2001 | Park | 375/240.16 |
| EP0651574 | May, 1995 | METHOD AND APPARATUS FOR CODING/DECODING MOTION VECTOR, AND METHOD AND APPARATUS FOR CODING/DECODING IMAGE SIGNAL. |
K. Rijske, “ITU Standardization of Very Low Bitrate Video Coding Algorithms”, vol. 7, No. 4, pp. 553-565, XP004047099, Nov. 1, 1995.
Kozu et al., “A New Technique for Block-Based Motion Compensation”, pp. V/217-20, vol. 5, XP002151093, Apr. 1, 1994.
“Transmission of Non-Telephone Signals Information Technology-Generic Coding of Moving Pictures and Associated Audio Information: Video”, ITU-T Telecommunication Standarization Sector of ITU, XX,XX, Jul. 1, 1995, pp. A-B, I-VIII, 1, XP000198491.
W. Lynch, “Bidirectional Motion Estimation Based On P Frame Motion Vectors and Area Overlap”, vol. CONF. 17, Mar. 23, 1992, pp. 445-448, XP000378964.
A. Nagata, “Moving Picture Coding System for Digital Storage Media Using Hybrid Coding”, vol. 2, No. 2, Aug. 1, 1990, pp. 109-116, XP000243471.
European Search Report dated Nov. 8, 2000, application No. EP 96120920.
Secretariat: Japan (JISC), “Coded Representation of Audio, Picture, Multimedia and Hypermedia Information,” ISO/IEC JTC 1/SC 29 N 313, dated May 20, 1993.
“Recommendation H.261—Video Codec for Audiovisual Services at px 64 kbit/s”, International Telegraph and Telephone Consulative Committee, Study Group XV—Report R 37, Aug. 1990.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13. A method for decoding a compressed video image sequence of a group of pictures including an I-frame followed by a plurality of P-frames and B-frames, comprising the steps of: decoding a block in the I-frame independently from any other frames in the group of pictures; predictively decoding a block in a P-frame based on the previous I-frame or a previous P-frame; bi-directionally predictively decoding a block in a B-frame based on the previous I-frame or a previous P-frame and a block in a P-frame positioned after the B-frame; deriving a scaled forward motion vector and a scaled backward motion vector for the block in the B-frame by scaling a motion vector of the block in the P-frame positioned after the B-frame; obtaining a final forward motion vector for the block in the B-frame by adding a delta motion vector to the scaled forward motion vector; and obtaining a final backward motion vector for the block in the B-frame by adding the delta motion vector to the scaled backward motion vector.
14. A method of decoding a sequence of video image frames according to claim 13, wherein the deriving step includes: scaling of the forward and backward motion vectors is based on a temporal reference of the B-frame and the P-frame.
15. A method for decoding a sequence of video image frames according to claim 13, further comprising the step of forming a decoded output, wherein the decoded output is responsive to a bitstream comprising: temporal reference information for the B-frame and the P-frame; motion vector information for the block in the P-frame; quantized residual error information for the block in the P-frame; the delta motion vector information for the block in the B-frame; and quantized residual error information for the block in the B-frame.
16. A method for decoding a sequence of video image frames according to claim 15, wherein the bitstream contains additional information indicating a presence of at least one of the delta motion vector information for the block in the B-frame; and the quantized residual error information for the block in the B-frame.
This application is a divisional reissue application of U.S. Pat. No. 5 , 825 , 421 , issued Oct. 20 , 1998 . This application also has a related reissue application Ser. No. 09 / 691 , 857 , filed on Oct. 18 , 2000 .
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention can be used in low bit rate video coding for tele-communicative applications. It improves the temporal frame rate of the decoder output as well as the overall picture quality.
2. Related art of the Invention
In a typical hybrid transform coding algorithm such as the ITU-T Recommendation H.261 [1] and MPEG [2] motion compensation is used to reduce the amount of temporal redundancy in the sequence. In the H.261 coding scheme, the frames are coded using only forward prediction, hereafter referred to as P-frames. In the MPEG doing scheme, some frames are coded using bi-direction prediction, hereafter referred to as B-frames. B-frames improve the efficiency of the coding scheme. Now the [1] is ITU-T Recommendation H.261 (Formerly CCITT Recommendation H.261) Codes for audiovisual services at p×64 kbit/s Geneva. 1990 , and the [2] is ISO/TEC 11172-2 1993 . Information technology—Coding of moving pictures and associated audio for digital storage media at up to about 1.5 Mbit/s—Part 2: Video.
However, it introduces delay in the encoding and decoding, making it unsuitable for applications in the communicative services where delay is an important parameter. FIG. 1a and 1b illustrates the frame prediction of H.261 and MPEG as described above. A new method of coding involving the coding of the P and B frames as a single unit, hereafter referred to as the PB-frame, was introduced. In this scheme the blocks in the PB-frames are coded and transmitted together thus reducing the total delay. In fact the total delay should not be more than a scheme using forward prediction only but at half the frame rate.
FIG. 2a shows the PB-frame prediction. A PB-frame consists of two pictures being coded as one unit. The name PB comes from the name of picture types in MPEG where there are P-frames and B-frames. Thus a PB-frame consists of one P-frame which is predicted from the last decoded P-frame and one B-frame which is predicted both from the last decoded P-frame and the P-frame currently being decoded. This last picture is called B-frame because parts of it may be bi-directionally predicted from the past and further P-frame.
FIG. 2b shows the forward and bi-directional prediction for a block in the B-frame, hereafter refereed to as a B-block. Only the region that overlaps with the corresponding block in the current P-frame, hereafter referred to as the P-block, is bi-directionally predicted. The rest of the B-block is forward predicted from the previous frame. Thus only the previous frame is required in the frame store. The information from the P-frame is obtained from the P-block currently being decoded.
In the PB-block only the motion vectors for the P-block is transmitted to the decoder. The forward and backward motion vectors for the B-block is derived from the P motion vectors. A linear motion model is used and the temporal reference of the B and P frame is used to scale the motion vector appropriately. FIG. 3a depicts the motion vector sealing and the formula is shown below:
MV F =(TR B ×MV)/TR P (1)
MV B =((TR B −TR P )×MV)/TR P (2)
where
-
- MV is the motion vector of the P-block.
- MV F and MV B are the forward and backward motion vectors for the B-block.
- TR B is the increment in the temporal reference from the last P-frame to the current B-frame, and
- TR P is the increment in the temporal reference from the last P-frame to the current P-frame.
Currently the method used in the prior art assumes a linear motion model. However this assumption is not valid in a normal scene where the motion is typically not linear. This is especially true when the camera shakes and when objects are not moving at constant velocities.
A second problem involves the quantization and transmission of the residual of the predication error in the B-block Currently the coefficients from the P-block and the B-block are interleaved in some scanning order which requires the B-block efficients to be transmitted even when they are all zero. This is not very efficient as it is quite often that there are no residual coefficients to transmit (all coefficients are zero).
SUMMARY OF THE INVENTION
In order to solve the first problem, the current invention employs a delta motion vector to compensate for the non-linear motion. Thus it becomes necessary for the encoder to perform an additional motion search to obtain the optimum delta motion vector that when added to the derived motion vectors would result in the best match in the prediction. This delta motion vectors are transmitted to the decoder at the block level only when necessary. A flag is used to indicate to the decoder if there are delta motion vectors present for the B-block.
For the second problem, this invention also uses a flag to indicate if there are coefficients for the B-block to be decoded.
The operation of the Invention is described as follows.
FIG. 3a shows the linear motion model used for the derivation of the forward and backward motion vectors from the P-block motion vector and the temporal reference information. As illustrated in FIG. 3b, this model breaks down when the motion is not linear. The derived forward and backward motion vector is different from the actual motion vector when the motion is not linear. This is especially true when objects in the scene are moving at changing velocities.
In the current invention the problem is solved by adding a small delta motion vector to the derived motion vector to compensate for the difference between the derived and true motion vector. Therefore the equation in (1) and (2) are now replaced by equations (3) and (4), respectively.
MV F ′=(TR B ×MV)/TR P +MV Delta (3)
MV B ′=((TR B =TR P )×MV)/TR P −MV Delta (4)
where
-
- MV is the motion vector of the P-block.
- MV Delta is the delta motion vector.
- MV F and MV B are the new forward and backward motion vectors for the B-block according to the current invention.
- TR B is the increment in the temporal reference from the last P-frame to the current B-frame. and
- TR p is the increment in the temporal reference from the last P-frame to the current P-frame.
- Note: Equations (3) and (4) are used for the motion vector in the horizontal as well as the vertical directions. Thus the motion vectors are in pairs and there are actually two independent delta motion vectors, one each for the horizontal and vertical directions.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1a is a prior art which illustrates the predictions mode used in the ITU-T Recommendation H.261 Standard.
FIG. 1b is a prior art which illustrates the prediction mode used in the ISO-IEC/JTC MPEG Standard.
FIG. 2a illustrates the PB-frame prediction mode.
FIG. 2b illustrates the B-block bi-directional prediction mode.
FIG. 3a illustrates the linear motion model.
FIG. 3b illustrates the non-linear motion model of the current invention.
FIG. 4 illustrates the encoder functionality block diagram.
FIG. 5 illustrates the B-block bi-directional prediction functionality block diagram.
FIG. 6 illustrates the decoder functionality block diagram.
PREFERRED EMBODIMENTS
The preferred embodiment of the current invention is described here. FIG. 4 illustrates the encoding functionality diagram. The present invention deals with the method for deriving the motion vectors for the B-block. The encoding functionality is presented here for completeness of the embodiment.
The encoding functionality block diagram depicts an encoder using a motion estimation and compensation for reducing the temporal redundancy in the sequence to be coded. The input sequences is organized into a first frame and pairs of subsequent frames. The first frame, hereafter referred to as the I-frame, is coded independent of all other frames. The pair of subsequent frames, hereafter referred to as PB-frame, consist of a B-frame followed by a P-frame. The P-frame is forward predicted based on the previously reconstructed I-frame or P-frame and the B-frame is bi-directionally predicted based on the previously reconstructed I-frame or P-frame and the information in the current P-frame.
The input frame image sequence, 1 , is placed in the Frame Memory 2 . If the frame is classified as an I-frame or a P-frame it is passed through line 14 to the Reference Memory 3 , for use as the reference frame in the motion estimation of the next PB-frame to be predictively encoded. The signal is then passed through line 13 to the Block Sampling module 4 , where it is partitioned into spatially non-overlapping blocks of pixel data for further processing.
If the frame is classified as an I-frame, the sampled blocks are passed through line 16 to the DCT module 7 . If the frame is classified as a PB-frame, the sampled blocks are passed through line 17 to the Motion Estimation module 5 . The Motion Estimation module 5 uses information from the Reference Frame Memory 3 and the current block 17 to obtain the motion vector for that provides the best match for the P-block. The motion vector and the local reconstructed frame, 12 , are passed through line 19 and 20 , respectively, to the Motion Compensation module 6 . The difference image is formed by subtracting the motion compensated decoded frame, 21 , from the current P-block 15 . This signal is then passed through line 22 to the DCT module 7 .
In the DCT module 7 , each block is transformed into the DCT domain coefficients. The transform coefficienta are passed through line 23 to Quantization module 8 , where they are quantizied. The quantizied coefficients are then passed though line 24 to the Run-length & Variable Length Coding module 9 . Here the coefficients are entropy coded to form the Output Bit Stream 25 .
If the current block is an I-block or a P-block, the quantized coefficients are also passed through line 26 to the Inverse Quantization module 10 . The output of the Inverse Quantization 10 , is then passed through line 27 to the Inverse DCT module 11 . If the current block is an I-block then the reconstructed block is placed, via line 28 , in the Local Decoded Frame Memory 12 . If the current block is a P-block then the output of the Inverse DCT 29 is added to the motion compensated output 21 , to from the reconstructed block 30 . The reconstructed block 30 , is then placed in the Local Decoded Frame Memory 12 , for the motion compensation of the subsequent frames.
After the P-block have been locally reconstructed, the information is passed again to the Motion Compensation Module 6 , where the prediction of the B-block is formed. FIG. 5 shows a more detailed functional diagram for the B-block prediction process. The P-motion vector derived in the Motion Estimation module 51 , is passed through line 57 to the Motion Vector Sealing Module 53 . Here the forward and backward motion vectors of the B-block is derived using the formula (1) and (2), respectively. In the present embodiment, an additional motion search around these vectors is performed in the Delta Motion Search module 54 , to obtain the delta motion vector. In this embodiment the motion vector is obtained by performing the search for all detail motion vector values between −3 and 3. The delta motion vector value that gives the best prediction in terms of the samples mean absolute difference in the pixel values of the B-block and the prediction block is chosen. The prediction is formed in the Bi-directional Motion Compensation module 55 , according to FIG. 2b using the information from the Local Decoded Frame Memory 52 , and the Current Reconstructed P-block 59 . In the bi-directional prediction, only information available in the corresponding P-block is used to predict the B-block. The average of the P-block information and the information from the Local Decoded Frame is used to predict the B-block. The rest of the B-block is predicted using information from the Local Decoded Frame only.
The reduction difference block is then passed through line 22 to the DCT module 7 . The DCT coefficients are then passed through line 23 to the Quantization module 8 . The result of the Quantization module 8 , is passed through line 24 to the Run-length & Variable Length Coding 9 . In this module the presence of the delta motion vector and the quantized residual error in the Output Bitstream 25 , is indicated a variable length code. NOB which is the acronym for No B-block. This flag is generated in Run-length & Variable Length Coding module 9 based on whether there are residual error in the Quantization module 8 and delta motion vectors found in the Delta Motion Search module 54 is not zero. Table 1 provides the preferred embodiment of the variable length code for the NOB flag. The variable length code of the NOB flag is inserted in the Output Bitstream, 25 , prior to the delta motion vector an quantized residual error codes.
| TABLE 1 | ||
| (Variable length code for the NOB flag) | ||
| Quantized Residual | Delta Motion | |
| NOB | Error Coded | Vectors Coded |
| 0 | No | No |
| 10 | No | Yes |
| 110 | Yes | No |
| 111 | Yes | Yes |
FIG. 6 shows the functional block diagram for the decoder. The Input Bit Stream 31 , is passed to the Variable Length & Run Length Decoding module 32 . The block and side information are extracted in this module. If the frame is a PB-frame then the bitstream is checked if any delta motion vector and/or quantized residual error coefficients present. The output of the module 32 is passed through line 37 to the Inverse Quantization module 33 . The output of the Inverse Quantization 33 , is then passed through line 38 to the Inverse DCT module 34 . Here the coefficients are transformed back into the pixel values.
If the current frame is an I-frame then the output of Inverse DCT 34 , is passed through line 39 and stored in the Frame Memory 42 .
If the current frame is a PB-frame, the side information containing the motion vector are passe through line 45 to the Motion compensation module 36 . The motion Compensation module 36 , uses this information and the information in the Local Decoded Memory, 35 , to from the motion compensated signal, 44 . This signal is then added to the output of the Inverse DCT module 34 , to form the reconstruction of the P-block.
The Motion Compensation module 36 , then uses the additional information obtained in the reconstructed P-block to obtain the bi-directional prediction for the B-block. The B-block is then reconstructed and placed in the Frame Memory, 42 , together with the P-block.
By implementing this invention, the temporal frame rate of the decoded sequences can be effectively doubled at a fraction of the expected cost in bit rate. The delay is similar to that of the same sequence decoded at half the frame rate.
As described above in the present invention a new predictive coding is used to increase the temporal frame rate and coding efficiency without introducing excessive delay. Currently the motion vector for the blocks in the bi-directionally predicted frame is derived from the motion vector of the corresponding block in the forward predicted frame using a linear motion model. This however is not effective when the motion in the image sequence is not linear. According to this invention, the efficiency of this method can be further improved if a non-linear motion model is used. In this model a delta motion vector is added to or subtracted from the derived forward and backward motion vector, respectively. The encoder performs an additional search to determine if there is a need for the delta motion vector. The presence of this delta motion vector in the transmitted bitstream is sinaglled to the decoder which then takes the appropriate action to make use of the delta motion vector to derive the effective forward and backward motion vectors for the bi-directionally predicted block.